Merge pull request 'ffmpeg-v4l2-request-fourier: route H.264 chroma intra deblock (4:2:0) through daedalus-fourier (0013)' (#102) from claude-noether/marfrit-packages:noether/h264-substitute-deblock-chroma-intra into main

Reviewed-on: marfrit/marfrit-packages#102

.gitea/scripts/check-already-published.sh

@@ -0,0 +1,230 @@
+#!/bin/bash
+# check-already-published.sh <recipe-dir>
+#
+# Decide whether a given recipe (arch/<name> or debian/<name>) is already
+# present in https://packages.reauktion.de/.  Emits exactly one line to
+# stdout:
+#
+#   skip=1   — package with this version-pkgrel-arch tuple already lives in
+#              the pool; CI should short-circuit.
+#   skip=0   — file is missing or HEAD failed; CI should build + publish.
+#
+# Design notes:
+#   * For Arch recipes we source the PKGBUILD in a clean subshell so
+#     shell expansions (epoch=, ${_pkgver/-/}, pkgname=() arrays) resolve
+#     naturally.  Only the first element of pkgname[] is checked — split
+#     packages share one source tarball / one build, so any-one-missing
+#     forces the full rebuild anyway.
+#   * For Debian recipes we extract the bare top-level PKGVER= /
+#     PKGREL= assignments (plus any other top-level VAR=value lines they
+#     reference) via grep and re-evaluate them in an isolated subshell —
+#     sourcing the entire build-deb.sh would run curl/tar/dpkg-deb
+#     against a tempdir we don't want to materialise here.
+#   * Epoch handling differs by ecosystem: Arch keeps `<epoch>:` in the
+#     pool filename, Debian/reprepro strips it.
+#   * curl --head with -f maps non-2xx to non-zero exit, which is what we
+#     want — 404 means "build it".  -L follows mirrors.  --max-time caps
+#     the worst-case latency per HEAD.
+set -euo pipefail
+
+REPO_BASE="${REPO_BASE:-https://packages.reauktion.de}"
+HEAD_TIMEOUT="${HEAD_TIMEOUT:-15}"
+
+RECIPE_DIR="${1:?usage: $0 <recipe-dir>   (e.g. arch/distcc-avahi or debian/lmcp)}"
+
+# Resolve relative to repo root if a leading path is passed; allow
+# both `arch/foo` and absolute paths.
+if [ ! -d "$RECIPE_DIR" ]; then
+  echo "error: recipe dir not found: $RECIPE_DIR" >&2
+  exit 2
+fi
+
+ecosystem="${RECIPE_DIR%%/*}"
+
+http_head() {
+  local url="$1"
+  curl -sS -L --max-time "$HEAD_TIMEOUT" -o /dev/null \
+       -w '%{http_code}' --head "$url" || echo "000"
+}
+
+emit() {
+  # one-line GITHUB_OUTPUT-compatible kv
+  echo "skip=$1"
+  exit 0
+}
+
+case "$ecosystem" in
+arch)
+  pkgbuild="$RECIPE_DIR/PKGBUILD"
+  [ -f "$pkgbuild" ] || { echo "error: $pkgbuild missing" >&2; exit 2; }
+
+  # Source in a fresh bash to capture variables. Some PKGBUILDs run
+  # functions or call commands at top level — keep this fast by
+  # restricting PATH and trapping side effects.
+  eval "$(
+    bash --noprofile --norc -c "
+      set +e
+      # Stub out anything that might shell out; we only need variable
+      # assignments to land.
+      cd '$RECIPE_DIR'
+      source ./PKGBUILD >/dev/null 2>&1 || true
+      # pkgname may be array; print first element.
+      if declare -p pkgname 2>/dev/null | grep -q 'declare -a'; then
+        first_name=\"\${pkgname[0]}\"
+      else
+        first_name=\"\$pkgname\"
+      fi
+      if declare -p arch 2>/dev/null | grep -q 'declare -a'; then
+        first_arch=\"\${arch[0]}\"
+      else
+        first_arch=\"\$arch\"
+      fi
+      printf 'PB_NAME=%q\n'   \"\$first_name\"
+      printf 'PB_VER=%q\n'    \"\$pkgver\"
+      printf 'PB_REL=%q\n'    \"\$pkgrel\"
+      printf 'PB_EPOCH=%q\n'  \"\${epoch:-}\"
+      printf 'PB_ARCH=%q\n'   \"\$first_arch\"
+    "
+  )"
+
+  if [ -z "${PB_NAME:-}" ] || [ -z "${PB_VER:-}" ] || [ -z "${PB_REL:-}" ]; then
+    echo "error: failed to parse PKGBUILD ($RECIPE_DIR)" >&2
+    emit 0
+  fi
+
+  # Pool arch:
+  #   arch=('any')                → any
+  #   arch=('aarch64' 'x86_64')   → aarch64 (we publish for both, but the
+  #                                 aarch64 artifact is the canonical CI build)
+  #   arch=('aarch64')            → aarch64
+  case "$PB_ARCH" in
+    any) pool_arch=any ;;
+    *)   pool_arch=aarch64 ;;
+  esac
+
+  # Version string with optional epoch (epoch:pkgver-pkgrel).
+  if [ -n "${PB_EPOCH:-}" ]; then
+    ver_full="${PB_EPOCH}:${PB_VER}-${PB_REL}"
+  else
+    ver_full="${PB_VER}-${PB_REL}"
+  fi
+
+  # Pool URL path (arch keeps any/aarch64 split; 'any' lands in the
+  # aarch64 dir per current marfrit layout — both arches share the
+  # blob via the publish-to-both-arches step in build.yml).
+  pool_dir="arch/aarch64"
+
+  base_url="${REPO_BASE}/${pool_dir}/${PB_NAME}-${ver_full}-${pool_arch}.pkg.tar"
+  for ext in zst xz gz; do
+    code=$(http_head "${base_url}.${ext}")
+    if [ "$code" = "200" ]; then
+      emit 1
+    fi
+  done
+  emit 0
+  ;;
+
+debian)
+  bd="$RECIPE_DIR/build-deb.sh"
+  ctrl="$RECIPE_DIR/control"
+  [ -f "$bd" ] || { echo "error: $bd missing" >&2; exit 2; }
+
+  # Pull top-level `VAR=value` lines until we've passed PKGREL, and
+  # only those whose RHS is safe to re-evaluate (no command
+  # substitution `$(...)`, no escaped `\$`, no embedded commands like
+  # `DESTDIR=... meson ...`).  This deliberately undershoots: we just
+  # need PKGVER/PKGREL plus any version vars they reference. Anything
+  # else (HERE=$(readlink ...), KERNELVER=\$(uname -r) inside a
+  # HEREDOC, etc.) gets dropped.
+  assigns=$(awk '
+    /^[A-Z_][A-Z0-9_]*=/ {
+      # split into LHS and RHS
+      eq = index($0, "=")
+      lhs = substr($0, 1, eq - 1)
+      rhs = substr($0, eq + 1)
+      # strip inline `# comment`
+      hash = index(rhs, "#")
+      if (hash > 1 && substr(rhs, hash-1, 1) == " ") rhs = substr(rhs, 1, hash - 2)
+      # reject lines with command-subst or escaped-dollar or naked commands
+      if (rhs ~ /\$\(/)   next
+      if (rhs ~ /\\\$/)   next
+      if (rhs ~ / [a-z]/) next   # e.g. `DESTDIR="$ROOT" meson ...`
+      print lhs "=" rhs
+      if (lhs == "PKGREL") exit
+    }
+  ' "$bd")
+
+  eval "$(
+    bash --noprofile --norc -c "
+      set +e
+      $assigns
+      printf 'PKGVER=%q\n' \"\${PKGVER:-}\"
+      printf 'PKGREL=%q\n' \"\${PKGREL:-}\"
+    "
+  )"
+
+  if [ -z "${PKGVER:-}" ] || [ -z "${PKGREL:-}" ]; then
+    echo "error: failed to parse PKGVER/PKGREL from $bd" >&2
+    emit 0
+  fi
+
+  # Strip epoch (`N:` prefix) — debian pool filenames omit it.
+  ver_no_epoch="${PKGVER#*:}"
+  # If PKGVER had no colon, ${PKGVER#*:} returns PKGVER unchanged (bash quirk:
+  # the pattern must match for the prefix to be stripped). Guard explicitly.
+  case "$PKGVER" in
+    *:*) : ;;
+    *)   ver_no_epoch="$PKGVER" ;;
+  esac
+
+  ver_full="${ver_no_epoch}-${PKGREL}"
+
+  # Architecture: parse control's `Architecture:` field.
+  if [ ! -f "$ctrl" ]; then
+    # Some recipes ship debian/control instead of ./control
+    ctrl="$RECIPE_DIR/debian/control"
+  fi
+  ctrl_arch=$(grep -m1 '^Architecture:' "$ctrl" 2>/dev/null | awk '{print $2}')
+  case "$ctrl_arch" in
+    all)            file_arch=all ;;
+    arm64|any)      file_arch=arm64 ;;
+    amd64)          file_arch=amd64 ;;
+    *)              file_arch=arm64 ;;  # conservative default
+  esac
+
+  pkg_name=$(basename "$RECIPE_DIR")
+
+  # Compare against the canonical Packages index (what apt actually
+  # consults).  reprepro refuses lower-version uploads, so checking
+  # only an exact source-pkgrel URL produces an endless-rebuild trap
+  # whenever source PKGREL has rolled back below pool head.  We skip
+  # if pools published version >= source version-tuple.
+  source_full="${ver_full}"
+  if [ -n "${PKGVER#*:}" ] && [ "${PKGVER}" != "${PKGVER#*:}" ]; then
+    # PKGVER had an epoch — keep it for dpkg --compare-versions.
+    source_full="${PKGVER}-${PKGREL}"
+  fi
+
+  # Determine suite: most recipes publish to both bookworm and trixie;
+  # checking trixie is sufficient (changelogs share Distribution).
+  suite="trixie"
+  pkg_arch_label="$file_arch"
+  [ "$file_arch" = "all" ] && pkg_arch_label="all"
+  packages_url="${REPO_BASE}/debian/dists/${suite}/main/binary-arm64/Packages"
+  [ "$file_arch" = "amd64" ] && packages_url="${REPO_BASE}/debian/dists/${suite}/main/binary-amd64/Packages"
+
+  pool_ver=$(set +o pipefail; curl -sS --max-time "$HEAD_TIMEOUT" "$packages_url" 2>/dev/null     | awk -v p="$pkg_name" '$1=="Package:" && $2==p {found=1; next} found && $1=="Version:" {print $2; exit}')
+
+  if [ -n "$pool_ver" ] && command -v dpkg >/dev/null &&        dpkg --compare-versions "$pool_ver" ge "$source_full"; then
+    echo "pool has $pool_ver >= source $source_full" >&2
+    emit 1
+  fi
+  echo "pool has $pool_ver, source wants $source_full — build" >&2
+  emit 0
+  ;;
+
+*)
+  echo "error: unsupported ecosystem '$ecosystem' (recipe-dir=$RECIPE_DIR)" >&2
+  emit 0
+  ;;
+esac

arch/aish/PKGBUILD

@@ -0,0 +1,53 @@
+# Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+# aish — AI-augmented conversational shell in LuaJIT.
+# Source of truth: git.reauktion.de/marfrit/aish
+
+pkgname=aish
+pkgver=0.1.0
+pkgrel=1
+pkgdesc="AI-augmented conversational shell (LuaJIT, FFI-only)"
+arch=('any')
+url="https://git.reauktion.de/marfrit/aish"
+license=('MIT')
+depends=('luajit' 'readline' 'curl')
+# The _tag back-translation handles both clean releases (no '_') and
+# pre-release pkgvers (e.g. 0.1.0_rc1 → v0.1.0-rc1).
+_tag="v${pkgver//_/-}"
+source=("${pkgname}-${pkgver}.tar.gz::https://git.reauktion.de/marfrit/aish/archive/${_tag}.tar.gz")
+sha256sums=('9ebc3939e028832e39391ae33efacb5ec9bcd99d123cbc8ca1cd6ca9a640b5b5')
+
+package() {
+    cd "${pkgname}"
+    local libdir="${pkgdir}/usr/share/lua/5.1/aish"
+
+    # Top-level modules
+    install -Dm644 main.lua     "${libdir}/main.lua"
+    install -Dm644 broker.lua   "${libdir}/broker.lua"
+    install -Dm644 context.lua  "${libdir}/context.lua"
+    install -Dm644 executor.lua "${libdir}/executor.lua"
+    install -Dm644 history.lua  "${libdir}/history.lua"
+    install -Dm644 mcp.lua      "${libdir}/mcp.lua"
+    install -Dm644 renderer.lua "${libdir}/renderer.lua"
+    install -Dm644 repl.lua     "${libdir}/repl.lua"
+    install -Dm644 router.lua   "${libdir}/router.lua"
+    install -Dm644 safety.lua   "${libdir}/safety.lua"
+    install -Dm644 secrets.lua  "${libdir}/secrets.lua"
+
+    # FFI bindings
+    install -Dm644 ffi/curl.lua     "${libdir}/ffi/curl.lua"
+    install -Dm644 ffi/libc.lua     "${libdir}/ffi/libc.lua"
+    install -Dm644 ffi/pty.lua      "${libdir}/ffi/pty.lua"
+    install -Dm644 ffi/readline.lua "${libdir}/ffi/readline.lua"
+
+    # Vendored dependencies
+    install -Dm644 vendor/dkjson.lua "${libdir}/vendor/dkjson.lua"
+
+    # Launch wrapper
+    install -Dm755 bin/aish "${pkgdir}/usr/bin/aish"
+
+    # Documentation + example config
+    install -Dm644 README.md  "${pkgdir}/usr/share/doc/${pkgname}/README.md"
+    install -Dm644 LICENSE    "${pkgdir}/usr/share/doc/${pkgname}/LICENSE"
+    install -Dm644 examples/config.lua \
+        "${pkgdir}/usr/share/doc/${pkgname}/examples/config.lua"
+}

arch/chromium-fourier/PKGBUILD

@@ -8,13 +8,13 @@
 # NEXT.md alongside this PKGBUILD for the full rationale and the
 # validation log on PineTab2 (RK3566).
 #
-# Multi-arch: builds natively on x86_64 and aarch64. The x86_64 path
-# is primarily a development / CI host; the runtime target audience is
-# aarch64. The two patches are architecture-independent.
+# Cross-compiled from x86_64 using chromium's bundled clang (upstream
+# LLVM doesn't ship clang 23+ yet; chromium's internal fork is required).
+# Runtime target is aarch64. The three patches are architecture-independent.
 
 pkgname=chromium-fourier
-pkgver=147.0.7727.116
-pkgrel=2
+pkgver=148.0.7778.178
+pkgrel=1
 epoch=1
 pkgdesc='Chromium with V4L2VDA HW video decode unlocked for mainline Linux Wayland on Rockchip'
 arch=('aarch64' 'x86_64')
@@ -150,7 +150,6 @@ build() {
     'symbol_level=0'
     'is_cfi=false'
     'treat_warnings_as_errors=false'
-    'enable_nacl=false'
     'enable_widevine=false'
 
     # System toolchain (clang/lld from pacman)

arch/chromium-fourier/patches/nv12-external-oes-on-modifier-external-only.patch

@@ -73,16 +73,15 @@ diff --git a/ui/ozone/common/native_pixmap_egl_binding.cc b/ui/ozone/common/nati
 index 31877f4459..6855c1093e 100644
 --- a/ui/ozone/common/native_pixmap_egl_binding.cc
 +++ b/ui/ozone/common/native_pixmap_egl_binding.cc
-@@ -6,10 +6,13 @@
- 
+@@ -6,9 +6,12 @@
+
  #include <array>
- 
+
 +#include "base/containers/flat_map.h"
  #include "base/logging.h"
  #include "base/memory/scoped_refptr.h"
 +#include "base/no_destructor.h"
  #include "base/notreached.h"
- #include "base/numerics/safe_conversions.h"
 +#include "base/synchronization/lock.h"
  #include "ui/gfx/linux/drm_util_linux.h"
  #include "ui/gl/gl_bindings.h"

arch/daedalus-v4l2-dkms/PKGBUILD

@@ -0,0 +1,76 @@
+# Maintainer: Markus Fritsche <fritsche.markus@gmail.com>
+#
+# daedalus-v4l2-dkms — DKMS package for the daedalus_v4l2 out-of-tree
+# kernel module (V4L2 stateless decoder shim for Pi 5 / CM5).
+#
+# Pair to daedalus-v4l2 (userspace daemon).  When loaded, the module
+# registers /dev/videoNN (V4L2 m2m) + /dev/mediaNN (media controller) +
+# /dev/daedalus-v4l2 (chardev to the userspace daemon).  Userspace
+# clients drive the V4L2 m2m + request API path; the daemon does the
+# actual FFmpeg-backed decode on /dev/daedalus-v4l2.
+#
+# Project: https://git.reauktion.de/reauktion/daedalus-v4l2
+# Sibling userspace package: daedalus-v4l2
+# Sibling consumer: libva-v4l2-request-fourier
+
+pkgname=daedalus-v4l2-dkms
+_module=daedalus_v4l2
+
+# Same pin as arch/daedalus-v4l2 — keep kernel module + daemon
+# bit-versioned together so the chardev wire protocol stays in sync.
+# 5d8b436 reverts PRs #7 + #8 (parking design that broke libva's
+# 1:1 contract — see daedalus-v4l2#9 + #10).  Tree is
+# content-equivalent to f0d4186 plus PR #4 (cosmetic menu ctrls).
+# PROTO_VERSION drops 1 → 0; lock-step install with
+# daedalus-v4l2 0.1.0.r33.5d8b436 REQUIRED.
+_commit=872eec505eb91b561892d02a0526749348ddc121
+
+pkgver=0.1.0.r45.872eec5
+pkgrel=1  # reset for new upstream pin (872eec5 — PROTO_MAX_PAYLOAD 64 KiB -> 1 MiB, closes #19); lock-step with daedalus-v4l2 0.1.0.r45.872eec5 REQUIRED
+pkgdesc="V4L2 stateless decoder shim kernel module (DKMS) — Pi 5 / CM5"
+arch=('any')
+url="https://git.reauktion.de/reauktion/daedalus-v4l2"
+license=('GPL-2.0-or-later')
+depends=('dkms')
+makedepends=('git')
+install="${pkgname}.install"
+
+source=("git+https://git.reauktion.de/reauktion/daedalus-v4l2.git#commit=${_commit}"
+        "dkms.conf"
+        "${pkgname}.install")
+sha256sums=('SKIP'
+            'SKIP'
+            'SKIP')
+
+pkgver() {
+    cd "${srcdir}/daedalus-v4l2"
+    printf '0.1.0.r%s.%s' \
+        "$(git rev-list --count HEAD)" \
+        "$(git rev-parse --short=7 HEAD)"
+}
+
+package() {
+    local _src="${pkgdir}/usr/src/${_module}-${pkgver}"
+
+    install -dm755 "${_src}"
+
+    # Install the kernel/ subdir and embed the shared proto header in
+    # the same tree.  The in-tree Makefile uses
+    # `ccflags-y += -I$(src)/../include` (assuming the parent
+    # daedalus-v4l2 layout); for DKMS we flatten by copying the header
+    # into kernel/include/ and patching the Makefile to point there.
+    cp -r "${srcdir}/daedalus-v4l2/kernel/." "${_src}/"
+    install -Dm644 "${srcdir}/daedalus-v4l2/include/daedalus_v4l2_proto.h" \
+        "${_src}/include/daedalus_v4l2_proto.h"
+    sed -i 's|-I\$(src)/\.\./include|-I$(src)/include|' "${_src}/Makefile"
+
+    # dkms.conf at the root of the source tree (DKMS convention).
+    # Substitute #MODULE_VERSION# placeholder with the actual pkgver
+    # so dkms install/uninstall match what's on disk.
+    install -Dm644 "${srcdir}/dkms.conf" "${_src}/dkms.conf"
+    sed -i "s/#MODULE_VERSION#/${pkgver}/" "${_src}/dkms.conf"
+
+    # License
+    install -Dm644 "${srcdir}/daedalus-v4l2/kernel/daedalus_v4l2_main.c" \
+        "${pkgdir}/usr/share/licenses/${pkgname}/SPDX-HEADER"
+}

arch/daedalus-v4l2-dkms/daedalus-v4l2-dkms.install

@@ -0,0 +1,61 @@
+# post-install / post-upgrade hook for daedalus-v4l2-dkms.
+#
+# pacman + the dkms-helpers alpm hook will already attempt
+# `dkms install` on its own.  This script's job is to emit a
+# loud, actionable warning when the module didn't actually
+# build for the running kernel — most commonly because the
+# kernel headers package isn't installed yet.
+#
+# Without this you get a silent failure: the package looks
+# installed but `modprobe daedalus_v4l2` returns ENOENT.
+
+_check_dkms_built() {
+    local name=daedalus_v4l2
+    local ver=$1
+    local kernelver=$(uname -r)
+
+    if ! command -v dkms >/dev/null 2>&1; then
+        return 1  # the hard-dep should have caught this
+    fi
+
+    local status
+    status=$(dkms status -m "$name" -v "$ver" -k "$kernelver" 2>/dev/null || true)
+    if printf '%s\n' "$status" | grep -q -E 'installed|loaded'; then
+        return 0  # all good
+    fi
+
+    cat >&2 <<EOF
+==> daedalus-v4l2-dkms: DKMS build did NOT land for kernel $kernelver.
+==>   dkms status -m $name -v $ver -k $kernelver:
+==>     $(printf '%s' "$status" | head -1)
+==>
+==> Most likely cause: kernel headers package is missing.
+==>   Arch / ALARM:     pacman -S linux-rpi-headers (or linux-rpi5-headers)
+==>   Raspberry Pi OS:  apt install linux-headers-rpi-2712
+==>
+==> After installing headers, finish the install with:
+==>   sudo dkms autoinstall $name/$ver
+==>   sudo modprobe daedalus_v4l2
+==>
+==> Until then daedalus_v4l2 will NOT be loadable and the
+==> userspace daedalus-v4l2 daemon will have nothing to talk to.
+EOF
+    return 1
+}
+
+post_install() {
+    _check_dkms_built "$1" || true
+}
+
+post_upgrade() {
+    # New version pinned by the bump may have built fine, but if
+    # a kernel-headers package was uninstalled / pruned since the
+    # last upgrade we'd silently regress.  Re-check.
+    _check_dkms_built "$1" || true
+}
+
+pre_remove() {
+    # The dkms alpm hook handles dkms remove on its own; nothing
+    # we need to add here.
+    :
+}

arch/daedalus-v4l2-dkms/dkms.conf

@@ -0,0 +1,19 @@
+# DKMS configuration for daedalus_v4l2 — V4L2 stateless decoder shim.
+#
+# Built against /lib/modules/$kernelver/build with the in-tree Makefile.
+# The Makefile uses `obj-m := daedalus_v4l2.o` and links
+# daedalus_v4l2_main.o + daedalus_v4l2_chardev.o into the final .ko.
+
+PACKAGE_NAME="daedalus_v4l2"
+PACKAGE_VERSION="#MODULE_VERSION#"
+
+# Single module produced by the Makefile.
+BUILT_MODULE_NAME[0]="daedalus_v4l2"
+DEST_MODULE_LOCATION[0]="/updates"
+
+# Use the package's own Makefile — it already does
+# `$(MAKE) -C $(KERNELDIR) M=$(PWD) modules`.
+MAKE[0]="make KERNELDIR=/lib/modules/${kernelver}/build all"
+CLEAN="make KERNELDIR=/lib/modules/${kernelver}/build clean"
+
+AUTOINSTALL="yes"

arch/daedalus-v4l2/PKGBUILD

@@ -0,0 +1,118 @@
+# Maintainer: Markus Fritsche <fritsche.markus@gmail.com>
+#
+# daedalus-v4l2 — userspace daemon + V4L2 m2m test tools.
+#
+# Pair to daedalus-v4l2-dkms (kernel module).  Together they expose
+# /dev/videoNN + /dev/mediaNN as a V4L2 stateless decoder shim on Pi 5 /
+# CM5, decoding VP9 / AV1 / H.264 via dlopen'd FFmpeg in a single-
+# threaded daemon and shipping decoded NV12 / P010 back through dmabuf.
+# Consumed end-to-end by libva-v4l2-request-fourier (>= 1.0.0.r376) so
+# `ffmpeg -hwaccel vaapi` against vp9_small.ivf produces byte-exact NV12.
+#
+# Project: https://git.reauktion.de/reauktion/daedalus-v4l2
+# Sibling kernel package: daedalus-v4l2-dkms
+# Sibling consumer: libva-v4l2-request-fourier
+
+pkgname=daedalus-v4l2
+_upstreampkg=daedalus-v4l2
+
+# 6e6dfa1 = picks up daedalus-v4l2 PR #16 — daemon now dlopens
+# the Kwiboo fourier fork's libavcodec.so.62 / libavformat.so.62 /
+# libavutil.so.60 at /opt/fourier instead of Debian-stock soname
+# 61/61/59.  First step on the daedalus-fourier substitution arc
+# (daedalus-v4l2#11).  Daemon still needs daedalus-fourier at
+# build time (Arch packaging for that is a follow-up; Debian side
+# fetches inline via build-deb.sh).
+_commit=872eec505eb91b561892d02a0526749348ddc121
+
+# 0.1.0 (pre-1.0) + commit count + short sha.  Bump the .Y on each
+# Phase 8.x close.  pkgver() recomputes at build time.
+pkgver=0.1.0.r45.872eec5
+pkgrel=1  # reset for new upstream pin (872eec5 — PROTO_MAX_PAYLOAD 64 KiB -> 1 MiB, closes #19); lock-step with daedalus-v4l2-dkms 0.1.0.r45.872eec5 REQUIRED
+pkgdesc="Userspace daemon for the daedalus-v4l2 V4L2 stateless decoder shim (VP9/AV1/H.264 on Pi 5 / CM5)"
+arch=('aarch64')
+url="https://git.reauktion.de/reauktion/daedalus-v4l2"
+license=('BSD-2-Clause' 'GPL-2.0-or-later')
+# Daemon dlopens libavformat.so.61 / libavcodec.so.61 / libavutil.so.59
+# at runtime (Option γ — see daemon/src/ffmpeg_loader.h).  ffmpeg
+# provides those; we don't link them.
+depends=('ffmpeg-v4l2-request-fourier' 'libdrm')
+# Headers from libav*-dev needed at compile time for type-safe function
+# pointer signatures; pkg-config locates them.
+makedepends=('cmake' 'ninja' 'pkgconf' 'git' 'ffmpeg')
+optdepends=('daedalus-v4l2-dkms: kernel module providing /dev/video0 + /dev/daedalus-v4l2'
+            'libva-v4l2-request-fourier: VA-API consumer routing through this daemon')
+install="${pkgname}.install"
+
+source=("git+https://git.reauktion.de/reauktion/daedalus-v4l2.git#commit=${_commit}"
+        "${pkgname}.install")
+sha256sums=('SKIP'
+            'SKIP')
+
+pkgver() {
+    cd "${srcdir}/${_upstreampkg}"
+    printf '0.1.0.r%s.%s' \
+        "$(git rev-list --count HEAD)" \
+        "$(git rev-parse --short=7 HEAD)"
+}
+
+build() {
+    cd "${srcdir}/${_upstreampkg}/daemon"
+    cmake -B build -G Ninja \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DCMAKE_INSTALL_PREFIX=/usr
+    cmake --build build
+
+    cd "${srcdir}/${_upstreampkg}/tools"
+    make
+}
+
+package() {
+    cd "${srcdir}/${_upstreampkg}"
+
+    # Daemon binary
+    install -Dm755 daemon/build/daedalus_v4l2_daemon \
+        "${pkgdir}/usr/bin/daedalus_v4l2_daemon"
+
+    # Test tools (under /usr/libexec to keep them out of the default PATH
+    # — they're for verification, not daily use).
+    install -Dm755 tools/test_chardev_pingpong \
+        "${pkgdir}/usr/libexec/daedalus-v4l2/test_chardev_pingpong"
+    install -Dm755 tools/test_m2m_decode \
+        "${pkgdir}/usr/libexec/daedalus-v4l2/test_m2m_decode"
+    install -Dm755 tools/test_m2m_stream \
+        "${pkgdir}/usr/libexec/daedalus-v4l2/test_m2m_stream"
+
+    # Shared wire-protocol header (kernel ↔ daemon); useful for
+    # third-party clients of the chardev.
+    install -Dm644 include/daedalus_v4l2_proto.h \
+        "${pkgdir}/usr/include/daedalus_v4l2_proto.h"
+
+    # systemd unit + module autoload — without these the daemon never
+    # starts and the libva/VAAPI consumer's REQ_DECODE has nobody on
+    # the other end of /dev/daedalus-v4l2.
+    install -Dm644 packaging/systemd/daedalus-v4l2.service \
+        "${pkgdir}/usr/lib/systemd/system/daedalus-v4l2.service"
+    install -Dm644 packaging/systemd/daedalus-v4l2.modules-load \
+        "${pkgdir}/usr/lib/modules-load.d/daedalus-v4l2.conf"
+
+    # Documentation
+    install -Dm644 README.md \
+        "${pkgdir}/usr/share/doc/${pkgname}/README.md"
+    for d in docs/*.md; do
+        install -Dm644 "$d" "${pkgdir}/usr/share/doc/${pkgname}/$(basename "$d")"
+    done
+
+    # Licenses: BSD-2-Clause for daemon/tools, GPL for the kernel proto
+    # header; the SPDX headers in src/ are the canonical declaration but
+    # ship a short note here for package-manager-driven license queries.
+    install -dm755 "${pkgdir}/usr/share/licenses/${pkgname}"
+    cat > "${pkgdir}/usr/share/licenses/${pkgname}/LICENSE" <<'EOF'
+daedalus-v4l2 userspace components are BSD-2-Clause licensed.
+The shared kernel↔daemon wire protocol header
+(/usr/include/daedalus_v4l2_proto.h) is GPL-2.0-or-later WITH
+Linux-syscall-note for kernel-side compatibility.  See SPDX
+headers on individual source files for the canonical
+per-file declaration.
+EOF
+}

arch/daedalus-v4l2/daedalus-v4l2.install

@@ -0,0 +1,42 @@
+# post_install / post_upgrade for daedalus-v4l2.
+#
+# Enables (but doesn't start — leave that to the operator) the
+# daemon service so it comes up on next boot.  Reloads systemd
+# so the new unit file is visible.  Triggers modules-load so the
+# kernel module loads without a reboot if daedalus-v4l2-dkms is
+# also installed.
+
+_activate() {
+    systemctl daemon-reload >/dev/null 2>&1 || true
+    systemctl enable daedalus-v4l2.service >/dev/null 2>&1 || true
+
+    # Trigger /usr/lib/modules-load.d/daedalus-v4l2.conf without a
+    # reboot.  Safe if the module is already loaded.
+    systemd-modules-load >/dev/null 2>&1 || true
+
+    cat >&2 <<EOF
+==> daedalus-v4l2: service enabled, will start on next boot.
+==> To start now (requires the kernel module to be loaded):
+==>   sudo systemctl start daedalus-v4l2.service
+==> Verify decode path:
+==>   journalctl -u daedalus-v4l2.service -f
+EOF
+}
+
+post_install() {
+    _activate
+}
+
+post_upgrade() {
+    _activate
+    systemctl try-restart daedalus-v4l2.service >/dev/null 2>&1 || true
+}
+
+pre_remove() {
+    systemctl stop daedalus-v4l2.service    >/dev/null 2>&1 || true
+    systemctl disable daedalus-v4l2.service >/dev/null 2>&1 || true
+}
+
+post_remove() {
+    systemctl daemon-reload >/dev/null 2>&1 || true
+}

arch/ffmpeg-v4l2-request-fourier/0002-nv15-to-p010-unpack.patch

@@ -0,0 +1,178 @@
+From 0cd6e669735e453ec8772f111065bbb2f70a5bc6 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Mon, 18 May 2026 07:27:10 +0000
+Subject: [PATCH] avutil/hwcontext_v4l2request: unpack NV15 to P010 in
+ transfer_data_from
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+V4L2_PIX_FMT_NV15 (RK3399/RK3588 rkvdec 10-bit 4:2:0 capture) is mapped to
+sw_format = AV_PIX_FMT_YUV420P10 in v4l2request_capture_pixelformats[]. The
+existing transfer_get_formats explicitly blanked the format list for that
+sw_format, so 'ffmpeg -hwaccel v4l2request -vf hwdownload,format=p010le' on
+a Hi10P / Main10 input failed at filter init with EINVAL before reaching
+the actual decode (which itself succeeds — 2 frames decoded cleanly).
+
+Expose AV_PIX_FMT_P010 as the transfer target for NV15-backed surfaces and
+unpack the packed 10-bit samples into the standard high-bits-of-16 layout
+inside transfer_data_from. Luma and chroma share the same packing format
+(5 bytes per 4 samples, little endian); chroma plane is W × H/2 samples
+for 4:2:0.
+
+The other 'needs custom unpack' sw_formats (YUV420P / Allwinner NV12_32L32
+tiled and YUV422P10 / rkvdec NV20) keep the original ENOSYS path because
+they need different unpack code that isn't covered by this patch.
+
+Closes marfrit/marfrit-packages#21.
+---
+ libavutil/hwcontext_v4l2request.c | 111 +++++++++++++++++++++++++++++-
+ 1 file changed, 110 insertions(+), 1 deletion(-)
+
+diff --git a/libavutil/hwcontext_v4l2request.c b/libavutil/hwcontext_v4l2request.c
+index b6633d9081..3842160dfb 100644
+--- a/libavutil/hwcontext_v4l2request.c
++++ b/libavutil/hwcontext_v4l2request.c
+@@ -1073,6 +1073,56 @@ fail:
+     return ret;
+ }
+ 
++/*
++ * Unpack one NV15-packed 10-bit plane (5 bytes per 4 samples, little endian)
++ * into a P010-style plane (10 bits in the high bits of a 16-bit container).
++ * `dst_stride` is in bytes; `src_stride` is bytes per row of NV15 data.
++ */
++static void v4l2request_nv15_unpack_plane_to_p010(const uint8_t *src,
++                                                  uint16_t *dst,
++                                                  unsigned width,
++                                                  unsigned height,
++                                                  unsigned src_stride,
++                                                  unsigned dst_stride)
++{
++    for (unsigned y = 0; y < height; y++) {
++        const uint8_t *s = src + y * src_stride;
++        uint16_t *d = (uint16_t *)((uint8_t *)dst + y * dst_stride);
++        unsigned x;
++
++        for (x = 0; x + 4 <= width; x += 4) {
++            uint16_t a = (uint16_t)s[0] | ((uint16_t)(s[1] & 0x03) << 8);
++            uint16_t b = ((uint16_t)s[1] >> 2) | ((uint16_t)(s[2] & 0x0F) << 6);
++            uint16_t c = ((uint16_t)s[2] >> 4) | ((uint16_t)(s[3] & 0x3F) << 4);
++            uint16_t e = ((uint16_t)s[3] >> 6) | ((uint16_t)s[4] << 2);
++
++            d[0] = (uint16_t)(a << 6);
++            d[1] = (uint16_t)(b << 6);
++            d[2] = (uint16_t)(c << 6);
++            d[3] = (uint16_t)(e << 6);
++
++            d += 4;
++            s += 5;
++        }
++
++        if (x < width) {
++            unsigned rem = width - x;
++            uint16_t pix[4] = { 0, 0, 0, 0 };
++
++            pix[0] = (uint16_t)s[0] | ((uint16_t)(s[1] & 0x03) << 8);
++            if (rem >= 2)
++                pix[1] = ((uint16_t)s[1] >> 2) | ((uint16_t)(s[2] & 0x0F) << 6);
++            if (rem >= 3)
++                pix[2] = ((uint16_t)s[2] >> 4) | ((uint16_t)(s[3] & 0x3F) << 4);
++            if (rem >= 4)
++                pix[3] = ((uint16_t)s[3] >> 6) | ((uint16_t)s[4] << 2);
++
++            for (unsigned j = 0; j < rem; j++)
++                d[j] = (uint16_t)(pix[j] << 6);
++        }
++    }
++}
++
+ static int v4l2request_transfer_get_formats(AVHWFramesContext *hwfc,
+                                             enum AVHWFrameTransferDirection dir,
+                                             enum AVPixelFormat **formats)
+@@ -1082,6 +1132,22 @@ static int v4l2request_transfer_get_formats(AVHWFramesContext *hwfc,
+     if (dir == AV_HWFRAME_TRANSFER_DIRECTION_TO)
+         return AVERROR(ENOSYS);
+ 
++    /*
++     * NV15-backed surfaces (sw_format = YUV420P10) are exposed as P010 to
++     * downstream filters: the unpack below converts the packed 10-bit
++     * samples into the standard high-bits-of-16 layout. Hi10P / Main10
++     * VAAPI/v4l2-request decode reaches userspace through this path.
++     */
++    if (hwfc->sw_format == AV_PIX_FMT_YUV420P10) {
++        fmts = av_malloc_array(2, sizeof(*fmts));
++        if (!fmts)
++            return AVERROR(ENOMEM);
++        fmts[0] = AV_PIX_FMT_P010;
++        fmts[1] = AV_PIX_FMT_NONE;
++        *formats = fmts;
++        return 0;
++    }
++
+     fmts = av_malloc_array(2, sizeof(*fmts));
+     if (!fmts)
+         return AVERROR(ENOMEM);
+@@ -1089,8 +1155,13 @@ static int v4l2request_transfer_get_formats(AVHWFramesContext *hwfc,
+     fmts[0] = hwfc->sw_format;
+     fmts[1] = AV_PIX_FMT_NONE;
+ 
++    /*
++     * Tiled-NV12-32L32 (Allwinner) and NV20 (rkvdec 4:2:2 10-bit) still need
++     * dedicated unpacks before hwdownload can consume them; leave them as
++     * "no transfer formats" so the filter graph reports the limitation
++     * rather than silently producing garbage.
++     */
+     if (hwfc->sw_format == AV_PIX_FMT_YUV420P ||
+-        hwfc->sw_format == AV_PIX_FMT_YUV420P10 ||
+         hwfc->sw_format == AV_PIX_FMT_YUV422P10)
+         fmts[0] = AV_PIX_FMT_NONE;
+ 
+@@ -1110,6 +1181,44 @@ static int v4l2request_transfer_data_from(AVHWFramesContext *hwfc,
+     map = av_frame_alloc();
+     if (!map)
+         return AVERROR(ENOMEM);
++
++    /*
++     * For NV15→P010, map the raw NV15 bytes (sw_format) and unpack into
++     * dst's P010 storage. Otherwise fall through to the original byte-copy
++     * path used for 1:1 sw_format matches (NV12, NV16, AFBC handled by DRM).
++     */
++    if (hwfc->sw_format == AV_PIX_FMT_YUV420P10) {
++        /*
++         * Only P010 is advertised by transfer_get_formats for this sw_format;
++         * a caller that bypasses get_formats and asks for anything else would
++         * silently corrupt output via av_frame_copy on NV15-packed bytes.
++         * Reject explicitly.
++         */
++        if (dst->format != AV_PIX_FMT_P010) {
++            ret = AVERROR(ENOSYS);
++            goto fail;
++        }
++
++        map->format = hwfc->sw_format;
++        ret = v4l2request_map_frame(hwfc, map, src);
++        if (ret)
++            goto fail;
++
++        v4l2request_nv15_unpack_plane_to_p010(map->data[0],
++                                              (uint16_t *)dst->data[0],
++                                              dst->width, dst->height,
++                                              map->linesize[0],
++                                              dst->linesize[0]);
++        /* NV15 chroma plane is W × H/2 samples (4:2:0, UV interleaved). */
++        v4l2request_nv15_unpack_plane_to_p010(map->data[1],
++                                              (uint16_t *)dst->data[1],
++                                              dst->width, dst->height / 2,
++                                              map->linesize[1],
++                                              dst->linesize[1]);
++        ret = 0;
++        goto fail;
++    }
++
+     map->format = dst->format;
+ 
+     ret = v4l2request_map_frame(hwfc, map, src);
+-- 
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0003-h264-idct4-daedalus-fourier.patch

@@ -0,0 +1,137 @@
+From f760c0541586f43334c02611fcb4c212c08ad576 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Thu, 21 May 2026 21:40:22 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 4x4 IDCT through
+ daedalus-fourier
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+H264DSPContext.idct_add (called per 4x4 block from the intra-4x4
+decode path in h264_mb.c) now dispatches through
+daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
+
+The recipe layer picks the substrate; for cycle 6 (H.264 IDCT 4x4)
+the recipe is CPU NEON, so this is effectively a NEON-to-NEON
+substitution with one extra dispatch call and recipe-table lookup.
+Provides the first end-to-end exercise of the daedalus-fourier
+kernel pack inside the libavcodec.so decode hot path; follow-up
+patches wire IDCT 8x8, luma-v deblock, and qpel mc20.
+
+The library context is process-global, lazily initialised under
+pthread_once on first call.  We pick the no-QPU constructor because
+libavcodec.so is loaded into arbitrary host processes
+(firefox-fourier, mpv-fourier, daedalus_v4l2_daemon, ...) and we
+cannot assume the host has a usable Vulkan instance.  Higher cycles
+(deblock luma-v, MC) that benefit from the QPU will provision their
+own recipe-selected context once that path is wired.
+
+Bulk paths (idct_add16, idct_add16intra, idct_add8 — used for
+non-intra4x4 macroblocks) remain on the stock NEON .S implementations
+and will be batched through daedalus_recipe_dispatch_h264_idct4 with
+n_blocks>1 in a follow-up.
+
+Bit-exact against ff_h264_idct_add_neon (daedalus-fourier cycle 6
+green; see marfrit/daedalus-fourier/CYCLE_LOGS.md).
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2.
+---
+ libavcodec/aarch64/Makefile               |  3 +-
+ libavcodec/aarch64/h264_idct_daedalus.c   | 49 +++++++++++++++++++++++
+ libavcodec/aarch64/h264dsp_init_aarch64.c |  3 +-
+ 3 files changed, 53 insertions(+), 2 deletions(-)
+ create mode 100644 libavcodec/aarch64/h264_idct_daedalus.c
+
+diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
+index 41ab025..7b95fb1 100644
+--- a/libavcodec/aarch64/Makefile
++++ b/libavcodec/aarch64/Makefile
+@@ -3,7 +3,8 @@ OBJS-$(CONFIG_AC3DSP)                   += aarch64/ac3dsp_init_aarch64.o
+ OBJS-$(CONFIG_FDCTDSP)                  += aarch64/fdctdsp_init_aarch64.o
+ OBJS-$(CONFIG_FMTCONVERT)               += aarch64/fmtconvert_init.o
+ OBJS-$(CONFIG_H264CHROMA)               += aarch64/h264chroma_init_aarch64.o
+-OBJS-$(CONFIG_H264DSP)                  += aarch64/h264dsp_init_aarch64.o
++OBJS-$(CONFIG_H264DSP)                  += aarch64/h264dsp_init_aarch64.o \
++                                           aarch64/h264_idct_daedalus.o
+ OBJS-$(CONFIG_HUFFYUVDSP)               += aarch64/huffyuvdsp_init_aarch64.o
+ OBJS-$(CONFIG_H264PRED)                 += aarch64/h264pred_init.o
+ OBJS-$(CONFIG_H264QPEL)                 += aarch64/h264qpel_init_aarch64.o
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+new file mode 100644
+index 0000000..538d223
+--- /dev/null
++++ b/libavcodec/aarch64/h264_idct_daedalus.c
+@@ -0,0 +1,49 @@
++/*
++ * H.264 4x4 IDCT + add — daedalus-fourier substitution shim.
++ *
++ * Routes H264DSPContext.idct_add through
++ * daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
++ * The recipe layer picks the substrate (CPU NEON by default for
++ * cycle 6; future cycles may dispatch to V3D opportunistically).
++ *
++ * FFmpeg's 4x4 block memory layout matches daedalus's column-major
++ * convention: block[r + 4*c] = coefficient at (row r, col c).  Both
++ * sides destructively zero the block after the transform.
++ *
++ * The library context is process-global and lazily initialised under
++ * pthread_once.  We pick the no-QPU constructor here because
++ * libavcodec.so is loaded into arbitrary host processes
++ * (firefox-fourier, mpv-fourier, daedalus_v4l2_daemon, ...) and we
++ * cannot assume the host has a usable Vulkan instance.  Higher cycles
++ * (deblock, MC) that benefit from the QPU initialise their own
++ * recipe-selected context once that path is wired.
++ */
++
++#include <pthread.h>
++#include <stddef.h>
++#include <stdint.h>
++
++#include <daedalus.h>
++
++#include "libavutil/attributes.h"
++#include "libavcodec/h264dsp.h"
++
++static daedalus_ctx     *g_dctx;
++static pthread_once_t    g_dctx_once = PTHREAD_ONCE_INIT;
++
++static void daedalus_ctx_init_once(void)
++{
++    g_dctx = daedalus_ctx_create_no_qpu();
++}
++
++void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
++
++void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
++{
++    static const daedalus_h264_block_meta meta = { .dst_off = 0 };
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_idct4(g_dctx, dst, (size_t)stride,
++                                        block, 1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+index c684574..b993df2 100644
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
++++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
+@@ -66,6 +66,7 @@ void ff_biweight_h264_pixels_4_neon(uint8_t *dst, uint8_t *src, ptrdiff_t stride
+                                     int weights, int offset);
+ 
+ void ff_h264_idct_add_neon(uint8_t *dst, int16_t *block, int stride);
++void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct_dc_add_neon(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct_add16_neon(uint8_t *dst, const int *block_offset,
+                              int16_t *block, int stride,
+@@ -139,7 +140,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
+         c->biweight_pixels_tab[1] = ff_biweight_h264_pixels_8_neon;
+         c->biweight_pixels_tab[2] = ff_biweight_h264_pixels_4_neon;
+ 
+-        c->idct_add        = ff_h264_idct_add_neon;
++        c->idct_add        = ff_h264_idct_add_daedalus;
+         c->idct_dc_add     = ff_h264_idct_dc_add_neon;
+         c->idct_add16      = ff_h264_idct_add16_neon;
+         c->idct_add16intra = ff_h264_idct_add16intra_neon;
+-- 
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0004-h264-idct8-daedalus-fourier.patch

@@ -0,0 +1,107 @@
+From 1b286ddb4efaca26ec9b9e290e989fec77dc1c77 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Fri, 22 May 2026 10:18:21 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 8x8 IDCT through
+ daedalus-fourier
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+H264DSPContext.idct8_add (called per 8x8 block from the High-profile
+intra-8x8-DCT decode path in h264_mb.c) now dispatches through
+daedalus_recipe_dispatch_h264_idct8 instead of ff_h264_idct8_add_neon.
+
+The recipe layer picks the substrate; for cycle 7 (H.264 IDCT 8x8)
+the recipe is CPU NEON, so this is effectively a NEON-to-NEON
+substitution layered on top of the cycle-6 IDCT 4x4 wiring.  Same
+pthread_once global context, same destructive-zero semantics; FFmpeg
+column-major 8x8 storage block[r + 8*c] matches daedalus's convention.
+
+Bulk path c->idct8_add4 (used for inter 8x8-DCT macroblocks) remains
+on the in-tree NEON .S code and will be batched through
+daedalus_recipe_dispatch_h264_idct8 with n_blocks>1 in a follow-up.
+
+Bit-exact against ff_h264_idct8_add_neon (daedalus-fourier cycle 7
+green).
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 7.
+---
+ libavcodec/aarch64/h264_idct_daedalus.c   | 29 ++++++++++++++++-------
+ libavcodec/aarch64/h264dsp_init_aarch64.c |  3 ++-
+ 2 files changed, 23 insertions(+), 9 deletions(-)
+
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+index 538d223..cbb98af 100644
+--- a/libavcodec/aarch64/h264_idct_daedalus.c
++++ b/libavcodec/aarch64/h264_idct_daedalus.c
+@@ -1,14 +1,16 @@
+ /*
+- * H.264 4x4 IDCT + add — daedalus-fourier substitution shim.
++ * H.264 4x4 / 8x8 IDCT + add — daedalus-fourier substitution shims.
+  *
+- * Routes H264DSPContext.idct_add through
+- * daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
+- * The recipe layer picks the substrate (CPU NEON by default for
+- * cycle 6; future cycles may dispatch to V3D opportunistically).
++ * Routes H264DSPContext.idct_add  → daedalus_recipe_dispatch_h264_idct4
++ *        H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
++ * instead of the in-tree ff_h264_idct{,8}_add_neon assembly.  The
++ * recipe layer picks the substrate (CPU NEON by default for cycles
++ * 6 + 7; future cycles may dispatch to V3D opportunistically).
+  *
+- * FFmpeg's 4x4 block memory layout matches daedalus's column-major
+- * convention: block[r + 4*c] = coefficient at (row r, col c).  Both
+- * sides destructively zero the block after the transform.
++ * FFmpeg's 4x4 and 8x8 block memory layouts match daedalus's
++ * column-major convention: block[r + N*c] = coefficient at
++ * (row r, col c) for N ∈ {4, 8}.  Both sides destructively zero the
++ * block after the transform.
+  *
+  * The library context is process-global and lazily initialised under
+  * pthread_once.  We pick the no-QPU constructor here because
+@@ -37,6 +39,7 @@ static void daedalus_ctx_init_once(void)
+ }
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
++void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+ {
+@@ -47,3 +50,13 @@ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+     daedalus_recipe_dispatch_h264_idct4(g_dctx, dst, (size_t)stride,
+                                         block, 1, &meta);
+ }
++
++void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride)
++{
++    static const daedalus_h264_block_meta meta = { .dst_off = 0 };
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_idct8(g_dctx, dst, (size_t)stride,
++                                        block, 1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+index b993df2..741e551 100644
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
++++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
+@@ -79,6 +79,7 @@ void ff_h264_idct_add8_neon(uint8_t **dest, const int *block_offset,
+                             const uint8_t nnzc[15 * 8]);
+ 
+ void ff_h264_idct8_add_neon(uint8_t *dst, int16_t *block, int stride);
++void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct8_dc_add_neon(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct8_add4_neon(uint8_t *dst, const int *block_offset,
+                              int16_t *block, int stride,
+@@ -146,7 +147,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
+         c->idct_add16intra = ff_h264_idct_add16intra_neon;
+         if (chroma_format_idc <= 1)
+             c->idct_add8   = ff_h264_idct_add8_neon;
+-        c->idct8_add       = ff_h264_idct8_add_neon;
++        c->idct8_add       = ff_h264_idct8_add_daedalus;
+         c->idct8_dc_add    = ff_h264_idct8_dc_add_neon;
+         c->idct8_add4      = ff_h264_idct8_add4_neon;
+     } else if (have_neon(cpu_flags) && bit_depth == 10) {
+-- 
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0005-h264-deblock-luma-v-daedalus-fourier.patch

@@ -0,0 +1,121 @@
+From 68731c41d7ea68be0e912b128cb4e71fb56e8263 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Fri, 22 May 2026 12:15:16 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma-v deblock through
+ daedalus-fourier
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+H264DSPContext.v_loop_filter_luma (non-intra bS<4 vertical luma
+deblock, called per macroblock-row edge from the slice deblock
+loop) now dispatches through
+daedalus_recipe_dispatch_h264_deblock_luma_v instead of
+ff_h264_v_loop_filter_luma_neon.
+
+The recipe layer picks the substrate; for cycle 8 the daedalus
+docstring marks the kernel "CPU primary; QPU opportunistic", but
+the libavcodec.so context here is built with
+daedalus_ctx_create_no_qpu — process-global pthread_once init,
+shared with cycles 6/7.  QPU opportunism stays gated off until a
+follow-up adds an explicit feature flag (no implicit Vulkan init
+in arbitrary host processes).  In the meantime cycle 8 is a
+plumbing-only substitution, NEON-to-NEON via the daedalus recipe.
+
+Intra (bS=4) loop filter — c->v_loop_filter_luma_intra — stays on
+the in-tree NEON .S code; daedalus's daedalus_h264_deblock_meta
+only covers the non-intra path per its docstring.
+
+FFmpeg `int alpha/beta/int8_t tc0[4]` → daedalus_h264_deblock_meta
+(int32_t alpha/beta + inline int8_t tc0[4]).  pix already points
+to row 0 of the bottom block per FFmpeg's deblock convention,
+satisfying daedalus's `dst_off >= 4 * dst_stride` constraint.
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8.
+---
+ libavcodec/aarch64/h264_idct_daedalus.c   | 36 +++++++++++++++++++----
+ libavcodec/aarch64/h264dsp_init_aarch64.c |  4 ++-
+ 2 files changed, 33 insertions(+), 7 deletions(-)
+
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+index cbb98af..92365fa 100644
+--- a/libavcodec/aarch64/h264_idct_daedalus.c
++++ b/libavcodec/aarch64/h264_idct_daedalus.c
+@@ -1,11 +1,14 @@
+ /*
+- * H.264 4x4 / 8x8 IDCT + add — daedalus-fourier substitution shims.
++ * H.264 4x4 / 8x8 IDCT + luma-v deblock — daedalus-fourier substitution shims.
+  *
+- * Routes H264DSPContext.idct_add  → daedalus_recipe_dispatch_h264_idct4
+- *        H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
+- * instead of the in-tree ff_h264_idct{,8}_add_neon assembly.  The
+- * recipe layer picks the substrate (CPU NEON by default for cycles
+- * 6 + 7; future cycles may dispatch to V3D opportunistically).
++ * Routes H264DSPContext.idct_add           → daedalus_recipe_dispatch_h264_idct4
++ *        H264DSPContext.idct8_add          → daedalus_recipe_dispatch_h264_idct8
++ *        H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
++ * instead of the in-tree ff_h264_*_neon assembly.  The recipe layer
++ * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
++ * is CPU primary with QPU opportunistic — the ctx below is no-QPU,
++ * so cycle 8 stays on the CPU NEON path until a separate change
++ * gates QPU init on a daedalus-fourier feature flag).
+  *
+  * FFmpeg's 4x4 and 8x8 block memory layouts match daedalus's
+  * column-major convention: block[r + N*c] = coefficient at
+@@ -40,6 +43,8 @@ static void daedalus_ctx_init_once(void)
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
++void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+ {
+@@ -60,3 +65,22 @@ void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+     daedalus_recipe_dispatch_h264_idct8(g_dctx, dst, (size_t)stride,
+                                         block, 1, &meta);
+ }
++
++void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++    meta.tc0[0] = tc0[0];
++    meta.tc0[1] = tc0[1];
++    meta.tc0[2] = tc0[2];
++    meta.tc0[3] = tc0[3];
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_deblock_luma_v(g_dctx, pix, (size_t)stride,
++                                                 1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+index 741e551..85ac381 100644
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
++++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
+@@ -27,6 +27,8 @@
+ 
+ void ff_h264_v_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                      int beta, int8_t *tc0);
++void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0);
+ void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                      int beta, int8_t *tc0);
+ void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+@@ -114,7 +116,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
+     int cpu_flags = av_get_cpu_flags();
+ 
+     if (have_neon(cpu_flags) && bit_depth == 8) {
+-        c->v_loop_filter_luma   = ff_h264_v_loop_filter_luma_neon;
++        c->v_loop_filter_luma   = ff_h264_v_loop_filter_luma_daedalus;
+         c->h_loop_filter_luma   = ff_h264_h_loop_filter_luma_neon;
+         c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
+         c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
+-- 
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0006-h264-restore-low-delay.patch

@@ -0,0 +1,82 @@
+From 0d1292ea99bc4e5fa2da438259fa01a2374e3e04 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Fri, 22 May 2026 14:18:25 +0200
+Subject: [PATCH] avcodec/h264: restore AV_CODEC_FLAG_LOW_DELAY semantics
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+FFmpeg 8.x dropped the H.264 decoder's low_delay path —
+AV_CODEC_FLAG_LOW_DELAY no longer prevents
+h264_select_output_frame from running the display-order DPB
+output queue.  V4L2-stateless-style consumers (daedalus-v4l2
+daemon, libva-v4l2-request-fourier) that set the flag end up
+seeing the 2-1-4-3 pair-swap pattern on B-frame streams again.
+
+Restore the documented semantics:
+
+  - Early-exit at the top of h264_select_output_frame when the
+    flag is set: emit the just-decoded picture immediately as
+    next_output_pic, mirror the corruption / recovery-point
+    tracking the main path performs, and skip the entire
+    delayed_pic[] / POC reorder machinery.
+
+  - Suppress the SPS-driven has_b_frames clobber in
+    h264_field_start when the flag is set, so the per-slice
+    bitstream_restriction_flag re-pickup cannot reintroduce a
+    nonzero reorder buffer mid-stream.
+
+This is a fork-only change required by the daedalus-v4l2 daemon's
+one-frame-per-send_packet contract; upstream FFmpeg consumers that
+expect display-order output remain untouched (flag default = off).
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 deblock
++ flag-restoration follow-up.
+---
+ libavcodec/h264_slice.c | 23 +++++++++++++++++++++++
+ 1 file changed, 23 insertions(+)
+
+diff --git a/libavcodec/h264_slice.c b/libavcodec/h264_slice.c
+index 97fab70..a7bfbd6 100644
+--- a/libavcodec/h264_slice.c
++++ b/libavcodec/h264_slice.c
+@@ -1308,6 +1308,28 @@ static int h264_select_output_frame(H264Context *h)
+     cur->mmco_reset = h->mmco_reset;
+     h->mmco_reset = 0;
+ 
++    /* AV_CODEC_FLAG_LOW_DELAY restore (FFmpeg 8.x dropped the H.264
++     * decoder's low_delay path).  Bypass the display-order DPB
++     * output queue: emit the just-decoded picture immediately, in
++     * decode order, one per send_packet.  V4L2-stateless-style
++     * consumers (daedalus-v4l2 daemon, libva-v4l2-request-fourier)
++     * do their own POC-based reorder downstream and require this
++     * behaviour. */
++    if (h->avctx->flags & AV_CODEC_FLAG_LOW_DELAY) {
++        h->next_output_pic    = cur;
++        h->next_outputed_poc  = cur->poc;
++        h->frame_recovered   |= cur->recovered;
++        cur->recovered       |= h->frame_recovered & FRAME_RECOVERED_SEI;
++        if (!cur->recovered) {
++            if (!(h->avctx->flags  & AV_CODEC_FLAG_OUTPUT_CORRUPT) &&
++                !(h->avctx->flags2 & AV_CODEC_FLAG2_SHOW_ALL))
++                h->next_output_pic = NULL;
++            else
++                cur->f->flags |= AV_FRAME_FLAG_CORRUPT;
++        }
++        return 0;
++    }
++
+     if (sps->bitstream_restriction_flag ||
+         h->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT) {
+         h->avctx->has_b_frames = FFMAX(h->avctx->has_b_frames, sps->num_reorder_frames);
+@@ -1415,6 +1437,7 @@ static int h264_field_start(H264Context *h, const H264SliceContext *sl,
+     sps = h->ps.sps;
+ 
+     if (sps->bitstream_restriction_flag &&
++        !(h->avctx->flags & AV_CODEC_FLAG_LOW_DELAY) &&
+         h->avctx->has_b_frames < sps->num_reorder_frames) {
+         h->avctx->has_b_frames = sps->num_reorder_frames;
+     }
+-- 
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0007-h264-qpel-mc20-daedalus-fourier.patch

@@ -0,0 +1,139 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Sat, 23 May 2026 12:00:00 +0200
+Subject: [PATCH] avcodec/aarch64/h264qpel: route 8x8 mc20 through
+ daedalus-fourier
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+H264QpelContext.put_h264_qpel_pixels_tab[1][2] (8x8 luma horizontal
+half-pel, 6-tap "put" variant — the canonical representative of the
+H.264 luma motion-compensation family) now dispatches through
+daedalus_recipe_dispatch_h264_qpel_mc20 instead of
+ff_put_h264_qpel8_mc20_neon.
+
+Cycle 9 of the daedalus-v4l2#11 step 2 substitution arc; closes the
+4-cycle libavcodec.so substitution sequence (6 IDCT 4x4 / 7 IDCT 8x8 /
+8 luma-v deblock / 9 qpel mc20).
+
+The recipe layer picks the substrate. Per docs/k9_h264qpel_mc20.md
+the verdict is CPU NEON: per-block 7.6 ns at 131 Mblock/s gives 135x
+margin over 30 fps 1080p, and the QPU dispatch floor (~250 ns)
+makes any V3D shader strictly worse. Substitution is plumbing-only,
+NEON-by-recipe — same daedalus_ctx_create_no_qpu pthread_once
+context shape the cycles 6/7/8 shims already own (kept SEPARATE
+from the H264DSP shim's ctx because H264QPEL is its own libavcodec
+Makefile module and link order does not guarantee a single .o
+owns the ctx symbol; one extra ~µs init per process, paid lazily).
+
+Other H.264 luma MC variants (mc02, mc11, mc22 etc.) and the 16x16
+size tier stay on the in-tree NEON .S code. Per the cycle-9 phase-1
+rationale, mc20 8x8 is representative of the whole family's per-block
+cost — extending the substitution to other variants would multiply
+recipe-lookup overhead without changing the substrate verdict.
+
+Bit-exact against ff_put_h264_qpel8_mc20_neon (daedalus-fourier
+cycle 9 green; M1 = 100% bit-exact across 10000 random blocks).
+
+No SONAME change, no Depends change.
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 9.
+---
+ libavcodec/aarch64/Makefile                |  3 +-
+ libavcodec/aarch64/h264_qpel_daedalus.c    | 50 ++++++++++++++++++++++
+ libavcodec/aarch64/h264qpel_init_aarch64.c |  4 +-
+ 3 files changed, 55 insertions(+), 2 deletions(-)
+ create mode 100644 libavcodec/aarch64/h264_qpel_daedalus.c
+
+diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
+--- a/libavcodec/aarch64/Makefile
++++ b/libavcodec/aarch64/Makefile
+@@ -7,7 +7,8 @@ OBJS-$(CONFIG_H264DSP)                  += aarch64/h264dsp_init_aarch64.o \
+                                            aarch64/h264_idct_daedalus.o
+ OBJS-$(CONFIG_HUFFYUVDSP)               += aarch64/huffyuvdsp_init_aarch64.o
+ OBJS-$(CONFIG_H264PRED)                 += aarch64/h264pred_init.o
+-OBJS-$(CONFIG_H264QPEL)                 += aarch64/h264qpel_init_aarch64.o
++OBJS-$(CONFIG_H264QPEL)                 += aarch64/h264qpel_init_aarch64.o \
++                                           aarch64/h264_qpel_daedalus.o
+ OBJS-$(CONFIG_HPELDSP)                  += aarch64/hpeldsp_init_aarch64.o
+ OBJS-$(CONFIG_IDCTDSP)                  += aarch64/idctdsp_init_aarch64.o
+ OBJS-$(CONFIG_ME_CMP)                   += aarch64/me_cmp_init_aarch64.o
+diff --git a/libavcodec/aarch64/h264_qpel_daedalus.c b/libavcodec/aarch64/h264_qpel_daedalus.c
+new file mode 100644
+--- /dev/null
++++ b/libavcodec/aarch64/h264_qpel_daedalus.c
+@@ -0,0 +1,50 @@
++/*
++ * H.264 luma qpel mc20 (8x8, horizontal half-pel, 6-tap "put")
++ * — daedalus-fourier substitution shim.
++ *
++ * Routes H264QpelContext.put_h264_qpel_pixels_tab[1][2] through
++ * daedalus_recipe_dispatch_h264_qpel_mc20 instead of
++ * ff_put_h264_qpel8_mc20_neon.  The recipe layer picks the substrate
++ * (CPU NEON for cycle 9; QPU not viable — per-block 7.6 ns vs
++ * ~250 ns QPU dispatch floor, see docs/k9_h264qpel_mc20.md).
++ *
++ * Sibling to libavcodec/aarch64/h264_idct_daedalus.c.  We keep a
++ * SEPARATE process-global pthread_once context here instead of
++ * sharing the H264DSP one because H264QPEL is its own libavcodec
++ * Makefile module and link order does not guarantee a single .o
++ * owns the ctx symbol.  The cost is one extra
++ * daedalus_ctx_create_no_qpu (~µs) per process; daemon and host
++ * processes pay this lazily on first MC call.
++ *
++ * FFmpeg H264QpelContext convention: both dst and src use a SINGLE
++ * stride and `src` already points at the leftmost OUTPUT column
++ * (col 0); the 6-tap filter reads cols -2..+3.  This matches
++ * daedalus_recipe_dispatch_h264_qpel_mc20's documented contract
++ * directly, so dst_off = src_off = 0.
++ */
++
++#include <pthread.h>
++#include <stddef.h>
++#include <stdint.h>
++
++#include <daedalus.h>
++
++#include "libavutil/attributes.h"
++
++static daedalus_ctx     *g_dctx;
++static pthread_once_t    g_dctx_once = PTHREAD_ONCE_INIT;
++
++static void daedalus_ctx_init_once(void)
++{
++    g_dctx = daedalus_ctx_create_no_qpu();
++}
++
++void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
++
++void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)
++{
++    static const daedalus_h264_qpel_meta meta = { .dst_off = 0, .src_off = 0 };
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++    daedalus_recipe_dispatch_h264_qpel_mc20(g_dctx, dst, src, (size_t)stride,
++                                            1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264qpel_init_aarch64.c b/libavcodec/aarch64/h264qpel_init_aarch64.c
+--- a/libavcodec/aarch64/h264qpel_init_aarch64.c
++++ b/libavcodec/aarch64/h264qpel_init_aarch64.c
+@@ -47,6 +47,8 @@ void ff_put_h264_qpel8_mc00_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t str
+ void ff_put_h264_qpel8_mc10_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc20_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc30_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
++void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
++                                     ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc01_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc11_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc21_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+@@ -184,7 +186,7 @@ av_cold void ff_h264qpel_init_aarch64(H264QpelContext *c, int bit_depth)
+
+         c->put_h264_qpel_pixels_tab[1][ 0] = ff_put_h264_qpel8_mc00_neon;
+         c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_neon;
+-        c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_neon;
++        c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_daedalus;
+         c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_neon;
+         c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_neon;
+         c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_neon;
+--
+2.47.3

arch/ffmpeg-v4l2-request-fourier/0008-h264-deblock-luma-h-daedalus-fourier.patch

@@ -0,0 +1,92 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: claude-noether <claude-noether@noreply.localhost>
+Date: Sun, 25 May 2026 12:00:00 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma-h deblock through daedalus-fourier
+
+Sibling of 0005 (which substituted v_loop_filter_luma).  Same
+NEON-to-NEON substitution: H264DSPContext.h_loop_filter_luma →
+daedalus_recipe_dispatch_h264_deblock_luma_h.  The H kernel landed
+in daedalus-fourier PR #9 (CPU NEON only — no QPU shader yet).
+
+libavcodec.so ctx is no-QPU per the existing 0003-0005 / 0007
+pattern; we cannot assume Vulkan in arbitrary host processes
+(firefox-fourier RDD, mpv-fourier, etc.).
+
+Intra (bS=4) h_loop_filter_luma_intra stays on the in-tree NEON .S
+code; daedalus_h264_deblock_meta only covers the non-intra path.
+An intra-h substitution can land once daedalus-fourier exposes a
+dispatch helper (the kernel already exists internally per PR #11).
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 H.
+---
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+--- a/libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 13:09:33.694760715 +0200
++++ libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 13:09:33.715603719 +0200
+@@ -1,9 +1,10 @@
+ /*
+- * H.264 4x4 / 8x8 IDCT + luma-v deblock — daedalus-fourier substitution shims.
++ * H.264 4x4 / 8x8 IDCT + luma v/h deblock — daedalus-fourier substitution shims.
+  *
+  * Routes H264DSPContext.idct_add           → daedalus_recipe_dispatch_h264_idct4
+  *        H264DSPContext.idct8_add          → daedalus_recipe_dispatch_h264_idct8
+  *        H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
++ *        H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
+  * instead of the in-tree ff_h264_*_neon assembly.  The recipe layer
+  * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
+  * is CPU primary with QPU opportunistic — the ctx below is no-QPU,
+@@ -45,6 +46,8 @@
+ void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+                                          int alpha, int beta, int8_t *tc0);
++void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+ {
+@@ -84,3 +87,22 @@
+     daedalus_recipe_dispatch_h264_deblock_luma_v(g_dctx, pix, (size_t)stride,
+                                                  1, &meta);
+ }
++
++void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++    meta.tc0[0] = tc0[0];
++    meta.tc0[1] = tc0[1];
++    meta.tc0[2] = tc0[2];
++    meta.tc0[3] = tc0[3];
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_deblock_luma_h(g_dctx, pix, (size_t)stride,
++                                                 1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 13:09:33.695937103 +0200
++++ libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 13:09:33.715541700 +0200
+@@ -31,6 +31,8 @@
+                                          int alpha, int beta, int8_t *tc0);
+ void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                      int beta, int8_t *tc0);
++void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0);
+ void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                            int beta);
+ void ff_h264_h_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+@@ -117,7 +119,7 @@
+ 
+     if (have_neon(cpu_flags) && bit_depth == 8) {
+         c->v_loop_filter_luma   = ff_h264_v_loop_filter_luma_daedalus;
+-        c->h_loop_filter_luma   = ff_h264_h_loop_filter_luma_neon;
++        c->h_loop_filter_luma   = ff_h264_h_loop_filter_luma_daedalus;
+         c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
+         c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
+ 
+--
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0009-h264-deblock-chroma-daedalus-fourier.patch

@@ -0,0 +1,127 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: claude-noether <claude-noether@noreply.localhost>
+Date: Sun, 25 May 2026 12:00:00 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma v/h deblock through daedalus-fourier
+
+Chroma siblings of 0005 (luma_v) and 0008 (luma_h).  Same
+NEON-to-NEON pattern via the daedalus recipe layer:
+
+  H264DSPContext.v_loop_filter_chroma →
+    daedalus_recipe_dispatch_h264_deblock_chroma_v
+  H264DSPContext.h_loop_filter_chroma →
+    daedalus_recipe_dispatch_h264_deblock_chroma_h
+
+Both kernels landed in daedalus-fourier PR #10.  Recipe table
+routes AUTO to CPU NEON (no chroma QPU shaders yet), so this
+is plumbing-only and stays bit-exact against the in-tree NEON.
+
+Intra chroma (bS=4) loop filters remain on in-tree NEON;
+daedalus_h264_deblock_meta covers the non-intra (bS<4) path.
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 chroma.
+---
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+--- a/libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 13:15:45.995368233 +0200
++++ libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 13:15:46.015839177 +0200
+@@ -1,10 +1,12 @@
+ /*
+- * H.264 4x4 / 8x8 IDCT + luma v/h deblock — daedalus-fourier substitution shims.
++ * H.264 4x4 / 8x8 IDCT + luma v/h + chroma v/h deblock — daedalus-fourier substitution shims.
+  *
+  * Routes H264DSPContext.idct_add           → daedalus_recipe_dispatch_h264_idct4
+  *        H264DSPContext.idct8_add          → daedalus_recipe_dispatch_h264_idct8
+- *        H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
+- *        H264DSPContext.h_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_h
++ *        H264DSPContext.v_loop_filter_luma   → daedalus_recipe_dispatch_h264_deblock_luma_v
++ *        H264DSPContext.h_loop_filter_luma   → daedalus_recipe_dispatch_h264_deblock_luma_h
++ *        H264DSPContext.v_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_v
++ *        H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
+  * instead of the in-tree ff_h264_*_neon assembly.  The recipe layer
+  * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
+  * is CPU primary with QPU opportunistic — the ctx below is no-QPU,
+@@ -48,6 +50,10 @@
+                                          int alpha, int beta, int8_t *tc0);
+ void ff_h264_h_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
+                                          int alpha, int beta, int8_t *tc0);
++void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                           int alpha, int beta, int8_t *tc0);
++void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                           int alpha, int beta, int8_t *tc0);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+ {
+@@ -106,3 +112,41 @@
+     daedalus_recipe_dispatch_h264_deblock_luma_h(g_dctx, pix, (size_t)stride,
+                                                  1, &meta);
+ }
++
++void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                           int alpha, int beta, int8_t *tc0)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++    meta.tc0[0] = tc0[0];
++    meta.tc0[1] = tc0[1];
++    meta.tc0[2] = tc0[2];
++    meta.tc0[3] = tc0[3];
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_deblock_chroma_v(g_dctx, pix, (size_t)stride,
++                                                   1, &meta);
++}
++
++void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                           int alpha, int beta, int8_t *tc0)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++    meta.tc0[0] = tc0[0];
++    meta.tc0[1] = tc0[1];
++    meta.tc0[2] = tc0[2];
++    meta.tc0[3] = tc0[3];
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_deblock_chroma_h(g_dctx, pix, (size_t)stride,
++                                                   1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 13:15:45.996482360 +0200
++++ libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 13:15:46.025604910 +0200
+@@ -39,8 +39,12 @@
+                                            int beta);
+ void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                        int beta, int8_t *tc0);
++void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                           int alpha, int beta, int8_t *tc0);
+ void ff_h264_h_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                        int beta, int8_t *tc0);
++void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                           int alpha, int beta, int8_t *tc0);
+ void ff_h264_h_loop_filter_chroma422_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                           int beta, int8_t *tc0);
+ void ff_h264_v_loop_filter_chroma_intra_neon(uint8_t *pix, ptrdiff_t stride,
+@@ -123,11 +127,11 @@
+         c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
+         c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
+ 
+-        c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_neon;
++        c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
+         c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
+ 
+         if (chroma_format_idc <= 1) {
+-            c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_neon;
++            c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
+             c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
+             c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
+         } else {
+--
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0010-h264-deblock-luma-intra-daedalus-fourier.patch

@@ -0,0 +1,126 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: claude-noether <claude-noether@noreply.localhost>
+Date: Sun, 25 May 2026 12:30:00 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma intra deblock through daedalus-fourier
+
+Adds the bS=4 intra-strength variants of the already-substituted
+luma_v / luma_h deblock (0005, 0008).  Intra MBs and certain
+inter-MB edges (4x4 transform boundaries inside an Intra_NxN
+neighbour) force boundary strength to 4 per H.264 §8.7.2.1.
+
+  H264DSPContext.v_loop_filter_luma_intra →
+    daedalus_recipe_dispatch_h264_deblock_luma_v_intra
+  H264DSPContext.h_loop_filter_luma_intra →
+    daedalus_recipe_dispatch_h264_deblock_luma_h_intra
+
+Both kernels landed in daedalus-fourier PR #11.  Recipe table
+routes AUTO to CPU NEON (no intra QPU shaders yet) — plumbing-
+only NEON-to-NEON via daedalus, bit-exact against the in-tree
+FFmpeg NEON path.
+
+Signature differs from bS<4: no tc0 argument.  The wrapper
+passes daedalus_h264_deblock_meta with alpha/beta set; tc0[] is
+ignored by the intra dispatch (bS=4 hardcodes the strength).
+
+Chroma intra variants are deferred to a follow-up PR because the
+chroma path has a 4:2:0 / 4:2:2 split (chroma_format_idc gating)
+that needs explicit conditional substitution to avoid running
+the 4:2:0-only daedalus dispatch on 4:2:2 chroma.
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8 intra.
+---
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+--- a/libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 13:18:54.992244965 +0200
++++ libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 13:20:12.338122217 +0200
+@@ -1,5 +1,5 @@
+ /*
+- * H.264 4x4 / 8x8 IDCT + luma v/h + chroma v/h deblock — daedalus-fourier substitution shims.
++ * H.264 4x4 / 8x8 IDCT + luma v/h (inter + intra) + chroma v/h deblock — daedalus-fourier substitution shims.
+  *
+  * Routes H264DSPContext.idct_add           → daedalus_recipe_dispatch_h264_idct4
+  *        H264DSPContext.idct8_add          → daedalus_recipe_dispatch_h264_idct8
+@@ -7,6 +7,8 @@
+  *        H264DSPContext.h_loop_filter_luma   → daedalus_recipe_dispatch_h264_deblock_luma_h
+  *        H264DSPContext.v_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_v
+  *        H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
++ *        H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
++ *        H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
+  * instead of the in-tree ff_h264_*_neon assembly.  The recipe layer
+  * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
+  * is CPU primary with QPU opportunistic — the ctx below is no-QPU,
+@@ -54,6 +56,10 @@
+                                            int alpha, int beta, int8_t *tc0);
+ void ff_h264_h_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+                                            int alpha, int beta, int8_t *tc0);
++void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                int alpha, int beta);
++void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                int alpha, int beta);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+ {
+@@ -150,3 +156,34 @@
+     daedalus_recipe_dispatch_h264_deblock_chroma_h(g_dctx, pix, (size_t)stride,
+                                                    1, &meta);
+ }
++
++void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                int alpha, int beta)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++    /* tc0[] is ignored by the intra-strength dispatch (bS=4 hardcodes the strength). */
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_deblock_luma_v_intra(g_dctx, pix, (size_t)stride,
++                                                        1, &meta);
++}
++
++void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                int alpha, int beta)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_deblock_luma_h_intra(g_dctx, pix, (size_t)stride,
++                                                        1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 13:18:54.993349573 +0200
++++ libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 13:20:12.338265830 +0200
+@@ -35,8 +35,12 @@
+                                          int alpha, int beta, int8_t *tc0);
+ void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                            int beta);
++void ff_h264_v_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                int alpha, int beta);
+ void ff_h264_h_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                            int beta);
++void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                int alpha, int beta);
+ void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                        int beta, int8_t *tc0);
+ void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+@@ -124,8 +128,8 @@
+     if (have_neon(cpu_flags) && bit_depth == 8) {
+         c->v_loop_filter_luma   = ff_h264_v_loop_filter_luma_daedalus;
+         c->h_loop_filter_luma   = ff_h264_h_loop_filter_luma_daedalus;
+-        c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
+-        c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
++        c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_daedalus;
++        c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_daedalus;
+ 
+         c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
+         c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
+--
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0011-h264-chroma-dc-hadamard-daedalus-fourier.patch

@@ -0,0 +1,101 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: claude-noether <claude-noether@noreply.localhost>
+Date: Sun, 25 May 2026 13:00:00 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma DC Hadamard through daedalus-fourier
+
+Substitutes H264DSPContext.chroma_dc_dequant_idct in the
+4:2:0 / bit_depth=8 init path with a wrapper that composes
+the daedalus chroma DC Hadamard primitive (fourier PR #25)
+with qmul scaling FFmpeg does in one fused function.
+
+Bit-exact against ff_h264_chroma_dc_dequant_idct_8_c.
+Hadamard correctness gated by fourier PR #23 test suite.
+
+4:2:2 chroma stays on the in-tree 422 variant (same
+gating shape as 0009 chroma deblock substitution).
+
+Requires daedalus-fourier commit b9f9ff2 or later (PR #25
+exposing the public Hadamard symbol).  Pin bumps in PKGBUILD
+and build-deb.sh come in the same commit.
+---
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+--- a/libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 13:38:32.019491484 +0200
++++ libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 13:38:32.033821507 +0200
+@@ -1,5 +1,5 @@
+ /*
+- * H.264 4x4 / 8x8 IDCT + luma v/h (inter + intra) + chroma v/h deblock — daedalus-fourier substitution shims.
++ * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
+  *
+  * Routes H264DSPContext.idct_add           → daedalus_recipe_dispatch_h264_idct4
+  *        H264DSPContext.idct8_add          → daedalus_recipe_dispatch_h264_idct8
+@@ -9,6 +9,7 @@
+  *        H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
+  *        H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
+  *        H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
++ *        H264DSPContext.chroma_dc_dequant_idct   → daedalus_h264_chroma_dc_hadamard_2x2 + caller-side qmul
+  * instead of the in-tree ff_h264_*_neon assembly.  The recipe layer
+  * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
+  * is CPU primary with QPU opportunistic — the ctx below is no-QPU,
+@@ -60,6 +61,7 @@
+                                                 int alpha, int beta);
+ void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+                                                 int alpha, int beta);
++void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+ {
+@@ -187,3 +189,32 @@
+     daedalus_recipe_dispatch_h264_deblock_luma_h_intra(g_dctx, pix, (size_t)stride,
+                                                         1, &meta);
+ }
++
++/* Composes daedalus_h264_chroma_dc_hadamard_2x2 with the qmul scaling
++ * that FFmpeg's reference does in one fused function (h264idct_template.c
++ * ff_h264_chroma_dc_dequant_idct).
++ *
++ * The 4 DC coefficients are scattered across the per-MB coefficient
++ * buffer at offsets [r*stride + c*xStride] (stride=32, xStride=16).
++ * Extract into a contiguous int16[4], run the Hadamard, then apply
++ * the qmul scale and write back to the original positions.
++ *
++ * No daedalus ctx needed; the Hadamard is a pure stateless primitive.
++ */
++void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul)
++{
++    enum { stride = 32, xStride = 16 };
++    int16_t dc[4];
++
++    dc[0] = block[stride*0 + xStride*0];
++    dc[1] = block[stride*0 + xStride*1];
++    dc[2] = block[stride*1 + xStride*0];
++    dc[3] = block[stride*1 + xStride*1];
++
++    daedalus_h264_chroma_dc_hadamard_2x2(dc);
++
++    block[stride*0 + xStride*0] = (int16_t)((int)dc[0] * qmul >> 7);
++    block[stride*0 + xStride*1] = (int16_t)((int)dc[1] * qmul >> 7);
++    block[stride*1 + xStride*0] = (int16_t)((int)dc[2] * qmul >> 7);
++    block[stride*1 + xStride*1] = (int16_t)((int)dc[3] * qmul >> 7);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 13:38:32.020346459 +0200
++++ libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 13:38:32.033909804 +0200
+@@ -41,6 +41,7 @@
+                                            int beta);
+ void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+                                                 int alpha, int beta);
++void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
+ void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                        int beta, int8_t *tc0);
+ void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+@@ -135,6 +136,7 @@
+         c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
+ 
+         if (chroma_format_idc <= 1) {
++            c->chroma_dc_dequant_idct = ff_h264_chroma_dc_dequant_idct_daedalus;
+             c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
+             c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
+             c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
+--
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0012-h264-qpel-rest-daedalus-fourier.patch

@@ -0,0 +1,245 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: claude-noether <claude-noether@noreply.localhost>
+Date: Sun, 25 May 2026 14:00:00 +0200
+Subject: [PATCH] avcodec/aarch64/h264qpel: route remaining qpel 8x8 positions through daedalus-fourier
+
+Closes the H.264 qpel substitution.  Extends 0007 (which routed only
+mc20 put_) to ALL 15 useful positions in BOTH the put_ and avg_
+tables, skipping mc00 (integer copy / pointer-only fast path).
+
+29 substitutions total: 14 new put_ + 15 avg_.  Each is a uniform
+wrapper around daedalus_recipe_dispatch_h264_qpel_{avg_,}mcXY exposed
+by daedalus-fourier PRs #15-#20.
+
+All recipe-table entries route AUTO to CPU NEON (no QPU shaders
+for any qpel position other than mc20 yet), so this is plumbing-only
+NEON-to-NEON — bit-exact against the in-tree ff_*_h264_qpel8_*_neon
+path.
+
+16x16 qpel tables ([0][...]) stay on the in-tree NEON.  daedalus
+only exposes 8x8 today; 16x16 substitution can land once fourier
+provides those variants (likely just dispatching the 8x8 path four
+times with shifted dst/src offsets).
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc qpel buildout.
+---
+diff --git a/libavcodec/aarch64/h264_qpel_daedalus.c b/libavcodec/aarch64/h264_qpel_daedalus.c
+--- a/libavcodec/aarch64/h264_qpel_daedalus.c	2026-05-25 14:05:05.789298250 +0200
++++ libavcodec/aarch64/h264_qpel_daedalus.c	2026-05-25 14:05:05.818358374 +0200
+@@ -1,10 +1,13 @@
+ /*
+- * H.264 luma qpel mc20 (8x8, horizontal half-pel, 6-tap "put")
+- * — daedalus-fourier substitution shim.
++ * H.264 luma qpel 8x8 — daedalus-fourier substitution shims (put_ + avg_).
+  *
+- * Routes H264QpelContext.put_h264_qpel_pixels_tab[1][2] through
+- * daedalus_recipe_dispatch_h264_qpel_mc20 instead of
+- * ff_put_h264_qpel8_mc20_neon.  The recipe layer picks the substrate
++ * Routes ALL 15 useful positions in H264QpelContext's 8x8 put_ and
++ * avg_ tables through daedalus_recipe_dispatch_h264_qpel_mc{XY}
++ * (skipping mc00 which is integer copy / FFmpeg's pointer-only fast
++ * path).  Plumbing-only NEON-by-recipe — daedalus-fourier PRs #15-#20
++ * exposed each variant via the same dispatch signature, so the
++ * substitution is a uniform macro across put_/avg_ and across all
++ * 15 mc positions.  The recipe layer picks the substrate
+  * (CPU NEON for cycle 9; QPU not viable — per-block 7.6 ns vs
+  * ~250 ns QPU dispatch floor, see docs/k9_h264qpel_mc20.md).
+  *
+@@ -48,3 +51,53 @@
+     daedalus_recipe_dispatch_h264_qpel_mc20(g_dctx, dst, src, (size_t)stride,
+                                             1, &meta);
+ }
++
++
++/* All other 8x8 qpel positions follow the same dispatch shape as mc20
++ * above.  The macro collapses ~600 LOC of one-wrapper-per-variant
++ * boilerplate (29 variants total: 14 put_ + 15 avg_). */
++#define DEFINE_QPEL_WRAPPER(type, suffix, dispatch_fn)                          \
++void ff_ ## type ## _h264_qpel8_ ## suffix ## _daedalus(uint8_t *dst,           \
++    const uint8_t *src, ptrdiff_t stride);                                      \
++void ff_ ## type ## _h264_qpel8_ ## suffix ## _daedalus(uint8_t *dst,           \
++    const uint8_t *src, ptrdiff_t stride)                                       \
++{                                                                               \
++    static const daedalus_h264_qpel_meta meta = { .dst_off = 0, .src_off = 0 }; \
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);                         \
++    dispatch_fn(g_dctx, dst, src, (size_t)stride, 1, &meta);                    \
++}
++
++/* put_ variants (mc20 stays on the explicit definition above). */
++DEFINE_QPEL_WRAPPER(put, mc10, daedalus_recipe_dispatch_h264_qpel_mc10)
++DEFINE_QPEL_WRAPPER(put, mc30, daedalus_recipe_dispatch_h264_qpel_mc30)
++DEFINE_QPEL_WRAPPER(put, mc01, daedalus_recipe_dispatch_h264_qpel_mc01)
++DEFINE_QPEL_WRAPPER(put, mc11, daedalus_recipe_dispatch_h264_qpel_mc11)
++DEFINE_QPEL_WRAPPER(put, mc21, daedalus_recipe_dispatch_h264_qpel_mc21)
++DEFINE_QPEL_WRAPPER(put, mc31, daedalus_recipe_dispatch_h264_qpel_mc31)
++DEFINE_QPEL_WRAPPER(put, mc02, daedalus_recipe_dispatch_h264_qpel_mc02)
++DEFINE_QPEL_WRAPPER(put, mc12, daedalus_recipe_dispatch_h264_qpel_mc12)
++DEFINE_QPEL_WRAPPER(put, mc22, daedalus_recipe_dispatch_h264_qpel_mc22)
++DEFINE_QPEL_WRAPPER(put, mc32, daedalus_recipe_dispatch_h264_qpel_mc32)
++DEFINE_QPEL_WRAPPER(put, mc03, daedalus_recipe_dispatch_h264_qpel_mc03)
++DEFINE_QPEL_WRAPPER(put, mc13, daedalus_recipe_dispatch_h264_qpel_mc13)
++DEFINE_QPEL_WRAPPER(put, mc23, daedalus_recipe_dispatch_h264_qpel_mc23)
++DEFINE_QPEL_WRAPPER(put, mc33, daedalus_recipe_dispatch_h264_qpel_mc33)
++
++/* avg_ variants — all 15 useful positions. */
++DEFINE_QPEL_WRAPPER(avg, mc10, daedalus_recipe_dispatch_h264_qpel_avg_mc10)
++DEFINE_QPEL_WRAPPER(avg, mc20, daedalus_recipe_dispatch_h264_qpel_avg_mc20)
++DEFINE_QPEL_WRAPPER(avg, mc30, daedalus_recipe_dispatch_h264_qpel_avg_mc30)
++DEFINE_QPEL_WRAPPER(avg, mc01, daedalus_recipe_dispatch_h264_qpel_avg_mc01)
++DEFINE_QPEL_WRAPPER(avg, mc11, daedalus_recipe_dispatch_h264_qpel_avg_mc11)
++DEFINE_QPEL_WRAPPER(avg, mc21, daedalus_recipe_dispatch_h264_qpel_avg_mc21)
++DEFINE_QPEL_WRAPPER(avg, mc31, daedalus_recipe_dispatch_h264_qpel_avg_mc31)
++DEFINE_QPEL_WRAPPER(avg, mc02, daedalus_recipe_dispatch_h264_qpel_avg_mc02)
++DEFINE_QPEL_WRAPPER(avg, mc12, daedalus_recipe_dispatch_h264_qpel_avg_mc12)
++DEFINE_QPEL_WRAPPER(avg, mc22, daedalus_recipe_dispatch_h264_qpel_avg_mc22)
++DEFINE_QPEL_WRAPPER(avg, mc32, daedalus_recipe_dispatch_h264_qpel_avg_mc32)
++DEFINE_QPEL_WRAPPER(avg, mc03, daedalus_recipe_dispatch_h264_qpel_avg_mc03)
++DEFINE_QPEL_WRAPPER(avg, mc13, daedalus_recipe_dispatch_h264_qpel_avg_mc13)
++DEFINE_QPEL_WRAPPER(avg, mc23, daedalus_recipe_dispatch_h264_qpel_avg_mc23)
++DEFINE_QPEL_WRAPPER(avg, mc33, daedalus_recipe_dispatch_h264_qpel_avg_mc33)
++
++#undef DEFINE_QPEL_WRAPPER
+diff --git a/libavcodec/aarch64/h264qpel_init_aarch64.c b/libavcodec/aarch64/h264qpel_init_aarch64.c
+--- a/libavcodec/aarch64/h264qpel_init_aarch64.c	2026-05-25 14:05:05.790403989 +0200
++++ libavcodec/aarch64/h264qpel_init_aarch64.c	2026-05-25 14:05:05.819136071 +0200
+@@ -50,6 +50,64 @@
+ void ff_put_h264_qpel8_mc30_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
+                                      ptrdiff_t stride);
++void ff_put_h264_qpel8_mc10_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc30_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc01_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc11_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc21_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc31_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc02_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc12_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc22_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc32_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc03_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc13_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc23_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_put_h264_qpel8_mc33_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc10_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc20_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc30_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc01_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc11_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc21_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc31_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc02_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc12_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc22_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc32_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc03_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc13_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc23_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
++void ff_avg_h264_qpel8_mc33_daedalus(uint8_t *dst, const uint8_t *src,
++                                  ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc01_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc11_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+ void ff_put_h264_qpel8_mc21_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);
+@@ -164,21 +222,21 @@
+         c->put_h264_qpel_pixels_tab[0][15] = ff_put_h264_qpel16_mc33_neon;
+ 
+         c->put_h264_qpel_pixels_tab[1][ 0] = ff_put_h264_qpel8_mc00_neon;
+-        c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_neon;
++        c->put_h264_qpel_pixels_tab[1][ 1] = ff_put_h264_qpel8_mc10_daedalus;
+         c->put_h264_qpel_pixels_tab[1][ 2] = ff_put_h264_qpel8_mc20_daedalus;
+-        c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_neon;
+-        c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_neon;
+-        c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_neon;
+-        c->put_h264_qpel_pixels_tab[1][ 6] = ff_put_h264_qpel8_mc21_neon;
+-        c->put_h264_qpel_pixels_tab[1][ 7] = ff_put_h264_qpel8_mc31_neon;
+-        c->put_h264_qpel_pixels_tab[1][ 8] = ff_put_h264_qpel8_mc02_neon;
+-        c->put_h264_qpel_pixels_tab[1][ 9] = ff_put_h264_qpel8_mc12_neon;
+-        c->put_h264_qpel_pixels_tab[1][10] = ff_put_h264_qpel8_mc22_neon;
+-        c->put_h264_qpel_pixels_tab[1][11] = ff_put_h264_qpel8_mc32_neon;
+-        c->put_h264_qpel_pixels_tab[1][12] = ff_put_h264_qpel8_mc03_neon;
+-        c->put_h264_qpel_pixels_tab[1][13] = ff_put_h264_qpel8_mc13_neon;
+-        c->put_h264_qpel_pixels_tab[1][14] = ff_put_h264_qpel8_mc23_neon;
+-        c->put_h264_qpel_pixels_tab[1][15] = ff_put_h264_qpel8_mc33_neon;
++        c->put_h264_qpel_pixels_tab[1][ 3] = ff_put_h264_qpel8_mc30_daedalus;
++        c->put_h264_qpel_pixels_tab[1][ 4] = ff_put_h264_qpel8_mc01_daedalus;
++        c->put_h264_qpel_pixels_tab[1][ 5] = ff_put_h264_qpel8_mc11_daedalus;
++        c->put_h264_qpel_pixels_tab[1][ 6] = ff_put_h264_qpel8_mc21_daedalus;
++        c->put_h264_qpel_pixels_tab[1][ 7] = ff_put_h264_qpel8_mc31_daedalus;
++        c->put_h264_qpel_pixels_tab[1][ 8] = ff_put_h264_qpel8_mc02_daedalus;
++        c->put_h264_qpel_pixels_tab[1][ 9] = ff_put_h264_qpel8_mc12_daedalus;
++        c->put_h264_qpel_pixels_tab[1][10] = ff_put_h264_qpel8_mc22_daedalus;
++        c->put_h264_qpel_pixels_tab[1][11] = ff_put_h264_qpel8_mc32_daedalus;
++        c->put_h264_qpel_pixels_tab[1][12] = ff_put_h264_qpel8_mc03_daedalus;
++        c->put_h264_qpel_pixels_tab[1][13] = ff_put_h264_qpel8_mc13_daedalus;
++        c->put_h264_qpel_pixels_tab[1][14] = ff_put_h264_qpel8_mc23_daedalus;
++        c->put_h264_qpel_pixels_tab[1][15] = ff_put_h264_qpel8_mc33_daedalus;
+ 
+         c->avg_h264_qpel_pixels_tab[0][ 0] = ff_avg_h264_qpel16_mc00_neon;
+         c->avg_h264_qpel_pixels_tab[0][ 1] = ff_avg_h264_qpel16_mc10_neon;
+@@ -198,21 +256,21 @@
+         c->avg_h264_qpel_pixels_tab[0][15] = ff_avg_h264_qpel16_mc33_neon;
+ 
+         c->avg_h264_qpel_pixels_tab[1][ 0] = ff_avg_h264_qpel8_mc00_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 1] = ff_avg_h264_qpel8_mc10_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 2] = ff_avg_h264_qpel8_mc20_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 3] = ff_avg_h264_qpel8_mc30_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 4] = ff_avg_h264_qpel8_mc01_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 5] = ff_avg_h264_qpel8_mc11_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 6] = ff_avg_h264_qpel8_mc21_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 7] = ff_avg_h264_qpel8_mc31_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 8] = ff_avg_h264_qpel8_mc02_neon;
+-        c->avg_h264_qpel_pixels_tab[1][ 9] = ff_avg_h264_qpel8_mc12_neon;
+-        c->avg_h264_qpel_pixels_tab[1][10] = ff_avg_h264_qpel8_mc22_neon;
+-        c->avg_h264_qpel_pixels_tab[1][11] = ff_avg_h264_qpel8_mc32_neon;
+-        c->avg_h264_qpel_pixels_tab[1][12] = ff_avg_h264_qpel8_mc03_neon;
+-        c->avg_h264_qpel_pixels_tab[1][13] = ff_avg_h264_qpel8_mc13_neon;
+-        c->avg_h264_qpel_pixels_tab[1][14] = ff_avg_h264_qpel8_mc23_neon;
+-        c->avg_h264_qpel_pixels_tab[1][15] = ff_avg_h264_qpel8_mc33_neon;
++        c->avg_h264_qpel_pixels_tab[1][ 1] = ff_avg_h264_qpel8_mc10_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][ 2] = ff_avg_h264_qpel8_mc20_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][ 3] = ff_avg_h264_qpel8_mc30_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][ 4] = ff_avg_h264_qpel8_mc01_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][ 5] = ff_avg_h264_qpel8_mc11_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][ 6] = ff_avg_h264_qpel8_mc21_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][ 7] = ff_avg_h264_qpel8_mc31_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][ 8] = ff_avg_h264_qpel8_mc02_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][ 9] = ff_avg_h264_qpel8_mc12_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][10] = ff_avg_h264_qpel8_mc22_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][11] = ff_avg_h264_qpel8_mc32_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][12] = ff_avg_h264_qpel8_mc03_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][13] = ff_avg_h264_qpel8_mc13_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][14] = ff_avg_h264_qpel8_mc23_daedalus;
++        c->avg_h264_qpel_pixels_tab[1][15] = ff_avg_h264_qpel8_mc33_daedalus;
+     } else if (have_neon(cpu_flags) && bit_depth == 10) {
+         c->put_h264_qpel_pixels_tab[0][ 1] = ff_put_h264_qpel16_mc10_neon_10;
+         c->put_h264_qpel_pixels_tab[0][ 2] = ff_put_h264_qpel16_mc20_neon_10;
+--
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/0013-h264-deblock-chroma-intra-daedalus-fourier.patch

@@ -0,0 +1,120 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: claude-noether <claude-noether@noreply.localhost>
+Date: Sun, 25 May 2026 14:30:00 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 chroma intra deblock (4:2:0) through daedalus-fourier
+
+Substitutes c->v_loop_filter_chroma_intra and c->h_loop_filter_chroma_intra
+with daedalus wrappers in the bit_depth=8 / chroma_format_idc<=1 (4:2:0)
+branch.  4:2:2 stays on the in-tree NEON path (the daedalus chroma intra
+dispatch is 4:2:0-only).
+
+The fourier dispatches were exposed in PR #11 (DEFINE_INTRA_DISPATCH
+macro generates the public daedalus_dispatch_h264_deblock_chroma_*_intra
+symbols + recipe wrappers).
+
+Re-architects the chroma init: v_loop_filter_chroma_intra was previously
+assigned unconditionally to the NEON variant (which works for both 4:2:0
+and 4:2:2).  We now assign it INSIDE both branches of the chroma_format_idc
+conditional, with the 4:2:0 branch picking daedalus and the 4:2:2 branch
+keeping NEON.  No regression for 4:2:2 streams.
+
+Same NEON-to-NEON via recipe shape as 0010 luma intra.
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc chroma intra.
+---
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+--- a/libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 14:21:08.267156263 +0200
++++ libavcodec/aarch64/h264_idct_daedalus.c	2026-05-25 14:21:08.287745931 +0200
+@@ -1,5 +1,5 @@
+ /*
+- * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
++ * H.264 4x4 / 8x8 IDCT + luma v/h (inter+intra) + chroma v/h (inter+intra) deblock + chroma DC Hadamard — daedalus-fourier substitution shims.
+  *
+  * Routes H264DSPContext.idct_add           → daedalus_recipe_dispatch_h264_idct4
+  *        H264DSPContext.idct8_add          → daedalus_recipe_dispatch_h264_idct8
+@@ -9,6 +9,8 @@
+  *        H264DSPContext.h_loop_filter_chroma → daedalus_recipe_dispatch_h264_deblock_chroma_h
+  *        H264DSPContext.v_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_v_intra
+  *        H264DSPContext.h_loop_filter_luma_intra → daedalus_recipe_dispatch_h264_deblock_luma_h_intra
++ *        H264DSPContext.v_loop_filter_chroma_intra → daedalus_recipe_dispatch_h264_deblock_chroma_v_intra
++ *        H264DSPContext.h_loop_filter_chroma_intra → daedalus_recipe_dispatch_h264_deblock_chroma_h_intra
+  *        H264DSPContext.chroma_dc_dequant_idct   → daedalus_h264_chroma_dc_hadamard_2x2 + caller-side qmul
+  * instead of the in-tree ff_h264_*_neon assembly.  The recipe layer
+  * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
+@@ -61,6 +63,10 @@
+                                                 int alpha, int beta);
+ void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+                                                 int alpha, int beta);
++void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                 int alpha, int beta);
++void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                 int alpha, int beta);
+ void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+@@ -218,3 +224,30 @@
+     block[stride*1 + xStride*0] = (int16_t)((int)dc[2] * qmul >> 7);
+     block[stride*1 + xStride*1] = (int16_t)((int)dc[3] * qmul >> 7);
+ }
++
++void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                 int alpha, int beta)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++    /* tc0[] unused for intra (bS=4 hardcodes the strength). */
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++    daedalus_recipe_dispatch_h264_deblock_chroma_v_intra(g_dctx, pix, (size_t)stride,
++                                                          1, &meta);
++}
++
++void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                 int alpha, int beta)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++    daedalus_recipe_dispatch_h264_deblock_chroma_h_intra(g_dctx, pix, (size_t)stride,
++                                                          1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 14:21:08.268311057 +0200
++++ libavcodec/aarch64/h264dsp_init_aarch64.c	2026-05-25 14:21:08.287886563 +0200
+@@ -42,6 +42,10 @@
+ void ff_h264_h_loop_filter_luma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
+                                                 int alpha, int beta);
+ void ff_h264_chroma_dc_dequant_idct_daedalus(int16_t *block, int qmul);
++void ff_h264_v_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                 int alpha, int beta);
++void ff_h264_h_loop_filter_chroma_intra_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                                 int alpha, int beta);
+ void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                        int beta, int8_t *tc0);
+ void ff_h264_v_loop_filter_chroma_daedalus(uint8_t *pix, ptrdiff_t stride,
+@@ -133,14 +137,15 @@
+         c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_daedalus;
+ 
+         c->v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_daedalus;
+-        c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
+ 
+         if (chroma_format_idc <= 1) {
+             c->chroma_dc_dequant_idct = ff_h264_chroma_dc_dequant_idct_daedalus;
++            c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_daedalus;
+             c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_daedalus;
+-            c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_neon;
++            c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma_intra_daedalus;
+             c->h_loop_filter_chroma_mbaff_intra = ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
+         } else {
++            c->v_loop_filter_chroma_intra = ff_h264_v_loop_filter_chroma_intra_neon;
+             c->h_loop_filter_chroma = ff_h264_h_loop_filter_chroma422_neon;
+             c->h_loop_filter_chroma_mbaff = ff_h264_h_loop_filter_chroma_neon;
+             c->h_loop_filter_chroma_intra = ff_h264_h_loop_filter_chroma422_intra_neon;
+--
+2.47.3
+

arch/ffmpeg-v4l2-request-fourier/PKGBUILD

@@ -24,8 +24,13 @@ _srcname=FFmpeg
 _version='8.1'
 _commit='b57fbbe50c9b2656fad86a1a7eeabfd2b2a50935'  # v4l2-request-n8.1 tip 2026-04-24
 pkgver=8.1.r123329.b57fbbe
-pkgrel=4
+pkgrel=10  # pkgrel=10 — H.264 luma qpel mc20 daedalus-fourier substitution (cycle 9, 2026-05-23)
 epoch=2
+
+# daedalus-fourier pin.  209a421 = PR #2 merge (Phase 8c — public API
+# gains daedalus_recipe_dispatch_h264_qpel_mc20 + DAEDALUS_KERNEL_H264_QPEL_MC20).
+# Cycle 9 closes the libavcodec.so substitution arc started at cycle 6.
+_daedalus_fourier_commit='b9f9ff2a89c068aea54dcb52b543afddad28311e'  # PR #25 — public chroma DC Hadamard symbol
 pkgdesc='FFmpeg with V4L2 Request API hwaccel (Rockchip / Allwinner stateless decode)'
 arch=('aarch64')
 url='https://github.com/Kwiboo/FFmpeg'
@@ -34,6 +39,7 @@ depends=(
   alsa-lib
   bzip2
   fontconfig
+  vulkan-icd-loader
   fribidi
   gmp
   gnutls
@@ -59,10 +65,13 @@ depends=(
   zlib
 )
 makedepends=(
+  cmake
   git
   linux-api-headers
   mesa
   nasm
+  ninja
+  vulkan-headers
 )
 provides=(
   libavcodec.so
@@ -78,8 +87,21 @@ provides=(
 conflicts=(ffmpeg)
 replaces=(ffmpeg ffmpeg-v4l2-request-git)
 source=("git+https://github.com/Kwiboo/FFmpeg.git#commit=${_commit}"
-        '0001-libudev-bypass-fallback.patch')
-sha256sums=('SKIP' 'SKIP')
+        "daedalus-fourier-${_daedalus_fourier_commit}.tar.gz::https://git.reauktion.de/marfrit/daedalus-fourier/archive/${_daedalus_fourier_commit}.tar.gz"
+        '0001-libudev-bypass-fallback.patch'
+        '0002-nv15-to-p010-unpack.patch'
+        '0003-h264-idct4-daedalus-fourier.patch'
+        '0004-h264-idct8-daedalus-fourier.patch'
+        '0005-h264-deblock-luma-v-daedalus-fourier.patch'
+        '0006-h264-restore-low-delay.patch'
+        '0007-h264-qpel-mc20-daedalus-fourier.patch'
+        '0008-h264-deblock-luma-h-daedalus-fourier.patch'
+        '0009-h264-deblock-chroma-daedalus-fourier.patch'
+        '0010-h264-deblock-luma-intra-daedalus-fourier.patch'
+        '0011-h264-chroma-dc-hadamard-daedalus-fourier.patch'
+        '0012-h264-qpel-rest-daedalus-fourier.patch'
+        '0013-h264-deblock-chroma-intra-daedalus-fourier.patch')
+sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
 
 pkgver() {
   cd "${_srcname}"
@@ -91,9 +113,36 @@ pkgver() {
 prepare() {
   cd "${_srcname}"
   patch -Np1 -i "${srcdir}/0001-libudev-bypass-fallback.patch"
+  patch -Np1 -i "${srcdir}/0002-nv15-to-p010-unpack.patch"
+  patch -Np1 -i "${srcdir}/0003-h264-idct4-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0004-h264-idct8-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0005-h264-deblock-luma-v-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0006-h264-restore-low-delay.patch"
+  patch -Np1 -i "${srcdir}/0007-h264-qpel-mc20-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0008-h264-deblock-luma-h-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0009-h264-deblock-chroma-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0010-h264-deblock-luma-intra-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0011-h264-chroma-dc-hadamard-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0012-h264-qpel-rest-daedalus-fourier.patch"
+  patch -Np1 -i "${srcdir}/0013-h264-deblock-chroma-intra-daedalus-fourier.patch"
 }
 
 build() {
+  # --- daedalus-fourier: build static .a with PIC, install to a
+  # per-build prefix; libavcodec.so links it into the shared object so
+  # H264DSPContext.idct_add (and follow-up kernels) dispatch through
+  # the daedalus recipe layer instead of the in-tree NEON .S code. ---
+  local _fourier_prefix="${srcdir}/fourier-prefix"
+  mkdir -p "${_fourier_prefix}"
+  pushd "${srcdir}"/daedalus-fourier >/dev/null
+  cmake -B build -G Ninja \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+    -DCMAKE_INSTALL_PREFIX="${_fourier_prefix}"
+  cmake --build build --target daedalus_core
+  cmake --install build
+  popd >/dev/null
+
   cd "${_srcname}"
 
   # FFmpeg's configure resolves the compiler via `which` and bakes the
@@ -145,6 +194,9 @@ build() {
     --enable-libx265 \
     --enable-libwebp \
     \
+    --extra-cflags="-I${_fourier_prefix}/include" \
+    --extra-ldflags="-L${_fourier_prefix}/lib" \
+    --extra-libs="-ldaedalus_core -lvulkan -lpthread" \
     --host-cflags='-fPIC'
 
   make

arch/firefox-fourier/0003-ffmpegvideo-v4l2-request-route.patch

@@ -18,27 +18,30 @@ This patch adds a sibling init path, `InitV4L2RequestDecoder`, that:
   * looks up the codec via two complementary mechanisms libavcodec
     uses for v4l2_request:
       - **named codec** (`h264_v4l2request`, `vp8_v4l2request`, etc.):
-        the legacy AVCodec-per-hwaccel registration. ALARM, Debian,
-        and most distros building with --enable-v4l2-request expose
-        this (avcodec_find_decoder_by_name lookup).
-      - **generic codec + AV_HWDEVICE_TYPE_DRM** in `hw_configs`:
-        the modern hwaccel registration on some upstream-only ffmpeg
-        builds.
+        the legacy AVCodec-per-hwaccel registration.
+      - **generic codec + hw_configs walk**: the modern hwaccel
+        registration. Accepts EITHER AV_HWDEVICE_TYPE_DRM (legacy
+        ffmpeg-v4l2-request-fork output prior to FFmpeg 7.1) OR
+        AV_HWDEVICE_TYPE_V4L2REQUEST (FFmpeg 7.1+ dedicated enum,
+        value 13 on Kwiboo's no-AMF tree, 14 on upstream-AMF tree).
+        Mozilla's bundled libavutil headers may not have the V4L2REQUEST
+        enumerator, so the test is on the integer value via `(int)cast`.
     Probes named-codec first (explicit, portable) and falls back to
-    walking the generic codec's `hw_configs` for the DRM device type;
-  * creates an `AV_HWDEVICE_TYPE_DRM` hwdevice context bound to
-    `/dev/dri/renderD128` via the new `av_hwdevice_ctx_create` wrapper
-    (patch 2/4) and attaches it to the codec context;
+    walking the generic codec's `hw_configs` for either device type;
+  * creates an hwdevice context bound to `/dev/dri/renderD128`. Uses
+    integer 13 (V4L2REQUEST as defined by Kwiboo's v4l2-request-n7.1.3
+    tree, what our libavcodec61-fourier emits) cast to enum
+    AVHWDeviceType for the av_hwdevice_ctx_create call;
   * reuses the existing `ChooseV4L2PixelFormat` get-format callback
     (already returns `AV_PIX_FMT_DRM_PRIME`) and the existing
     `apply_cropping = 0` constraint.
 
 `InitV4L2RequestDecoder` is invoked **before** `InitV4L2Decoder` in
 `InitHWDecoderIfAllowed`. On Rockchip mainline it succeeds via either
-mechanism (ALARM uses the named codec). On Pi4 / Mediatek /
-vendor-MPP-stateful boards neither mechanism is registered for the
-codec, the function bails out, and the existing stateful
-`InitV4L2Decoder` runs as before. No regression of stateful boards.
+mechanism. On Pi4 / Mediatek / vendor-MPP-stateful boards neither
+mechanism is registered for the codec, the function bails out, and the
+existing stateful `InitV4L2Decoder` runs as before. No regression of
+stateful boards.
 
 `mDRMDeviceContext` is unconditionally `av_buffer_unref`'d in
 `ProcessShutdown` (no-op when null). Gated behind
@@ -46,9 +49,8 @@ codec, the function bails out, and the existing stateful
 
 Bug 1969297.
 
-diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h
---- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h	2026-03-18 19:22:14.000000000 +0000
-+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h	2026-04-27 20:43:39.347992674 +0000
+--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h	2026-05-21 04:57:59.570946601 +0000
++++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h	2026-05-21 04:57:59.876488776 +0000
 @@ -225,7 +225,12 @@
    bool IsLinuxHDR() const;
    MediaResult InitVAAPIDecoder();
@@ -73,9 +75,8 @@ diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.h b/dom/media/platfor
    // If video overlay is used we want to upload SW decoded frames to
    // DMABuf and present it as a external texture to rendering pipeline.
    bool mUploadSWDecodeToDMABuf = false;
-diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp
---- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp	2026-04-27 16:09:10.000000000 +0200
-+++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp	2026-04-29 00:10:00.098884335 +0200
+--- a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp	2026-05-21 04:57:59.566685221 +0000
++++ b/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp	2026-05-21 04:58:00.136004159 +0000
 @@ -403,6 +403,129 @@
    return NS_OK;
  }
@@ -90,7 +91,7 @@ diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp b/dom/media/platf
 +  }
 +  const char* drmDevice = "/dev/dri/renderD128";
 +  if (mLib->av_hwdevice_ctx_create(&mDRMDeviceContext,
-+                                   AV_HWDEVICE_TYPE_DRM, drmDevice,
++                                   (enum AVHWDeviceType)13, drmDevice,
 +                                   nullptr, 0) < 0) {
 +    FFMPEG_LOG("  av_hwdevice_ctx_create(DRM, %s) failed", drmDevice);
 +    return false;
@@ -143,7 +144,7 @@ diff --git a/dom/media/platforms/ffmpeg/FFmpegVideoDecoder.cpp b/dom/media/platf
 +      for (int i = 0;; i++) {
 +        const AVCodecHWConfig* cfg = mLib->avcodec_get_hw_config(generic, i);
 +        if (!cfg) break;
-+        if (cfg->device_type == AV_HWDEVICE_TYPE_DRM) {
++        if (cfg->device_type == AV_HWDEVICE_TYPE_DRM || (int)cfg->device_type == 13 || (int)cfg->device_type == 14) {
 +          codec = generic;
 +          FFMPEG_LOG("  using generic codec %s with DRM hwaccel", codec->name);
 +          break;

arch/firefox-fourier/PKGBUILD

@@ -13,7 +13,7 @@
 
 pkgname=firefox-fourier
 pkgver=150.0.1
-pkgrel=1
+pkgrel=7
 pkgdesc='Firefox with V4L2 stateless HW video decode unlocked for mainline Linux Rockchip'
 arch=('aarch64' 'x86_64')
 url='https://www.mozilla.org/firefox'
@@ -87,8 +87,15 @@ source=(
   '0002-libwrapper-hwdevice-ctx-create.patch'
   '0003-ffmpegvideo-v4l2-request-route.patch'
   '0004-prefs-v4l2-request.patch'
+  '0005-rdd-sandbox-v4l2-media-ctl.patch'
+  # Vendor-default prefs that gate the patched VAAPI path on RK3399 —
+  # widget.dmabuf.force-enabled etc. See marfrit-packages#8 for evidence.
+  'rockchip-fourier-defaults.js'
+  # Plasma/GNOME start-menu entry — categorises under Internet, picks the
+  # 128px firefox icon shipped under /usr/lib/firefox-fourier/browser/.
+  'firefox-fourier.desktop'
 )
-sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
+sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
 
 prepare() {
   cd "${srcdir}/firefox-${pkgver}"
@@ -103,6 +110,7 @@ prepare() {
   patch -Np1 -i "${srcdir}/0002-libwrapper-hwdevice-ctx-create.patch"
   patch -Np1 -i "${srcdir}/0003-ffmpegvideo-v4l2-request-route.patch"
   patch -Np1 -i "${srcdir}/0004-prefs-v4l2-request.patch"
+  patch -Np1 -i "${srcdir}/0005-rdd-sandbox-v4l2-media-ctl.patch"
 
   cp "${srcdir}/mozconfig" .mozconfig
 }
@@ -160,4 +168,26 @@ export MOZ_X11_EGL="${MOZ_X11_EGL:-1}"
 exec /usr/lib/firefox-fourier/firefox-fourier "$@"
 LAUNCHER
   chmod 0755 "${pkgdir}/usr/bin/firefox-fourier"
+
+  # Vendor-default prefs (RK3399 HW-decode unlock) — closes #8.
+  # Lower precedence than user prefs / about:config; loaded by Firefox
+  # at startup from the package install dir. The 0004 patch covers
+  # media.ffmpeg.v4l2-request.enabled; this file covers the three
+  # additional prefs that gate the path to the patched code.
+  # Vendor-prefs install path: /usr/lib/firefox-fourier/defaults/preferences/
+  # (Mozilla's canonical scan dir for third-party default-pref drops.) The
+  # browser/defaults/preferences/ alternative looked promising but is NOT a
+  # vendor-prefs scan location in Firefox 150 — empirically confirmed on
+  # fresnel: file shipped there, VAAPI never engaged. Same file under
+  # defaults/preferences/ → MOZ_LOG showed `Requesting pixel format
+  # VAAPI_VLD` + dmabuf surfaces locking end-to-end.
+  install -Dm644 "${srcdir}/rockchip-fourier-defaults.js" \
+    "${pkgdir}/usr/lib/firefox-fourier/defaults/preferences/rockchip-fourier-defaults.js"
+
+  # Desktop entry — fileless install would leave the package without a
+  # start-menu entry (stock firefox.desktop disappears when our `provides`
+  # replaces stock firefox). Plasma & GNOME pick this up via the
+  # `Categories=Network;WebBrowser;` line → "Internet" submenu.
+  install -Dm644 "${srcdir}/firefox-fourier.desktop" \
+    "${pkgdir}/usr/share/applications/firefox-fourier.desktop"
 }

arch/firefox-fourier/arch-0004-Bug-2023597-Use-wasm32-wasip1-target-for-clang-22.1-.patch

@@ -0,0 +1,28 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Alex Hochheiden <ahochheiden@mozilla.com>
+Date: Wed, 1 Apr 2026 18:11:37 +0000
+Subject: [PATCH] Bug 2023597 - Use `wasm32-wasip1` target for clang >= 22.1
+ r=firefox-build-system-reviewers,sergesanspaille
+
+https://github.com/llvm/llvm-project/pull/165345
+https://releases.llvm.org/22.1.0/tools/clang/docs/ReleaseNotes.html
+
+Differential Revision: https://phabricator.services.mozilla.com/D291023
+---
+ build/moz.configure/toolchain.configure | 3 +++
+ 1 file changed, 3 insertions(+)
+
+diff --git a/build/moz.configure/toolchain.configure b/build/moz.configure/toolchain.configure
+index a37ed610cc43..c7d0c8bdf75c 100644
+--- a/build/moz.configure/toolchain.configure
++++ b/build/moz.configure/toolchain.configure
+@@ -695,6 +695,9 @@ def check_compiler(configure_cache, compiler, language, target, android_version)
+         # This makes clang define __ANDROID_API__ and use versioned library
+         # directories from the NDK.
+         toolchain = "%s%d" % (target.toolchain, android_version)
++    elif target.kernel == "WASI" and info.type == "clang" and info.version >= Version("22.1"):
++        # The wasm32-wasi target was renamed to wasm32-wasip1 in LLVM 22.1.
++        toolchain = "wasm32-wasip1"
+     else:
+         toolchain = target.toolchain
+ 

arch/firefox-fourier/firefox-fourier.desktop

@@ -0,0 +1,22 @@
+[Desktop Entry]
+Version=1.0
+Type=Application
+Name=Firefox (Fourier — V4L2 HW decode)
+GenericName=Web Browser
+Comment=Browse the Web with V4L2 stateless HW video decode (RK3588/RK3566)
+Exec=/usr/bin/firefox-fourier %u
+Icon=/usr/lib/firefox-fourier/browser/chrome/icons/default/default128.png
+Terminal=false
+StartupNotify=true
+StartupWMClass=firefox-fourier
+Categories=Network;WebBrowser;
+MimeType=text/html;text/xml;application/xhtml+xml;application/vnd.mozilla.xul+xml;application/rss+xml;application/rdf+xml;image/gif;image/jpeg;image/png;image/svg+xml;image/webp;image/avif;application/json;application/pdf;audio/flac;audio/ogg;audio/webm;video/ogg;video/webm;x-scheme-handler/http;x-scheme-handler/https;x-scheme-handler/chrome;x-scheme-handler/mailto;
+Actions=new-window;new-private-window;
+
+[Desktop Action new-window]
+Name=Open a New Window
+Exec=/usr/bin/firefox-fourier --new-window %u
+
+[Desktop Action new-private-window]
+Name=Open a New Private Window
+Exec=/usr/bin/firefox-fourier --private-window %u

arch/firefox-fourier/mozconfig

@@ -34,3 +34,20 @@ ac_add_options --with-distribution-id=de.reauktion.fourier
 
 # Reduce build memory pressure on aarch64 — parallel link is heavy.
 mk_add_options MOZ_PARALLEL_BUILD=8
+
+# Explicit distcc routing. fermi's makepkg.conf has
+# BUILDENV=(distcc ...) which auto-prepends /usr/lib/distcc/bin to
+# $PATH, but mach's configure picks up CC/CXX from the env directly
+# and the distcc wrappers won't fire unless we set them. Mirrors the
+# ffmpeg-v4l2-request-fourier pattern.
+#
+# Note: only the C/C++ portion of the build distributes; rustc and
+# the host-only build steps stay local. Empirically that's still a
+# 30-40% wall-clock win on a 4-worker pool. DISTCC_HOSTS comes from
+# makepkg.conf ('+zeroconf' by default — Avahi-discovers tesla,
+# dcc1, dcc2, ampere).
+if [[ ":${PATH}:" == *":/usr/lib/distcc/bin:"* ]]; then
+    export CC="distcc gcc"
+    export CXX="distcc g++"
+fi
+

arch/firefox-fourier/rockchip-fourier-defaults.js

@@ -0,0 +1,19 @@
+// firefox-fourier — RK3399 V4L2-stateless HW-decode default prefs.
+//
+// The patch series (0001..0004) builds the VAAPI / V4L2-request routing
+// path through libavcodec, but the resulting code path is gated by three
+// other prefs that are 'false' upstream because the relevant probes don't
+// fire on panfrost EGL or trip the Intel-tuned cost heuristic. Without
+// these, firefox-fourier silently SW-decodes on a fresh profile despite
+// having all the unlock patches applied.
+//
+// Filed via marfrit/marfrit-packages#8 — see that issue for MOZ_LOG
+// evidence on fresnel (Pinebook Pro / RK3399).
+//
+// These are *vendor* defaults: lower precedence than user.js and
+// about:config user prefs. Power users who want to disable HW decode for
+// debugging can flip them in user prefs without touching this file.
+
+pref("widget.dmabuf.force-enabled",                  true);
+pref("media.hardware-video-decoding.force-enabled",  true);
+pref("media.ffvpx-hw.enabled",                       true);

arch/kwin-fourier/PKGBUILD

@@ -2,7 +2,7 @@
 # Upstream maintainers: Felix Yan, Antonio Rojas
 # Contributor: Andrea Scarpino <andrea@archlinux.org>
 #
-# kwin-fourier — KWin 6.6.4 with the V4L2-stateless implicit-sync
+# kwin-fourier — KWin 6.6.5 with the V4L2-stateless implicit-sync
 # transaction wait bypass. Hypothesis: KWin's
 # `Transaction::watchDmaBuf` calls DMA_BUF_IOCTL_EXPORT_SYNC_FILE on
 # every plane of every imported dmabuf and parks the transaction on a
@@ -21,9 +21,9 @@
 # ../chromium-fourier/KWIN_PIVOT.md for the full diagnosis thread.
 
 pkgname=kwin-fourier
-pkgver=6.6.4
+pkgver=6.6.5
 _dirver=$(echo $pkgver | cut -d. -f1-3)
-pkgrel=3
+pkgrel=1
 _upname=kwin
 epoch=1
 arch=(aarch64 x86_64)
@@ -103,7 +103,7 @@ conflicts=(kwin)
 replaces=(kwin)
 source=(https://download.kde.org/stable/plasma/$_dirver/$_upname-$pkgver.tar.xz{,.sig}
         0001-transaction-bypass-watchDmaBuf-fence-wait.patch)
-sha256sums=('3f9439760580a977d018daf4b35b62e5a1700def7b21c8dfbfc789d21378d7ad'
+sha256sums=('6c187ce7a5506090b438ef900103836fa0537674dde8b31e5b497ef321643cb4'
             'SKIP'
             'SKIP')
 validpgpkeys=('E0A3EB202F8E57528E13E72FD7574483BB57B18D'  # Jonathan Esk-Riddell <jr@jriddell.org>

arch/libva-v4l2-request-fourier/PKGBUILD

@@ -6,8 +6,11 @@
 # tracks the campaign fork's git history directly, so iteration sweeps
 # (DEBUG removal, follow-up bugfixes) land in a clean linear log.
 #
-# Campaign: ~/src/libva-multiplanar/ (eight closed iterations as of
-# 2026-05-06; iter8 close is the production tip pinned below).
+# Campaign: ~/src/libva-multiplanar/ (iter8 close 2026-05-06) plus
+# ~/src/fresnel-fourier/ which carried the fork to iter38b — multi-device
+# probe so a single libva session serves all 5 codecs (rkvdec H.264 +
+# HEVC + VP9, hantro MPEG-2 + VP8) plus a bounds-check fix for
+# MAX_PROFILES. Pinned below.
 # Fork repo: https://git.reauktion.de/marfrit/libva-v4l2-request-fourier
 # Bootlin upstream: https://github.com/bootlin/libva-v4l2-request
 #
@@ -18,18 +21,32 @@
 # Alternative: boltzmann via his subagent + marfrit-publish.
 
 pkgname=libva-v4l2-request-fourier
+epoch=1
 _upstreampkg=libva-v4l2-request
 
-# Pin the fork tip. 65969da = "iter8 Phase 4: tests/run_perf_binding_cell.sh"
-# — last commit on master before fresnel-fourier work started layering
-# MPEG-2 rewrites on top of it (2026-05-08). The libva-multiplanar campaign
-# closed iter8 at this commit; promote to a later pin only after a future
-# iteration closes cleanly.
-_commit=65969da3ee901442a8ca4e1f7f9697d5461d368a
+# Pin the fork tip. c454618 = PR #16 merge "picture, request_pool:
+# transparent OUTPUT-pool resize on bitstream overrun (#15)" —
+# follow-up root-cause fix to #13/#14. On a mid-stream bitstream-
+# budget overrun (typical cause: SPS-driven resolution upshift in an
+# adaptive-bitrate stream), codec_store_buffer now snapshots the in-
+# flight surface's accumulated bytes, releases its OUTPUT pool slot,
+# calls request_pool_resize (STREAMOFF → REQBUFS(0) → S_FMT with
+# 2×sizeimage hint, capped at 1 GiB, page-aligned → CREATE_BUFS →
+# mmap → media_request_alloc → STREAMON), re-acquires a slot, re-
+# mirrors the surface's source_{data,size,request_fd}, restores the
+# bytes, and continues. The frame survives instead of being dropped
+# back to libavcodec for surface recreation. CAPTURE side untouched
+# (per-queue V4L2 streaming independence).
+#
+# Prior pin (2860d75) = PR #14 merge — codec_store_buffer bounds-
+# check floor (#13).
+_commit=c454618ae11addce2e17b560f4deeacbed067d98
 
 # Project version from meson.build (1.0.0) + commit count + short sha,
-# matching the ffmpeg-v4l2-request-fourier convention.
-pkgver=1.0.0.r280.65969da
+# matching the ffmpeg-v4l2-request-fourier convention. Recomputed at
+# build time by pkgver() below; the static value here is a placeholder
+# so AUR-style consumers see something coherent before src/ exists.
+pkgver=1.0.0.r390.c454618
 pkgrel=1
 pkgdesc="VA-API backend for V4L2 stateless decoders (multiplanar fork — fourier umbrella)"
 arch=('aarch64')
@@ -55,7 +72,15 @@ build() {
     cd "${srcdir}/${_upstreampkg}-fourier"
     # meson_options.txt only exposes 'kernel_headers' — leave it empty to
     # use system /usr/include kernel UAPI headers. No per-codec toggles.
-    arch-meson build --buildtype=release
+    #
+    # b_lto=false: override arch-meson's wrapper default of `-D b_lto=true`,
+    # which the makepkg.conf OPTIONS=(..., !lto, ...) line does NOT actually
+    # override (arch-meson hard-codes b_lto=true). The hand-built reproducer
+    # from issue #17 shows: LTO/ICF kernel-folds per-codec helpers and HEVC's
+    # multi-control-struct chain (SPS+PPS+DECODE_PARAMS+SLICE_PARAMS) gets a
+    # wrong helper-instance pulled in at vaEndPicture → segfault. The 4 other
+    # codecs (single-control-struct) tolerate the folding by accident.
+    arch-meson build --buildtype=release -Db_lto=false
     meson compile -C build
 }
 

arch/libva-v4l2-request-ohm-gl-fix/PKGBUILD

@@ -1,6 +1,13 @@
 # Maintainer: Markus Fritsche <fritsche.markus@gmail.com>
 # Campaign: ohm_gl_fix Phase 6 Step 1
 #
+# DEPRECATED (2026-05-16): superseded by ../libva-v4l2-request-fourier/
+# which tracks the campaign fork's git history directly and adds the
+# iter38 multi-device probe (single libva session for rkvdec H.264/HEVC/VP9
+# + hantro MPEG-2/VP8). The successor declares
+# replaces=('libva-v4l2-request-ohm-gl-fix'), so installing it removes
+# this package automatically. See README.md for the full deprecation note.
+#
 # Forks libva-v4l2-request to add hantro-vpu multiplanar + modern
 # stateless UAPI support. Conflicts/replaces stock libva-v4l2-request.
 #
@@ -11,7 +18,7 @@ pkgname=libva-v4l2-request-ohm-gl-fix
 _upstreampkg=libva-v4l2-request
 pkgver=1.0.0.r0.ga3c2476
 pkgrel=2
-pkgdesc="VA-API backend for V4L2 stateless decoders, hantro-vpu multiplanar fork"
+pkgdesc="DEPRECATED — use libva-v4l2-request-fourier. VA-API backend for V4L2 stateless decoders, hantro-vpu multiplanar fork"
 arch=('aarch64')
 url="https://github.com/bootlin/libva-v4l2-request"
 license=('LGPL2.1' 'MIT')

arch/libva-v4l2-request-ohm-gl-fix/README.md

@@ -1,5 +1,24 @@
 # libva-v4l2-request-ohm-gl-fix
 
+> ## ⚠ DEPRECATED — use [`libva-v4l2-request-fourier`](../libva-v4l2-request-fourier/) instead
+>
+> This package is the **predecessor experimental** build (tarball pin
+> + 18 stacked patches) and is no longer maintained as of 2026-05-16.
+> Its successor `libva-v4l2-request-fourier` tracks the campaign fork's
+> git history directly
+> ([git.reauktion.de/marfrit/libva-v4l2-request-fourier](https://git.reauktion.de/marfrit/libva-v4l2-request-fourier))
+> so iteration sweeps (DEBUG removal, follow-up bugfixes) land in a clean
+> linear log, and adds the iter38 multi-device probe that lets a single
+> libva session serve rkvdec H.264/HEVC/VP9 + hantro MPEG-2/VP8 without
+> needing `LIBVA_V4L2_REQUEST_VIDEO_PATH` overrides.
+>
+> `libva-v4l2-request-fourier` declares
+> `replaces=('libva-v4l2-request-ohm-gl-fix')`, so installing it will
+> remove this package automatically. Kept in-tree as historical reference
+> for the ohm_gl_fix Phase 6 audit trail.
+
+---
+
 Bootlin's libva-v4l2-request VA-API backend, with hantro-vpu
 multi-planar + chromium-149-era stateless H.264 patches developed
 in the [ohm_gl_fix campaign](../../../ohm_gl_fix/) Phase 6 Step 1

arch/linux-ampere-fourier/PKGBUILD

@@ -0,0 +1,168 @@
+# Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+#
+# linux-ampere-fourier — CoolPi GenBook (RK3588) kernel built from the
+# kernel-agent fleet/ampere.yaml manifest applied to mainline v7.0-rc3.
+#
+# kafr2 baseline (2026-05-18): mainline v7.0-rc3 + the 10 scope-tagged
+# kernel-agent patches under patches/{soc,module,board,driver}/:
+#   - 1 soc/rk3588 pwm15 pinctrl
+#   - 6 board/coolpi-cm5-genbook DTS patches (pwm-fan, RK806 power-off,
+#     speaker, USB-C PD, lid switch + USB3 PHY, microphone)
+#   - 3 driver/media VP9-on-VDPU381 patches (Sarma's v8 series, imported
+#     via marfrit/kernel-agent#12 closure and PR #24)
+#
+# Drops the prior f8f3ad9 baseline ("18 commits ahead") because that tip
+# black-screens ampere — kernel-agent's ka-promote produces this 10-patch
+# minimal set from fleet/ampere.yaml. End-to-end VP9 + AV1 (av1-vpu-dec
+# is mainline-7.0) decode verified bit-exact via kdirect on the
+# hand-built tip 48a8c78 before this package iteration was cut.
+#
+# Coexists with the user's other extlinux labels in
+# /boot/firmware/extlinux/extlinux.conf; never edits them. Adds a
+# managed `linux-ampere-fourier` label (the user sets `default` manually
+# after verifying boot).
+#
+# Bootloader path: /boot/firmware/ (vfat on mmcblk0p1). Kernel +
+# initramfs + DTB land there directly. Reverting = boot a different
+# extlinux label (e.g. arch_mainline, ubuntu_mainline).
+
+pkgbase=linux-ampere-fourier
+pkgname=("$pkgbase" "$pkgbase-headers")
+pkgver=7.0rc3.kafr2
+pkgrel=1
+pkgdesc='CoolPi GenBook kernel (v7.0-rc3 + kernel-agent fleet/ampere.yaml — 6 board patches + 3 VP9-VDPU381 + 1 pwm15)'
+arch=(aarch64)
+url='https://git.reauktion.de/marfrit/kernel-agent'
+license=(GPL-2.0-only)
+makedepends=(
+  bc cpio gettext kmod libelf pahole perl python tar xz
+  ccache
+  uboot-tools dtc
+)
+options=('!strip')
+
+# Pinned tip of the kernel-agent-managed source tree for ampere.
+# 10 commits ahead of v7.0-rc3, exactly mirroring fleet/ampere.yaml's
+# manifest under apply order:
+#   - c57d069 soc/rk3588: pwm15 pinctrl entries
+#   - 05a915c board/genbook: pwm-fan with thermal cooling
+#   - d007b90 module/coolpi-cm5: RK806 system-power-controller
+#   - 3722eab board/genbook: speaker via audio-graph-card
+#   - 3e42ab6 board/genbook: USB-C PD via FUSB302
+#   - 7c241f2 board/genbook: lid switch + USB3 PHY lane
+#   - dd545fa board/genbook: wire internal microphone
+#   - 9ddcae5 driver/media: rkvdec-vp9 helper rename (Sarma)
+#   - c5063d9 driver/media: rkvdec move vp9 to common (Sarma)
+#   - 48a8c78 driver/media: rkvdec VP9 for VDPU381 (Sarma)
+#
+# This is the same tree state ka-promote ampere produces as cumulative.patch
+# (see marfrit/kernel-agent build/ampere/v7.0-rc3/manifest.lock for the
+# b2sum + per-patch sha256s).
+_commit=48a8c785de7f5320513052a64e544c6310d7b273
+
+source=(
+  # Local tarball produced by ./prebuild.sh from a local clone of the
+  # linux-rk3588-marfrit branch. Not fetched from a URL because the
+  # boltzmann working clone is shallow (gitea push rejects) and the
+  # 260MB tarball isn't committed to marfrit-packages. Run prebuild.sh
+  # before makepkg; see README in this dir.
+  "linux-rk3588-marfrit-${_commit:0:7}.tar.gz"
+  'config'  # snapshot of running ampere kernel's /proc/config.gz (7.0.0-rc3-ARCH+)
+  'linux-ampere-fourier.preset'
+  'extlinux-add.hook'
+  'extlinux-add.sh'
+)
+sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
+
+# kernelrelease becomes <Makefile-VERSION>.<PATCH>.<SUBLEVEL><EXTRAVERSION><LOCALVERSION>
+# i.e. 7.0.0-rc3-ampere-fourier. Module dir + EXTRAVERSION suffix keep
+# this disjoint from the hand-managed /boot/firmware/Image-7.0.0-rc3-ARCH+
+# that's currently on the host.
+_kernver=7.0.0-rc3-ampere-fourier
+_srcdir=linux-rk3588-marfrit
+
+prepare() {
+  cd "${_srcdir}"
+
+  echo ":: writing config"
+  cp "${srcdir}/config" .config
+
+  # LOCALVERSION suffix to differentiate from upstream-stock builds.
+  scripts/config --set-str LOCALVERSION "-ampere-fourier"
+  scripts/config -d LOCALVERSION_AUTO
+
+  echo ":: olddefconfig (accept new symbols sensibly)"
+  make olddefconfig
+
+  make -s kernelrelease > version
+  echo ":: kernel release: $(<version)"
+}
+
+build() {
+  cd "${_srcdir}"
+  unset LDFLAGS
+  # Native build only — no distcc per kernel-agent policy
+  # (feedback_kernel_agent_no_distcc.md). ccache stays.
+  export CC="ccache gcc"
+  export HOSTCC="ccache gcc"
+  make ${MAKEFLAGS:--j$(nproc)} Image modules dtbs
+}
+
+_package() {
+  pkgdesc='CoolPi GenBook overclocked kernel (ampere-fourier baseline)'
+  depends=(coreutils kmod mkinitcpio uboot-tools)
+  optdepends=('linux-firmware: firmware images needed for some devices')
+  backup=("etc/mkinitcpio.d/${pkgbase}.preset")
+
+  cd "${_srcdir}"
+  local _kver
+  _kver=$(<version)
+
+  # Kernel image into the vfat firmware partition (where extlinux looks).
+  install -Dm644 arch/arm64/boot/Image \
+    "${pkgdir}/boot/firmware/Image-ampere-fourier"
+
+  # Single DTB for the GenBook target — install directly under
+  # /boot/firmware/ (no subdir, matches existing host convention).
+  install -Dm644 arch/arm64/boot/dts/rockchip/rk3588-coolpi-cm5-genbook.dtb \
+    "${pkgdir}/boot/firmware/rk3588-coolpi-cm5-genbook.dtb-ampere-fourier"
+
+  ZSTD_CLEVEL=19 make INSTALL_MOD_PATH="${pkgdir}/usr" \
+    INSTALL_MOD_STRIP=1 modules_install
+  rm -f "${pkgdir}/usr/lib/modules/${_kver}/"{source,build}
+
+  install -Dm644 "${srcdir}/${pkgbase}.preset" \
+    "${pkgdir}/etc/mkinitcpio.d/${pkgbase}.preset"
+
+  install -Dm755 "${srcdir}/extlinux-add.hook" \
+    "${pkgdir}/usr/share/libalpm/hooks/95-${pkgbase}-extlinux.hook"
+  install -Dm755 "${srcdir}/extlinux-add.sh" \
+    "${pkgdir}/usr/share/libalpm/scripts/${pkgbase}-extlinux"
+}
+
+_package-headers() {
+  pkgdesc='Headers and scripts for the linux-ampere-fourier kernel'
+  depends=(pahole)
+
+  cd "${_srcdir}"
+  local _kver _builddir
+  _kver=$(<version)
+  _builddir="${pkgdir}/usr/lib/modules/${_kver}/build"
+
+  install -Dt "${_builddir}" -m644 .config Makefile Module.symvers System.map vmlinux version
+  install -Dt "${_builddir}/kernel" -m644 kernel/Makefile
+
+  cp -a scripts "${_builddir}"
+  install -Dt "${_builddir}/arch/arm64" -m644 arch/arm64/Makefile
+  cp -a arch/arm64/include "${_builddir}/arch/arm64/"
+  cp -a include "${_builddir}/"
+
+  find . -name 'Kbuild' -exec install -Dm644 {} "${_builddir}/{}" \;
+  find . -name 'Kconfig*' -exec install -Dm644 {} "${_builddir}/{}" \;
+
+  install -d "${pkgdir}/usr/src"
+  ln -sr "${_builddir}" "${pkgdir}/usr/src/${pkgbase}"
+}
+
+eval "package_${pkgbase}() { _package; }"
+eval "package_${pkgbase}-headers() { _package-headers; }"

arch/linux-ampere-fourier/README.md

@@ -0,0 +1,70 @@
+# linux-ampere-fourier
+
+Kernel package for ampere (CoolPi GenBook RK3588). Baselined on
+`marfrit/linux-rk3588-marfrit @ f8f3ad9` (mainline v7.0-rc3 + 18
+RK3588-fleet-relevant commits — 10 Markus, 8 upstream cherry-picks).
+
+See `marfrit/kernel-agent/fleet/ampere.yaml` for the manifest +
+`marfrit/kernel-agent/patches/{soc,module,board}/...` for the
+scope-tagged board patches in the baseline.
+
+## Build
+
+The kernel source isn't on Gitea — boltzmann's working clone is
+shallow (Gitea refuses shallow pushes) and a 260MB tarball doesn't
+belong in `marfrit-packages`. Stage the source locally from a
+clone of the `linux-rk3588-marfrit` branch:
+
+```sh
+cd arch/linux-ampere-fourier
+./prebuild.sh           # produces linux-rk3588-marfrit-f8f3ad9.tar.gz
+makepkg -s --noconfirm  # native aarch64 build; no distcc
+```
+
+`prebuild.sh` looks at `$LINUX_RK3588_MARFRIT_TREE` (default
+`~/src/linux-rockchip`) for the kernel working tree. The tip commit
+must be reachable in that clone — fetch the `linux-rk3588-marfrit`
+branch first if you cloned from elsewhere.
+
+## Build hosts
+
+Native aarch64 only (per kernel-agent `feedback_kernel_agent_no_distcc.md`).
+Either ampere itself (8C/2.4GHz, 32GB, native arch) or boltzmann
+(Rock 5 ITX+, same uarch). fermi as fallback.
+
+## Install
+
+Adds a managed label to `/boot/firmware/extlinux/extlinux.conf`:
+
+```
+label linux-ampere-fourier
+    menu label linux-ampere-fourier (managed)
+    kernel /Image-ampere-fourier
+    initrd /initramfs-ampere-fourier.img
+    fdt    /rk3588-coolpi-cm5-genbook.dtb-ampere-fourier
+    append <inherited from arch_mainline>
+```
+
+Default label is NOT changed. After verifying boot of the managed
+label at the u-boot menu, you flip `default` manually. Reverting =
+boot a different label (e.g. `arch_mainline`, `ubuntu_mainline`).
+
+## Boot path
+
+ampere uses `/boot/firmware/` (vfat on mmcblk0p1, ~1G), distinct
+from fresnel's `/boot/` on root partition. The PKGBUILD installs
+Image, initramfs, and DTB directly under `/boot/firmware/`. No
+DTB subdir — single board target.
+
+## Open follow-ups (per kernel-agent issue #6)
+
+- **Ask 2** (VP9 enablement on RK3588 rkvdec) — not addressed in this
+  iteration. Separate session.
+- **Ask 3** (AV1 decoder integration) — backend libva work, not kernel.
+- Hosting the source tarball publicly so `prebuild.sh` isn't needed —
+  candidate: Gitea release asset, or `packages.reauktion.de/sources/`.
+- Splitting the 12 non-board cherry-pick commits in the baseline
+  (4 Shawn Lin phy, 2 Cristian Ciocaltea, etc.) into scope-tagged
+  patches in kernel-agent — currently they ride along inside the
+  pinned baseline rather than being explicit `includes:` in the
+  manifest.

arch/linux-ampere-fourier/extlinux-add.hook

@@ -0,0 +1,13 @@
+[Trigger]
+Operation = Install
+Operation = Upgrade
+Operation = Remove
+Type = Path
+Target = boot/firmware/Image-ampere-fourier
+Target = boot/firmware/rk3588-coolpi-cm5-genbook.dtb-ampere-fourier
+Target = boot/firmware/initramfs-ampere-fourier.img
+
+[Action]
+Description = Updating extlinux entry for linux-ampere-fourier
+When = PostTransaction
+Exec = /usr/share/libalpm/scripts/linux-ampere-fourier-extlinux

arch/linux-ampere-fourier/extlinux-add.sh

@@ -0,0 +1,62 @@
+#!/bin/sh
+# Add / update / remove the linux-ampere-fourier entry in
+# /boot/firmware/extlinux/extlinux.conf. Idempotent. Coexists with
+# the hand-managed labels in that file; never edits them. Default
+# label is NOT touched — user picks at u-boot menu.
+#
+# ampere boots from a vfat partition (mmcblk0p1) mounted at
+# /boot/firmware/, distinct from fresnel's /boot/ on root.
+
+set -eu
+
+CONF="/boot/firmware/extlinux/extlinux.conf"
+TAG_BEGIN="# >>> linux-ampere-fourier (managed) >>>"
+TAG_END="# <<< linux-ampere-fourier (managed) <<<"
+
+# Copy APPEND from the user's `arch_mainline` label so the managed
+# entry inherits the same root= and console= settings the host's
+# bootloader already trusts. Falls back to a CoolPi GenBook default
+# if no arch_mainline label exists (first-time install on a fresh
+# bootloader config).
+EXISTING_APPEND=$(awk '
+  /^[[:space:]]*label[[:space:]]+arch_mainline[[:space:]]*$/ { found=1; next }
+  found && /^[[:space:]]*append[[:space:]]/ {
+    sub(/^[[:space:]]*append[[:space:]]+/, "")
+    print
+    exit
+  }
+  /^[[:space:]]*label[[:space:]]/ { found=0 }
+' "$CONF" 2>/dev/null || true)
+
+APPEND="${EXISTING_APPEND:-root=LABEL=arch rw rootwait rootfstype=btrfs rootflags=subvol=@,ssd,discard=async console=ttyS2,1500000 console=tty1 consoleblank=0 loglevel=7 cma=256M coherent_pool=2M}"
+
+ENTRY=$(cat <<EOF
+${TAG_BEGIN}
+label linux-ampere-fourier
+	menu label linux-ampere-fourier (managed)
+	kernel /Image-ampere-fourier
+	initrd /initramfs-ampere-fourier.img
+	fdt    /rk3588-coolpi-cm5-genbook.dtb-ampere-fourier
+	append ${APPEND}
+${TAG_END}
+EOF
+)
+
+# Strip any prior managed block, then append fresh
+TMP=$(mktemp)
+awk -v b="$TAG_BEGIN" -v e="$TAG_END" '
+  $0==b{skip=1; next}
+  $0==e{skip=0; next}
+  !skip{print}
+' "$CONF" > "$TMP"
+
+# Post-Remove: kernel files absent → don't re-add the entry
+if [ -f "/boot/firmware/Image-ampere-fourier" ] \
+   && [ -f "/boot/firmware/rk3588-coolpi-cm5-genbook.dtb-ampere-fourier" ]; then
+  printf '%s\n' "$ENTRY" >> "$TMP"
+  echo "linux-ampere-fourier: extlinux entry updated"
+else
+  echo "linux-ampere-fourier: kernel files absent, entry removed"
+fi
+
+mv "$TMP" "$CONF"

arch/linux-ampere-fourier/linux-ampere-fourier.preset

@@ -0,0 +1,13 @@
+# mkinitcpio preset for linux-ampere-fourier
+#
+# ampere boots from /boot/firmware/ (vfat partition on mmcblk0p1). The
+# initramfs lands there too so extlinux can pick it up. Only one PRESET
+# because /boot/firmware is ~1G total — no room for a fallback image
+# alongside the primary.
+
+ALL_kver="/boot/firmware/Image-ampere-fourier"
+ALL_microcode=()
+
+PRESETS=('default')
+
+default_image="/boot/firmware/initramfs-ampere-fourier.img"

arch/linux-ampere-fourier/prebuild.sh

@@ -0,0 +1,68 @@
+#!/bin/bash
+# prebuild — stage the kernel source tarball this PKGBUILD expects.
+#
+# linux-ampere-fourier's source is a snapshot of marfrit/linux-rk3588-marfrit
+# @ f8f3ad9 (260MB), too big to commit to marfrit-packages and currently
+# unpushable to Gitea (boltzmann's working clone is shallow; gitea push
+# refuses shallow updates). Hosting the tarball outside Gitea would need
+# infrastructure setup that's not in scope for the first iteration.
+#
+# So: produce the tarball locally from the kernel working tree just
+# before makepkg runs. Idempotent — if an existing tarball matches the
+# expected sha256 we skip the archive step.
+#
+# Run from this directory:  cd arch/linux-ampere-fourier && ./prebuild.sh
+# Override the kernel-tree location:  LINUX_RK3588_MARFRIT_TREE=/path ./prebuild.sh
+#
+# Default tree location matches the boltzmann/ampere convention from
+# kernel-agent issue #6: $HOME/src/linux-rockchip.
+set -euo pipefail
+
+TREE="${LINUX_RK3588_MARFRIT_TREE:-${HOME}/src/linux-rockchip}"
+COMMIT=48a8c785de7f5320513052a64e544c6310d7b273
+# Generated tarball sha varies with gzip version — script warns-not-fails.
+# Leave EXPECTED empty for fresh kafr2 builds; first successful build can
+# pin a canonical sha here if a reproducibility audit needs it.
+SHA256_EXPECTED=
+
+HERE="$(cd "$(dirname "$0")" && pwd)"
+OUTPUT="${HERE}/linux-rk3588-marfrit-${COMMIT:0:7}.tar.gz"
+
+if [ -f "$OUTPUT" ]; then
+    have=$(sha256sum "$OUTPUT" | cut -d' ' -f1)
+    if [ "$have" = "$SHA256_EXPECTED" ]; then
+        echo "prebuild: $OUTPUT already exists with correct sha256"
+        exit 0
+    else
+        echo "prebuild: existing $OUTPUT sha mismatch (have=$have, want=$SHA256_EXPECTED) — regenerating" >&2
+        rm -f "$OUTPUT"
+    fi
+fi
+
+if [ ! -d "$TREE/.git" ]; then
+    echo "prebuild: kernel tree not at $TREE" >&2
+    echo "          set LINUX_RK3588_MARFRIT_TREE=/path/to/linux-rockchip and retry" >&2
+    exit 2
+fi
+
+cd "$TREE"
+
+if ! git cat-file -e "$COMMIT" 2>/dev/null; then
+    echo "prebuild: commit $COMMIT not found in $TREE" >&2
+    echo "          fetch the linux-rk3588-marfrit branch first:" >&2
+    echo "          git fetch <remote> linux-rk3588-marfrit" >&2
+    exit 3
+fi
+
+echo "prebuild: generating archive from $TREE @ $COMMIT..."
+git archive --format=tar.gz --prefix=linux-rk3588-marfrit/ "$COMMIT" -o "$OUTPUT"
+
+# git archive emits a deterministic tar stream but gzip compression may
+# vary by version. The sha256 check is informational; warn-don't-fail.
+have=$(sha256sum "$OUTPUT" | cut -d' ' -f1)
+if [ "$have" != "$SHA256_EXPECTED" ]; then
+    echo "prebuild: WARNING $OUTPUT sha=$have (canonical=$SHA256_EXPECTED)" >&2
+    echo "          probably a gzip-version difference; tar payload should be identical" >&2
+fi
+
+echo "prebuild: wrote $OUTPUT ($(du -h "$OUTPUT" | cut -f1), sha=$have)"

arch/linux-fresnel-fourier/0004-media-videobuf2-add-opt-in-dma_resv-producer-fence-h.patch

@@ -0,0 +1,356 @@
+From a202de1646d4c8f8ee2ebc2e4c100b621975754a Mon Sep 17 00:00:00 2001
+In-Reply-To: <20260429195306.239666-1-mfritsche@reauktion.de>
+References: <20260429195306.239666-1-mfritsche@reauktion.de>
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Sat, 9 May 2026 16:16:07 +0200
+Subject: [PATCH RFC v2] media: videobuf2: add opt-in dma_resv producer fence
+ helper
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+V4L2 producers historically don't propagate buffer-state-done into
+the dmabuf's dma_resv exclusive fence. Userspace consumers that
+import V4L2-produced dmabufs and wait on the dmabuf's implicit-sync
+fence (poll(POLLIN), DMA_BUF_IOCTL_EXPORT_SYNC_FILE,
+EGL_LINUX_DMA_BUF_EXT) currently see either zero fences or a stub
+fence from dma_fence_get_stub(). This is correct by accident for the
+common DQBUF-then-import case but represents a contract gap that
+breaks Wayland compositors importing CAPTURE buffers from a stateless
+H.264 decoder under continuous playback on implicit-sync GPU stacks
+(observed on RK3566 + hantro VPU + Mali-G52 panfrost; manifests as
+green frames -- BT.709 limited-range YUV(0,0,0) -> RGB(0,77,0) -- when
+the GPU samples the dmabuf before the producer's decode completes).
+
+Add an opt-in API gated by both a per-driver runtime flag
+(vb2_queue::supports_release_fences) and a Kconfig
+(CONFIG_VIDEOBUF2_RELEASE_FENCES, default n) that lets producers
+populate a real dma_resv exclusive write fence on the dmabufs they
+export. Drivers call vb2_buffer_attach_release_fence(vb) at a
+finite-time-fenced point in their pipeline (typically m2m
+device_run, just before the HW kick); vb2_buffer_done() signals and
+puts the fence as part of its state transition.
+
+The publish and signal paths are wrapped in
+dma_fence_begin_signalling() / dma_fence_end_signalling() so
+PROVE_LOCKING can validate that nothing taken in those critical
+sections deadlocks against the signal path. dma_resv_lock is
+sleepable but not taken on the signal path, so taking it inside the
+publish critical section is safe under lockdep.
+
+Skips planes whose vb2_plane.dbuf is NULL -- buffers never exported
+via VIDIOC_EXPBUF (or imported via V4L2_MEMORY_DMABUF) have no
+dmabuf for userspace to wait on.
+
+Drivers that don't opt in pay nothing: the helper is a no-op stub
+when CONFIG_VIDEOBUF2_RELEASE_FENCES=n, and an early-return check
+of supports_release_fences when =y but the flag is unset.
+
+Validated on RK3566 PineTab2 with PROVE_LOCKING enabled: 30s of
+bbb_1080p30 H.264 stateless decode + zero-copy panfrost EGL import
+via dmabuf-wayland (mpv 0.41 + KWin 6.6.4 + Mesa panfrost 26.0.5)
+produces 31,816 dma_fence init/signal pairs across 5,724 vb2 buffer
+cycles with zero lockdep splats from videobuf2 / dma_resv code paths.
+
+Subsequent patches in this series opt the hantro and rockchip-rga
+drivers in.
+
+Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
+Cc: Christian König <christian.koenig@amd.com>
+Cc: Nicolas Dufresne <nicolas@ndufresne.ca>
+Cc: Sumit Semwal <sumit.semwal@linaro.org>
+Cc: Hans Verkuil <hverkuil@xs4all.nl>
+Cc: Tomasz Figa <tfiga@chromium.org>
+Cc: linux-media@vger.kernel.org
+Cc: dri-devel@lists.freedesktop.org
+Cc: linaro-mm-sig@lists.linaro.org
+Signed-off-by: Markus Fritsche <mfritsche@reauktion.de>
+---
+ drivers/media/common/videobuf2/Kconfig        |  29 ++++
+ .../media/common/videobuf2/videobuf2-core.c   | 135 ++++++++++++++++++
+ include/media/videobuf2-core.h                |  51 +++++++
+ 3 files changed, 215 insertions(+)
+
+diff --git a/drivers/media/common/videobuf2/Kconfig b/drivers/media/common/videobuf2/Kconfig
+index d2223a12c..bbfa26984 100644
+--- a/drivers/media/common/videobuf2/Kconfig
++++ b/drivers/media/common/videobuf2/Kconfig
+@@ -30,3 +30,32 @@ config VIDEOBUF2_DMA_SG
+ config VIDEOBUF2_DVB
+ 	tristate
+ 	select VIDEOBUF2_CORE
++
++config VIDEOBUF2_RELEASE_FENCES
++	bool "videobuf2: opt-in dma_resv producer fences for V4L2 dmabuf exports"
++	depends on VIDEOBUF2_CORE
++	depends on DMA_SHARED_BUFFER
++	default n
++	help
++	  Enables an opt-in API that lets vb2 producers populate a dma_resv
++	  exclusive write fence on the dmabufs they export to userspace.
++	  The fence is signalled when the buffer transitions to
++	  VB2_BUF_STATE_DONE.
++
++	  This gives userspace consumers that import V4L2-produced dmabufs
++	  and wait on the dmabuf's implicit-sync fence (poll(POLLIN),
++	  DMA_BUF_IOCTL_EXPORT_SYNC_FILE, EGL_LINUX_DMA_BUF_EXT) a real
++	  producer fence to wait on, instead of a stub fence from
++	  dma_fence_get_stub() that the dma_buf core substitutes when
++	  dma_resv is empty.
++
++	  Drivers individually opt in by setting
++	  vb2_queue::supports_release_fences = true and calling
++	  vb2_buffer_attach_release_fence() at the right point in their
++	  pipeline (typically m2m device_run, just before HW kick).
++
++	  Distributors leave this off unless targeting Wayland/EGL
++	  consumers of V4L2 stateless decoder output on
++	  implicit-sync-only GPU stacks (e.g. mainline panfrost).
++
++	  If unsure, say N.
+diff --git a/drivers/media/common/videobuf2/videobuf2-core.c b/drivers/media/common/videobuf2/videobuf2-core.c
+index adf668b21..85d7fddbd 100644
+--- a/drivers/media/common/videobuf2/videobuf2-core.c
++++ b/drivers/media/common/videobuf2/videobuf2-core.c
+@@ -26,6 +26,12 @@
+ #include <linux/freezer.h>
+ #include <linux/kthread.h>
+ 
++#ifdef CONFIG_VIDEOBUF2_RELEASE_FENCES
++#include <linux/dma-fence.h>
++#include <linux/dma-resv.h>
++#include <linux/dma-buf.h>
++#endif
++
+ #include <media/videobuf2-core.h>
+ #include <media/v4l2-mc.h>
+ 
+@@ -1173,6 +1179,120 @@ void *vb2_plane_cookie(struct vb2_buffer *vb, unsigned int plane_no)
+ }
+ EXPORT_SYMBOL_GPL(vb2_plane_cookie);
+ 
++#ifdef CONFIG_VIDEOBUF2_RELEASE_FENCES
++/*
++ * dma_resv release-fence integration.
++ *
++ * Optional, opt-in path that lets producers publish a real
++ * dma_fence on their CAPTURE-side dmabufs so userspace consumers
++ * (compositors, EGL importers) get spec-clean implicit-sync
++ * semantics instead of the dma_buf core's stub fence. Drivers
++ * call vb2_buffer_attach_release_fence() at a finite-time-fenced
++ * point (typically m2m device_run) and the fence is signalled by
++ * vb2_buffer_done(). Gated at runtime by
++ * vb2_queue::supports_release_fences and at compile time by
++ * CONFIG_VIDEOBUF2_RELEASE_FENCES.
++ */
++
++static const char *vb2_dma_resv_get_driver_name(struct dma_fence *fence)
++{
++	return "videobuf2";
++}
++
++static const char *vb2_dma_resv_get_timeline_name(struct dma_fence *fence)
++{
++	return "vb2-release-fence";
++}
++
++static const struct dma_fence_ops vb2_dma_resv_fence_ops = {
++	.get_driver_name = vb2_dma_resv_get_driver_name,
++	.get_timeline_name = vb2_dma_resv_get_timeline_name,
++};
++
++int vb2_buffer_attach_release_fence(struct vb2_buffer *vb)
++{
++	struct vb2_queue *q = vb->vb2_queue;
++	struct dma_fence *fence;
++	unsigned int plane;
++	bool cookie;
++
++	if (!q->supports_release_fences)
++		return 0;
++
++	if (WARN_ON(vb->release_fence))
++		return -EINVAL;
++
++	fence = kzalloc(sizeof(*fence), GFP_KERNEL);
++	if (!fence)
++		return -ENOMEM;
++
++	dma_fence_init(fence, &vb2_dma_resv_fence_ops, &q->dma_resv_fence_lock,
++		       q->dma_resv_fence_context,
++		       atomic64_inc_return(&q->dma_resv_fence_seqno));
++
++	/*
++	 * Annotate the publish-side critical section. Per
++	 * Documentation/driver-api/dma-buf.rst, lockdep validates
++	 * that nothing taken in this region can deadlock against
++	 * the signal path in vb2_buffer_signal_release_fence().
++	 * dma_resv_lock is sleepable but is not taken on the signal
++	 * path, so taking it inside the critical section is safe.
++	 */
++	cookie = dma_fence_begin_signalling();
++	for (plane = 0; plane < vb->num_planes; plane++) {
++		struct dma_buf *dbuf = vb->planes[plane].dbuf;
++
++		if (!dbuf)
++			continue;
++
++		dma_resv_lock(dbuf->resv, NULL);
++		dma_resv_add_fence(dbuf->resv, fence, DMA_RESV_USAGE_WRITE);
++		dma_resv_unlock(dbuf->resv);
++	}
++	dma_fence_end_signalling(cookie);
++
++	/* One reference for the eventual signal in vb2_buffer_done. */
++	vb->release_fence = dma_fence_get(fence);
++
++	/* The dma_resv held its own reference per plane. Drop ours. */
++	dma_fence_put(fence);
++
++	return 0;
++}
++EXPORT_SYMBOL_GPL(vb2_buffer_attach_release_fence);
++
++static void vb2_buffer_signal_release_fence(struct vb2_buffer *vb,
++					    enum vb2_buffer_state state)
++{
++	struct dma_fence *fence = vb->release_fence;
++	bool cookie;
++
++	if (!fence)
++		return;
++
++	cookie = dma_fence_begin_signalling();
++	if (state == VB2_BUF_STATE_ERROR)
++		dma_fence_set_error(fence, -EIO);
++	dma_fence_signal(fence);
++	dma_fence_end_signalling(cookie);
++
++	dma_fence_put(fence);
++	vb->release_fence = NULL;
++}
++#else /* !CONFIG_VIDEOBUF2_RELEASE_FENCES */
++
++int vb2_buffer_attach_release_fence(struct vb2_buffer *vb)
++{
++	return 0;
++}
++EXPORT_SYMBOL_GPL(vb2_buffer_attach_release_fence);
++
++static inline void vb2_buffer_signal_release_fence(struct vb2_buffer *vb,
++						   enum vb2_buffer_state state)
++{
++}
++#endif /* CONFIG_VIDEOBUF2_RELEASE_FENCES */
++
+ void vb2_buffer_done(struct vb2_buffer *vb, enum vb2_buffer_state state)
+ {
+ 	struct vb2_queue *q = vb->vb2_queue;
+@@ -1199,6 +1319,9 @@ void vb2_buffer_done(struct vb2_buffer *vb, enum vb2_buffer_state state)
+ 	if (state != VB2_BUF_STATE_QUEUED)
+ 		__vb2_buf_mem_finish(vb);
+ 
++	if (state != VB2_BUF_STATE_QUEUED)
++		vb2_buffer_signal_release_fence(vb, state);
++
+ 	spin_lock_irqsave(&q->done_lock, flags);
+ 	if (state == VB2_BUF_STATE_QUEUED) {
+ 		vb->state = VB2_BUF_STATE_QUEUED;
+@@ -2651,6 +2774,18 @@ int vb2_core_queue_init(struct vb2_queue *q)
+ 	mutex_init(&q->mmap_lock);
+ 	init_waitqueue_head(&q->done_wq);
+ 
++#ifdef CONFIG_VIDEOBUF2_RELEASE_FENCES
++	/*
++	 * Per-queue dma_resv release-fence context. Drivers that
++	 * opt in via supports_release_fences and call
++	 * vb2_buffer_attach_release_fence() use these to allocate
++	 * fences on a single per-queue timeline.
++	 */
++	q->dma_resv_fence_context = dma_fence_context_alloc(1);
++	atomic64_set(&q->dma_resv_fence_seqno, 0);
++	spin_lock_init(&q->dma_resv_fence_lock);
++#endif
++
+ 	q->memory = VB2_MEMORY_UNKNOWN;
+ 
+ 	if (q->buf_struct_size == 0)
+diff --git a/include/media/videobuf2-core.h b/include/media/videobuf2-core.h
+index 4424d481d..766ff2194 100644
+--- a/include/media/videobuf2-core.h
++++ b/include/media/videobuf2-core.h
+@@ -288,6 +288,16 @@ struct vb2_buffer {
+ 	unsigned int		skip_cache_sync_on_finish:1;
+ 
+ 	struct vb2_plane	planes[VB2_MAX_PLANES];
++#ifdef CONFIG_VIDEOBUF2_RELEASE_FENCES
++	/*
++	 * Producer release fence published on each plane's
++	 * dmabuf->resv when the driver opts in via
++	 * vb2_buffer_attach_release_fence(). Signalled and put by
++	 * vb2_buffer_done() on transition to DONE/ERROR. NULL when
++	 * the driver did not opt in for this buffer.
++	 */
++	struct dma_fence	*release_fence;
++#endif
+ 	struct list_head	queued_entry;
+ 	struct list_head	done_entry;
+ #ifdef CONFIG_VIDEO_ADV_DEBUG
+@@ -648,6 +658,19 @@ struct vb2_queue {
+ 	spinlock_t			done_lock;
+ 	wait_queue_head_t		done_wq;
+ 
++#ifdef CONFIG_VIDEOBUF2_RELEASE_FENCES
++	/*
++	 * dma_resv release-fence context. Drivers that set
++	 * supports_release_fences and call
++	 * vb2_buffer_attach_release_fence() use these to allocate
++	 * fences on a per-queue timeline.
++	 */
++	u64				dma_resv_fence_context;
++	atomic64_t			dma_resv_fence_seqno;
++	spinlock_t			dma_resv_fence_lock;
++#endif
++
++	unsigned int			supports_release_fences:1;
+ 	unsigned int			streaming:1;
+ 	unsigned int			start_streaming_called:1;
+ 	unsigned int			error:1;
+@@ -735,6 +758,34 @@ void *vb2_plane_cookie(struct vb2_buffer *vb, unsigned int plane_no);
+  */
+ void vb2_buffer_done(struct vb2_buffer *vb, enum vb2_buffer_state state);
+ 
++/**
++ * vb2_buffer_attach_release_fence() - opt-in dma_resv release fence.
++ * @vb: the buffer being committed to the producer.
++ *
++ * Drivers that have set vb2_queue::supports_release_fences may call
++ * this from any sleepable context where they have committed to
++ * running the operation in finite time -- typically m2m
++ * device_run(), just before the HW kick. The helper allocates a
++ * dma_fence on the queue's per-queue timeline, attaches it as
++ * DMA_RESV_USAGE_WRITE on each plane's dmabuf->resv, and stashes
++ * it in vb->release_fence. vb2_buffer_done() signals and puts the
++ * fence as part of the buffer's state transition.
++ *
++ * Skips planes whose vb2_plane.dbuf is NULL -- buffers never
++ * exported via VIDIOC_EXPBUF (or imported via V4L2_MEMORY_DMABUF)
++ * have no dmabuf for userspace to wait on.
++ *
++ * No-op when vb2_queue::supports_release_fences is not set
++ * (regardless of CONFIG_VIDEOBUF2_RELEASE_FENCES). When
++ * CONFIG_VIDEOBUF2_RELEASE_FENCES=n, this is a stub that returns 0.
++ *
++ * Returns 0 on success or when the no-op stub is in effect,
++ * negative errno on allocation failure when fence publishing was
++ * attempted. Best-effort: drivers should ignore the return value
++ * unless they want diagnostics.
++ */
++int vb2_buffer_attach_release_fence(struct vb2_buffer *vb);
++
+ /**
+  * vb2_discard_done() - discard all buffers marked as DONE.
+  * @q:		pointer to &struct vb2_queue with videobuf2 queue.
+-- 
+2.53.0
+

arch/linux-fresnel-fourier/0005-media-hantro-attach-dma_resv-release-fence-at-device.patch

@@ -0,0 +1,95 @@
+From 1844c263bde8dd244d7db46f8c508e7c70da459c Mon Sep 17 00:00:00 2001
+In-Reply-To: <20260429195306.239666-1-mfritsche@reauktion.de>
+References: <20260429195306.239666-1-mfritsche@reauktion.de>
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Sat, 9 May 2026 16:24:01 +0200
+Subject: [PATCH RFC v2] media: hantro: attach dma_resv release fence at
+ device_run
+
+Opt the hantro driver into the new vb2 release-fence helper so its
+CAPTURE-side dmabufs carry a real producer fence that wayland
+compositors and other implicit-sync consumers can wait on, instead
+of the dma_buf core's stub fence.
+
+Attach point is m2m device_run, immediately after
+v4l2_m2m_buf_copy_metadata() and before ctx->codec_ops->run().
+Per Nicolas Dufresne's v1 review (lore.kernel.org/linux-media/
+3d8deeb15581b754e4c061d4c4a13657aa08bc3c.camel@ndufresne.ca/),
+this satisfies the dma_fence finite-time contract: the m2m core
+has committed to running the job by this point, codec_ops->run
+either kicks the HW (decode-complete signals the fence via
+vb2_buffer_done) or fails immediately (job_finish with
+VB2_BUF_STATE_ERROR signals with -EIO). PM and clocks are already
+up by this point, so no allocation context restrictions.
+
+The CAPTURE queue is opted in with supports_release_fences=true at
+queue_init.
+
+Userspace consumers that import hantro CAPTURE dmabufs and wait on
+their implicit-sync fence (Wayland zwp_linux_dmabuf_v1 +
+panfrost EGL_LINUX_DMA_BUF_EXT) now wait on a real fence
+representing the producer's actual completion, fixing green-frame
+corruption observed on RK3566 PineTab2 + Mali-G52 panfrost (the
+GPU was sampling zero pages because the dmabuf's implicit fence
+was the dma_buf core's pre-signalled stub).
+
+Validated end-to-end on PineTab2 (RK3566 / hantro G1 / Mali-G52
+mainline panfrost): 30s of bbb_1080p30 H.264 stateless decode +
+zero-copy panfrost EGL import via dmabuf-wayland (mpv 0.41 +
+KWin 6.6.4 + Mesa panfrost 26.0.5) renders correctly with no
+green-frame corruption and no PROVE_LOCKING splats.
+
+Cc: Ezequiel Garcia <ezequiel@vanguardiasur.com.ar>
+Cc: Philipp Zabel <p.zabel@pengutronix.de>
+Cc: Nicolas Dufresne <nicolas@ndufresne.ca>
+Cc: linux-media@vger.kernel.org
+Cc: linux-rockchip@lists.infradead.org
+Signed-off-by: Markus Fritsche <mfritsche@reauktion.de>
+---
+ .../media/platform/verisilicon/hantro_drv.c   | 23 +++++++++++++++++++
+ 1 file changed, 23 insertions(+)
+
+diff --git a/drivers/media/platform/verisilicon/hantro_drv.c b/drivers/media/platform/verisilicon/hantro_drv.c
+index 2e81877f6..6a66c47ed 100644
+--- a/drivers/media/platform/verisilicon/hantro_drv.c
++++ b/drivers/media/platform/verisilicon/hantro_drv.c
+@@ -186,6 +186,22 @@ static void device_run(void *priv)
+ 
+ 	v4l2_m2m_buf_copy_metadata(src, dst);
+ 
++	/*
++	 * Attach a producer fence on the CAPTURE-side dmabuf so userspace
++	 * importers (e.g. Wayland compositors) get spec-clean implicit-sync
++	 * semantics. Called from device_run rather than buf_queue: the
++	 * dma_fence finite-time contract requires that once a fence is
++	 * published, the producer must signal it in finite time. By the
++	 * time we reach device_run, the m2m core has committed to running
++	 * this job, and the next hop (codec_ops->run) either kicks the HW
++	 * (decode-complete signals the fence via vb2_buffer_done) or
++	 * fails immediately (job_finish with VB2_BUF_STATE_ERROR signals
++	 * the fence with -EIO). Either path resolves the fence in finite
++	 * time. Best-effort: a NOMEM here means we lose implicit-sync
++	 * precision for this frame, no functional regression.
++	 */
++	(void)vb2_buffer_attach_release_fence(&dst->vb2_buf);
++
+ 	if (ctx->codec_ops->run(ctx))
+ 		goto err_cancel_job;
+ 
+@@ -249,6 +265,13 @@ queue_init(void *priv, struct vb2_queue *src_vq, struct vb2_queue *dst_vq)
+ 	dst_vq->lock = &ctx->dev->vpu_mutex;
+ 	dst_vq->dev = ctx->dev->v4l2_dev.dev;
+ 
++	/*
++	 * Opt the CAPTURE queue into vb2 release-fence publishing.
++	 * No-op unless CONFIG_VIDEOBUF2_RELEASE_FENCES=y; runtime cost
++	 * is one extra fence allocation + dma_resv update per device_run.
++	 */
++	dst_vq->supports_release_fences = true;
++
+ 	return vb2_queue_init(dst_vq);
+ }
+ 
+-- 
+2.53.0
+

arch/linux-fresnel-fourier/0006-media-rockchip-rga-attach-dma_resv-release-fence-at-.patch

@@ -0,0 +1,117 @@
+From 2c63a63bf65739763051dc4ce7ce2ffaf2d514c4 Mon Sep 17 00:00:00 2001
+In-Reply-To: <20260429195306.239666-1-mfritsche@reauktion.de>
+References: <20260429195306.239666-1-mfritsche@reauktion.de>
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Sat, 9 May 2026 16:50:51 +0200
+Subject: [PATCH RFC v2] media: rockchip-rga: attach dma_resv release fence at
+ device_run
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+Opt the rockchip-rga driver into the new vb2 release-fence helper.
+
+Same shape as the hantro patch: attach a producer fence on the
+CAPTURE-side dmabuf at m2m device_run, signalled by
+vb2_buffer_done() when RGA completes the m2m operation.
+
+Differs from hantro in one mechanical detail: rga's device_run
+wraps the entire body in spin_lock_irqsave(&rga->ctrl_lock). Our
+helper calls dma_resv_lock(), which is sleepable, so the
+buffer-fetch + fence-attach sequence has to run above the spinlock.
+Restructure device_run so:
+
+  - v4l2_m2m_next_src_buf / next_dst_buf,
+  - src->sequence increment,
+  - vb2_buffer_attach_release_fence(&dst->vb2_buf)
+
+run before spin_lock_irqsave; only the rga->curr assignment and
+rga_hw_start() (the actual HW kick) remain inside the spinlock.
+
+This is safe under the m2m-job ownership model: by the time
+device_run is called, the m2m core has selected this context and
+serializes one device_run per context, so v4l2_m2m_next_*_buf
+returns stable pointers until the corresponding *_buf_remove in
+rga_isr. ctrl_lock was previously protecting per-device state
+(rga->curr) and the HW register access, neither of which depends on
+the buffer-fetch happening inside the lock.
+
+The CAPTURE queue is opted in with supports_release_fences=true at
+queue_init.
+
+Userspace consumers of RGA-produced dmabufs (image-processing
+pipelines, screen-rotation servers, gstreamer flows on Rockchip
+boards) get spec-clean implicit-sync semantics, matching what
+hantro does in the previous patch in this series.
+
+Sven Püschel's ongoing "media: platform: rga: Add RGA3 support"
+v5 series (linux-rockchip 2026-04-28) restructures rga.c
+substantially. If that lands first, the device_run restructure
+here will need a rebase against the new shape; the locking story
+itself is invariant.
+
+Cc: Jacob Chen <jacob-chen@iotwrt.com>
+Cc: Ezequiel Garcia <ezequiel@vanguardiasur.com.ar>
+Cc: Sven Püschel <s.pueschel@pengutronix.de>
+Cc: Heiko Stuebner <heiko@sntech.de>
+Cc: Hans Verkuil <hverkuil@xs4all.nl>
+Cc: linux-media@vger.kernel.org
+Cc: linux-rockchip@lists.infradead.org
+Signed-off-by: Markus Fritsche <mfritsche@reauktion.de>
+---
+ drivers/media/platform/rockchip/rga/rga.c | 27 +++++++++++++++++++----
+ 1 file changed, 23 insertions(+), 4 deletions(-)
+
+diff --git a/drivers/media/platform/rockchip/rga/rga.c b/drivers/media/platform/rockchip/rga/rga.c
+index fea63b94c..03030c7ea 100644
+--- a/drivers/media/platform/rockchip/rga/rga.c
++++ b/drivers/media/platform/rockchip/rga/rga.c
+@@ -38,15 +38,28 @@ static void device_run(void *prv)
+ 	struct vb2_v4l2_buffer *src, *dst;
+ 	unsigned long flags;
+ 
+-	spin_lock_irqsave(&rga->ctrl_lock, flags);
+-
+-	rga->curr = ctx;
+-
++	/*
++	 * Fetch the next-job buffers and (best-effort) attach a producer
++	 * fence on CAPTURE before taking ctrl_lock below.
++	 * vb2_buffer_attach_release_fence() takes dma_resv_lock, which is
++	 * sleepable; ctrl_lock is taken with spin_lock_irqsave so any
++	 * sleepable call must happen above it. Buffer ownership is
++	 * already committed at this point: the m2m core has selected
++	 * this context for device_run and serializes one device_run per
++	 * context, so v4l2_m2m_next_*_buf returns stable pointers until
++	 * the corresponding *_buf_remove in rga_isr.
++	 */
+ 	src = v4l2_m2m_next_src_buf(ctx->fh.m2m_ctx);
+ 	src->sequence = ctx->osequence++;
+ 
+ 	dst = v4l2_m2m_next_dst_buf(ctx->fh.m2m_ctx);
+ 
++	(void)vb2_buffer_attach_release_fence(&dst->vb2_buf);
++
++	spin_lock_irqsave(&rga->ctrl_lock, flags);
++
++	rga->curr = ctx;
++
+ 	rga_hw_start(rga, vb_to_rga(src), vb_to_rga(dst));
+ 
+ 	spin_unlock_irqrestore(&rga->ctrl_lock, flags);
+@@ -123,6 +136,12 @@ queue_init(void *priv, struct vb2_queue *src_vq, struct vb2_queue *dst_vq)
+ 	dst_vq->lock = &ctx->rga->mutex;
+ 	dst_vq->dev = ctx->rga->v4l2_dev.dev;
+ 
++	/*
++	 * Opt the CAPTURE queue into vb2 release-fence publishing.
++	 * Compile-time gated by CONFIG_VIDEOBUF2_RELEASE_FENCES.
++	 */
++	dst_vq->supports_release_fences = true;
++
+ 	return vb2_queue_init(dst_vq);
+ }
+ 
+-- 
+2.53.0
+

arch/linux-fresnel-fourier/PKGBUILD

@@ -10,8 +10,8 @@
 pkgbase=linux-fresnel-fourier
 pkgname=("$pkgbase" "$pkgbase-headers")
 pkgver=7.0
-pkgrel=1
-pkgdesc='Pinebook Pro kernel (mmind/linux-rockchip v7.0 + OC OPP + PBP DTS tweaks)'
+pkgrel=14
+pkgdesc='Pinebook Pro kernel (mmind/linux-rockchip v7.0 + OC OPP + PBP DTS tweaks + vb2_dma_resv RFC v2)'
 arch=(aarch64)
 url='https://git.reauktion.de/marfrit/kernel-agent'
 license=(GPL-2.0-only)
@@ -23,15 +23,20 @@ makedepends=(
 options=('!strip')
 source=(
   "https://git.kernel.org/torvalds/t/linux-${pkgver}.tar.gz"
+  # board/pinebook-pro
   '0001-arm64-dts-rk3399-pinebook-pro-add-OC-OPP-tables-1704-2184.patch'
   '0002-arm64-dts-rk3399-pinebook-pro-enable-hdmi-sound.patch'
   '0003-arm64-dts-rk3399-pinebook-pro-spi1-max-freq-10MHz.patch'
+  # subsystem/media/videobuf2/dma-resv-release-fence (RFC v2, in kernel-agent)
+  '0004-media-videobuf2-add-opt-in-dma_resv-producer-fence-h.patch'
+  '0005-media-hantro-attach-dma_resv-release-fence-at-device.patch'
+  '0006-media-rockchip-rga-attach-dma_resv-release-fence-at-.patch'
   'config'  # snapshot of fresnel /usr/lib/modules/6.19.10-1-eos-arm/build/.config
   'linux-fresnel-fourier.preset'
   'extlinux-add.hook'
   'extlinux-add.sh'
 )
-sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
+sha256sums=('SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP' 'SKIP')
 
 _kernver=${pkgver}.0-fresnel-fourier
 _srcdir=linux-${pkgver}

arch/linux-pinetab2-danctnix-besser/.gitignore

@@ -0,0 +1,19 @@
+# makepkg outputs and intermediate state
+src/
+pkg/
+*.pkg.tar.zst
+*.pkg.tar.zst.sig
+*-build.log*
+*-prepare.log
+*-package_*.log
+*.tar.xz
+*.tar.sign
+*.patch.zst
+*.patch.zst.sig
+logpipe.*
+
+# build-time artifacts that must never be committed
+src/
+*.tar.xz
+*.tar.zst
+pkg/

arch/linux-pinetab2-danctnix-besser/PKGBUILD

@@ -0,0 +1,247 @@
+# Maintainer: Markus Fritsche <fritsche.markus@gmail.com>
+# Forked from: linux-pinetab2 by Danct12 <danct12@disroot.org>
+# Original Contributor: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
+#
+# linux-pinetab2-danctnix-besser: linux-pinetab2 + the BESser bes2600
+# driver patchset.  Soft-upstream fork of linux-pinetab2 — drop-in
+# replacement, same kernel version, only the bes2600 staging driver
+# differs.  See git.reauktion.de/marfrit/besser and
+# git.reauktion.de/marfrit/bes2600-dkms for full provenance.
+#
+# ─────────────────────────────────────────────────────────────────────
+# kernel-agent flow (effective pkgrel=4):
+#
+# The single 0001-bes2600-besser-kernel-agent-cumulative.patch is
+# regenerated by `ka-promote ohm` in
+# git.reauktion.de/marfrit/kernel-agent against fleet/ohm.yaml.  Its
+# b2sum below MUST match the cumulative.b2sum recorded in the
+# co-located manifest.lock.  Sibling manifest.lock pins the source-of-
+# truth manifest sha256 and per-include patch sha256s.
+#
+# To update for a new patch in the manifest:
+#   1. Edit kernel-agent:fleet/ohm.yaml (add/remove includes)
+#   2. `ka-promote ohm` → writes new cumulative.patch + manifest.lock
+#      to kernel-agent/build/ohm/v7.0-danctnix1/
+#   3. Copy both files into this directory (replacing existing ones)
+#   4. Update the b2sum in this PKGBUILD for the cumulative
+#   5. Bump pkgrel
+# ─────────────────────────────────────────────────────────────────────
+
+pkgbase=linux-pinetab2-danctnix-besser
+pkgver=7.0.danctnix1
+pkgrel=5
+pkgdesc='PineTab2 (BESser bes2600 driver patchset, kernel-agent managed)'
+_srcname=linux-pinetab2
+_srctag=v${pkgver%.*}-${pkgver##*.}
+arch=(aarch64)
+_url_git="https://codeberg.org/DanctNIX/${_srcname}"
+url="${_url_git}/commits/tag/$_srctag"
+license=(GPL-2.0-only)
+makedepends=(
+  bc
+  cpio
+  gettext
+  git
+  libelf
+  pahole
+  perl
+  python
+  tar
+  xz
+)
+options=(
+  !debug
+  !strip
+)
+source=(
+  https://cdn.kernel.org/pub/linux/kernel/v${pkgver%%.*}.x/linux-${pkgver%.*}.tar.{xz,sign}
+  ${_url_git}/releases/download/${_srctag}/${_srctag}.patch.zst{,.sig}
+  0001-bes2600-besser-kernel-agent-cumulative.patch
+  config
+)
+validpgpkeys=(
+  ABAF11C65A2970B130ABE3C479BE3E4300411886  # Linus Torvalds
+  647F28654894E3BD457199BE38DBBDC86092693E  # Greg Kroah-Hartman
+  F09A933C0FE0331E558CA4E166CAB7EAA45DD781  # Danct12
+)
+b2sums=('3d9795083c8938f80f480de0d10bfd9c525640e59d5c7f22983de3f12ee42c84c31be902cafb05579ddb1c32bac5ed06b0d4953f9705450be185bd2d9ab08f89'
+        'SKIP'
+        '71fe98221e802b315e54b4b10d3e8c8f376695a36bae3541d876e5776a37f3fa33c8f8dfa6e51fcbd6f5396add02e5166634165f2351836a0ea0453c172fe56c'
+        'SKIP'
+        '50397711a6a3ba522283685a9e7397aeed6663f353f7cba214d4bb88bc98516065b2fca9a36ce13c52644617879f69f39c5305e86db5d9fb25c4dae5434eb9c4'
+        '656a998ab40cb85ee4c00f087b071a91632a6c091da2c84b0f74236b51d2dea6e9db6886625f80ad81dc249d8494ec47cd79d6dd9ea4f5e44f3cde857f861e10')
+
+export KBUILD_BUILD_HOST=archlinux
+export KBUILD_BUILD_USER=$pkgbase
+export KBUILD_BUILD_TIMESTAMP="$(date -Ru${SOURCE_DATE_EPOCH:+d @$SOURCE_DATE_EPOCH})"
+
+prepare() {
+  cd linux-${pkgver%.*}
+
+  echo "Setting version..."
+  echo "-$pkgrel" > localversion.10-pkgrel
+  echo "${pkgbase#linux}" > localversion.20-pkgname
+
+  local src
+  for src in "${source[@]}"; do
+      src="${src%%::*}"
+      src="${src##*/}"
+      src="${src%.zst}"
+      [[ $src = *.patch ]] || continue
+      echo "Applying patch: $src..."
+      patch -Np1 < "../$src"
+  done
+
+  echo "Setting config..."
+  cp ../config .config
+  make olddefconfig
+  diff -u ../config .config || :
+
+  make -s kernelrelease > version
+  echo "Prepared $pkgbase version $(<version)"
+}
+
+build() {
+  cd linux-${pkgver%.*}
+  make DTC_FLAGS="-@" all
+  make -C tools/bpf/bpftool vmlinux.h feature-clang-bpf-co-re=1
+}
+
+_package() {
+  pkgdesc="The $pkgdesc kernel and modules"
+  depends=(
+    coreutils
+    kmod
+    mkinitcpio
+  )
+  optdepends=(
+    'wireless-regdb: to set the correct wireless channels of your country'
+    'linux-firmware: firmware images needed for some devices'
+  )
+  provides=(
+    KSMBD-MODULE
+    WIREGUARD-MODULE
+    "linux-pinetab2=$pkgver-$pkgrel"
+  )
+  conflicts=(linux-pinetab2)
+  replaces=(
+    wireguard-arch
+  )
+
+  cd linux-${pkgver%.*}
+  local modulesdir="$pkgdir/usr/lib/modules/$(<version)"
+
+  echo "Installing boot image..."
+  # systemd expects to find the kernel here to allow hibernation
+  # https://github.com/systemd/systemd/commit/edda44605f06a41fb86b7ab8128dcf99161d2344
+  install -Dm644 "$(make -s image_name)" "$modulesdir/vmlinuz"
+
+  # Used by mkinitcpio to name the kernel
+  echo "$pkgbase" | install -Dm644 /dev/stdin "$modulesdir/pkgbase"
+
+  echo "Installing modules..."
+  ZSTD_CLEVEL=19 make INSTALL_MOD_PATH="$pkgdir/usr" INSTALL_MOD_STRIP=1 \
+    DEPMOD=/doesnt/exist modules_install  # Suppress depmod
+
+  echo "Installing device trees..."
+  make INSTALL_DTBS_PATH="$pkgdir/boot/dtbs" dtbs_install
+
+  # Removing unnecessary device trees (keep only pinetab2 variants).
+  # find -delete is robust to nullglob/cwd quirks where a bash for-loop
+  # silently no-op'd in the makepkg environment.
+  find "$pkgdir"/boot/dtbs/rockchip/ -mindepth 1 -maxdepth 1 -type f \
+       ! -name 'rk3566-pinetab2-*' -delete
+
+  # remove build link
+  rm "$modulesdir"/build
+}
+
+_package-headers() {
+  pkgdesc="Headers and scripts for building modules for the $pkgdesc kernel"
+  depends=(pahole)
+
+  cd linux-${pkgver%.*}
+  local builddir="$pkgdir/usr/lib/modules/$(<version)/build"
+
+  echo "Installing build files..."
+  install -Dt "$builddir" -m644 .config Makefile Module.symvers System.map \
+    localversion.* version vmlinux tools/bpf/bpftool/vmlinux.h
+  install -Dt "$builddir/kernel" -m644 kernel/Makefile
+  install -Dt "$builddir/arch/arm64" -m644 arch/arm64/Makefile
+  cp -t "$builddir" -a scripts
+
+  # required when DEBUG_INFO_BTF_MODULES is enabled
+  install -Dt "$builddir/tools/bpf/resolve_btfids" tools/bpf/resolve_btfids/resolve_btfids
+
+  echo "Installing headers..."
+  cp -t "$builddir" -a include
+  cp -t "$builddir/arch/arm64" -a arch/arm64/include
+  install -Dt "$builddir/arch/arm64/kernel" -m644 arch/arm64/kernel/asm-offsets.s
+
+  install -Dt "$builddir/drivers/md" -m644 drivers/md/*.h
+  install -Dt "$builddir/net/mac80211" -m644 net/mac80211/*.h
+
+  # https://bugs.archlinux.org/task/13146
+  install -Dt "$builddir/drivers/media/i2c" -m644 drivers/media/i2c/msp3400-driver.h
+
+  # https://bugs.archlinux.org/task/20402
+  install -Dt "$builddir/drivers/media/usb/dvb-usb" -m644 drivers/media/usb/dvb-usb/*.h
+  install -Dt "$builddir/drivers/media/dvb-frontends" -m644 drivers/media/dvb-frontends/*.h
+  install -Dt "$builddir/drivers/media/tuners" -m644 drivers/media/tuners/*.h
+
+  # https://bugs.archlinux.org/task/71392
+  install -Dt "$builddir/drivers/iio/common/hid-sensors" -m644 drivers/iio/common/hid-sensors/*.h
+
+  echo "Installing KConfig files..."
+  find . -name 'Kconfig*' -exec install -Dm644 {} "$builddir/{}" \;
+
+  echo "Removing unneeded architectures..."
+  local arch
+  for arch in "$builddir"/arch/*/; do
+    [[ $arch = */arm64/ ]] && continue
+    echo "Removing $(basename "$arch")"
+    rm -r "$arch"
+  done
+
+  echo "Removing documentation..."
+  rm -r "$builddir/Documentation"
+
+  echo "Removing broken symlinks..."
+  find -L "$builddir" -type l -printf 'Removing %P\n' -delete
+
+  echo "Removing loose objects..."
+  find "$builddir" -type f -name '*.o' -printf 'Removing %P\n' -delete
+
+  echo "Stripping build tools..."
+  local file
+  while read -rd '' file; do
+    case "$(file -Sib "$file")" in
+      application/x-sharedlib\;*)      # Libraries (.so)
+        strip -v $STRIP_SHARED "$file" ;;
+      application/x-archive\;*)        # Libraries (.a)
+        strip -v $STRIP_STATIC "$file" ;;
+      application/x-executable\;*)     # Binaries
+        strip -v $STRIP_BINARIES "$file" ;;
+      application/x-pie-executable\;*) # Relocatable binaries
+        strip -v $STRIP_SHARED "$file" ;;
+    esac
+  done < <(find "$builddir" -type f -perm -u+x ! -name vmlinux -print0)
+
+  echo "Stripping vmlinux..."
+  strip -v $STRIP_STATIC "$builddir/vmlinux"
+
+  echo "Adding symlink..."
+  mkdir -p "$pkgdir/usr/src"
+  ln -sr "$builddir" "$pkgdir/usr/src/$pkgbase"
+}
+
+pkgname=(
+  "$pkgbase"
+  "$pkgbase-headers"
+)
+for _p in "${pkgname[@]}"; do
+  eval "package_$_p() {
+    $(declare -f "_package${_p#$pkgbase}")
+    _package${_p#$pkgbase}
+  }"
+done

arch/linux-pinetab2-danctnix-besser/README.md

@@ -0,0 +1,52 @@
+# linux-pinetab2-danctnix-besser
+
+DanctNIX PineTab2 kernel + the BESser bes2600 driver patchset.
+**kernel-agent managed** as of pkgrel=4 (2026-05-18).
+
+## TL;DR
+
+- **What**: drop-in replacement for `linux-pinetab2` with the BESser
+  bes2600 patchset applied.
+- **For**: PineTab2 (RK3566) with the on-board BES2600 SDIO WiFi/BT chip.
+- **Install**: `pacman -U linux-pinetab2-danctnix-besser-*.pkg.tar.zst
+  linux-pinetab2-danctnix-besser-headers-*.pkg.tar.zst`. It `provides`
+  and `conflicts` with `linux-pinetab2`, so it slots in cleanly.
+- **Source of truth for patches**:
+  [git.reauktion.de/marfrit/kernel-agent](https://git.reauktion.de/marfrit/kernel-agent)
+  `fleet/ohm.yaml`.
+
+## How the patch cumulative is built
+
+1. `kernel-agent/fleet/ohm.yaml` lists the patch series-dir includes
+   for this kernel.
+2. `kernel-agent/bin/ka-promote ohm` walks those includes, concatenates
+   them into a single `cumulative.patch`, and emits a `manifest.lock`
+   recording per-patch sha256s and the cumulative b2sum.
+3. The cumulative is dropped into this directory as
+   `0001-bes2600-besser-kernel-agent-cumulative.patch`, the b2sum in
+   `PKGBUILD` is updated to match, and `pkgrel` is bumped.
+4. `manifest.lock` is committed alongside the cumulative for an
+   audit trail (which patches went in, in what order).
+
+## pkgrel history
+
+| pkgrel | Date       | Notes                                           |
+|--------|------------|-------------------------------------------------|
+| 1–3    | 2026-05-08…05-18 | hand-managed in `marfrit/besser/danctnix-besser-pkgbuild/` (DEPRECATED) |
+| 4      | 2026-05-18 | first kernel-agent-managed release; same source contents as the hand-managed pkgrel=3 (Patch I + SCS workaround folded into single cumulative); fixes pkgrel=3's duplicated `0003-arm64-...patch` source-array bug |
+
+## Known caveats
+
+- **`CONFIG_SHADOW_CALL_STACK=n`** in the config: GCC 15.2.1 fails to
+  build `arch/arm64/lib/xor-neon.c` with `arm_neon.h` `#pragma GCC
+  target()` blocks and `-fsanitize=shadow-call-stack`. See
+  `reference_arm64_scs_arm_neon_gcc15` memory and the Makefile
+  workaround patch included in the cumulative. Track upstream GCC
+  resolution before re-enabling.
+- The `cumulative-c5x-danctnix` series-dir in kernel-agent is currently
+  a single-file interim (the working hand-curated cumulative) because
+  the per-series mirrors created in `kernel-agent#17` use DKMS-style
+  root paths instead of in-tree staging paths and do not apply
+  cleanly. Per-series reconstruction is a separate kernel-agent
+  follow-up. The b2sum integrity check in `PKGBUILD` still pins the
+  exact bytes shipped.

arch/linux-pinetab2-danctnix-besser/manifest.lock

@@ -0,0 +1,37 @@
+baseline:
+  ref: v7.0-danctnix1
+  tree: DanctNIX/linux-pinetab2
+  upstream_compat: linux-7.0
+  url: https://codeberg.org/DanctNIX/linux-pinetab2
+cumulative:
+  b2sum: 50397711a6a3ba522283685a9e7397aeed6663f353f7cba214d4bb88bc98516065b2fca9a36ce13c52644617879f69f39c5305e86db5d9fb25c4dae5434eb9c4
+  path: cumulative.patch
+  size: 162716
+generated_at: '2026-05-18T17:16:06.455474+00:00'
+host: ohm
+ka_promote_version: 1
+manifest:
+  path: fleet/ohm.yaml
+  sha256: da59ac2c965e5ad9c5004a115b10a37abf47ed3ecc8b7f5ab426470d2ee7b442
+resolved_patches:
+- apply_order: 1
+  from_series: true
+  include: driver/bes2600/cumulative-c5x-danctnix/0001-bes2600-besser-cumulative-series.patch
+  sha256: e477a170567487fef84fe13be5b0a1f0498247ff1f201000d0085a2e49ff9026
+  size: 148149
+- apply_order: 2
+  from_series: true
+  include: driver/bes2600/scan-filter-5ghz-danctnix/0001-bes2600-filter-5ghz-scan-and-allow-single-channel.patch
+  sha256: 31e67569e00daead0784214aced1e077d3270cf1407baa0b330d474e17ec3931
+  size: 7735
+- apply_order: 3
+  from_series: true
+  include: arch/arm64/xor-neon-ffixed-x18-scs-build-fix-danctnix/0001-arm64-xor-neon-ffixed-x18-build-fix.patch
+  sha256: a49c50f0ebffc499970c24908b832c3e61c96ed87de35b3a82178eff587f94f1
+  size: 1574
+- apply_order: 4
+  from_series: true
+  include: driver/bes2600/queue-pending-record-lock-bh-danctnix/0001-bes2600-take-pending-record-lock-with-bh.patch
+  sha256: 089862e5f6da5783ed0db979144e4fa07cff7f743809a0bebd715c75a3bb8eb5
+  size: 5258
+schema_version: 1

arch/lmcp/PKGBUILD

@@ -3,15 +3,18 @@
 # Source of truth: git.reauktion.de/marfrit/lmcp
 
 pkgname=lmcp
-pkgver=0.5.4
+pkgver=1.2.1
 pkgrel=1
 pkgdesc="Lightweight MCP (Model Context Protocol) server in pure Lua"
 arch=('any')
 url="https://git.reauktion.de/marfrit/lmcp"
 license=('MIT')
 depends=('lua' 'lua-socket')
-source=("${pkgname}-${pkgver}.tar.gz::https://git.reauktion.de/marfrit/lmcp/archive/v${pkgver}.tar.gz")
-sha256sums=('af72b8c1d88255456b75d2c53cd5c451a8923417e5498ef31858539397e09caf')
+# The _tag back-translation handles both clean releases (no '_') and
+# pre-release pkgvers (e.g. 1.2.0_rc1 → v1.2.0-rc1).
+_tag="v${pkgver//_/-}"
+source=("${pkgname}-${pkgver}.tar.gz::https://git.reauktion.de/marfrit/lmcp/archive/${_tag}.tar.gz")
+sha256sums=('bf9cce1a84c66b1b74c5aec923c5960d60ae33c221afc8d47ce0d74b8f7ee609')
 
 package() {
     cd "${pkgname}"

arch/mesa-panvk-bifrost-video/0001-panvk-expose-robustness2-nullDescriptor-bifrost.patch

@@ -0,0 +1,57 @@
+From: claude-noether (on behalf of mfritsche)
+Date: 2026-05-19
+Subject: panvk: expose VK_KHR/EXT_robustness2 + nullDescriptor on Bifrost (PAN_ARCH 6/7)
+
+Without this, Mesa's Zink driver refuses to use PanVk-Bifrost as its Vulkan
+backend, falling back silently to llvmpipe (software rasterizer) for all
+GL-via-Zink on Bifrost SBCs. That defeats the entire purpose of having a
+Vulkan driver on Bifrost — GL acceleration via Zink is the most natural
+near-term consumer.
+
+panvk_vX_nir_lower_descriptors.c:1309 and panvk_vX_shader.c:1355 already
+plumb dev->vk.enabled_features.nullDescriptor arch-agnostically — the gate
+at panvk_vX_physical_device.c was set conservatively when Bifrost was
+unmaintained, not because of hardware incapability.
+
+iter1–7 of the panvk-bifrost campaign proved fundamental driver functions
+on Mali-G52 r1 MC1 (PAN_ARCH=7). This patch is the iter8 follow-up.
+
+robustBufferAccess2 and robustImageAccess2 are NOT flipped — they're
+independent rb2 features Zink doesn't require, gated differently
+(robustBufferAccess2 = PAN_ARCH >= 11, robustImageAccess2 = false), and
+out of scope for iter8.
+
+---
+ src/panfrost/vulkan/panvk_vX_physical_device.c | 6 +++---
+ 1 file changed, 3 insertions(+), 3 deletions(-)
+
+diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
+--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
+@@ -91,7 +91,7 @@ get_device_extensions(const struct panvk_physical_device *device,
+       .KHR_pipeline_binary = true,
+       .KHR_pipeline_executable_properties = true,
+       .KHR_pipeline_library = true,
+-      .KHR_robustness2 = PAN_ARCH >= 10,
++      .KHR_robustness2 = true,
+       .KHR_sampler_mirror_clamp_to_edge = true,
+       .KHR_sampler_ycbcr_conversion = true,
+       .KHR_separate_depth_stencil_layouts = true,
+@@ -168,7 +168,7 @@ get_device_extensions(const struct panvk_physical_device *device,
+       .EXT_queue_family_foreign = true,
+       .EXT_robustness = pan_arch(device->kmod.dev->props.gpu_id) >= 9,
+       .EXT_image_robustness = true,
+-      .EXT_robustness2 = PAN_ARCH >= 10,
++      .EXT_robustness2 = true,
+       .EXT_sampler_filter_minmax = PAN_ARCH >= 10,
+       .EXT_scalar_block_layout = true,
+       .EXT_separate_stencil_usage = true,
+@@ -493,7 +493,7 @@ get_device_features(const struct panvk_physical_device *device,
+       /* VK_KHR_robustness2 */
+       .robustBufferAccess2 = PAN_ARCH >= 11,
+       .robustImageAccess2 = false,
+-      .nullDescriptor = PAN_ARCH >= 10,
++      .nullDescriptor = true,
+
+       /* VK_KHR_shader_clock */
+       .shaderSubgroupClock = device->kmod.dev->props.gpu_can_query_timestamp,

arch/mesa-panvk-bifrost-video/0002-panvk-expose-vulkan-1.1-1.2-on-bifrost.patch

@@ -0,0 +1,47 @@
+From: claude-noether (on behalf of mfritsche)
+Date: 2026-05-20
+Subject: panvk: expose Vulkan 1.1 + 1.2 on Bifrost (PAN_ARCH 6/7)
+
+ANGLE (Chromium's GL stack) requires apiVersion >= 1.1 to initialize. Without
+this, Brave / Chromium's GPU process fails at GL info collection:
+
+  vk_renderer.cpp:2659 (initialize): ANGLE Requires a minimum Vulkan device
+                                     version of 1.1
+  Display::initialize error 0: Internal Vulkan error (-9): The requested
+                               version of Vulkan is not supported by the driver
+
+Stack-up with iter8's robustness2 patch enables ANGLE → PanVk-Bifrost →
+Skia (via --enable-features=Vulkan) on Bifrost SBCs.
+
+PanVk-Bifrost already supports the bulk of 1.1-promoted features as extensions
+(multiview, maintenance1-3, descriptor update template, 16-bit storage,
+descriptor update template, sampler ycbcr, variable pointers, etc. — all
+visible in iter0 vulkaninfo). The version bump primarily bundles them.
+
+Risk: Vulkan 1.1 has features beyond what iter1–7 exercised (protected memory,
+full subgroup ops). Specific app failures will be characterizable.
+
+1.2 is also flipped — Brave's Vulkan path may want descriptor indexing,
+buffer device address, etc. (all listed in iter0 vulkaninfo as supported
+extensions, just gated as 1.0-with-extensions, not 1.2-core).
+
+---
+ src/panfrost/vulkan/panvk_vX_physical_device.c | 4 ++--
+ 1 file changed, 2 insertions(+), 2 deletions(-)
+
+diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
+--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
+@@ -38,8 +38,8 @@ get_device_extensions(const struct panvk_physical_device *device,
+                       struct vk_device_extension_table *ext)
+ {
+    *ext = (struct vk_device_extension_table){
+-      .KHR_8bit_storage = true,
+-      .KHR_16bit_storage = true,
+-      bool has_vk1_1 = PAN_ARCH >= 10;
+-      bool has_vk1_2 = PAN_ARCH >= 10;
++      .KHR_8bit_storage = true,
++      .KHR_16bit_storage = true,
++      bool has_vk1_1 = true;
++      bool has_vk1_2 = true;
+       *ext = (struct vk_device_extension_table){

arch/mesa-panvk-bifrost-video/0003-panvk-bifrost-vk-ext-transform-feedback.patch

@@ -0,0 +1,328 @@
+--- a/src/panfrost/vulkan/panvk_shader.h	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/panvk_shader.h	2026-05-20 18:52:53.312698258 +0200
+@@ -150,6 +150,10 @@
+    struct {
+ #if PAN_ARCH < 9
+       int32_t raw_vertex_offset;
++      uint32_t num_vertices;       /* iter13: XFB needs per-draw vertex count */
++      /* aligned_u64 attribute below inserts the 4-byte alignment gap
++       * after num_vertices automatically — no explicit pad needed. */
++      aligned_u64 xfb_address[4];  /* iter13: 4 transform feedback buffer base addresses */
+ #endif
+       int32_t first_vertex;
+       int32_t base_instance;
+--- a/src/panfrost/vulkan/panvk_vX_physical_device.c	2026-05-20 19:09:29.711145446 +0200
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c	2026-05-20 18:52:54.832720445 +0200
+@@ -169,6 +169,7 @@
+       .EXT_provoking_vertex = true,
+       .EXT_queue_family_foreign = true,
+       .EXT_robustness2 = true,
++      .EXT_transform_feedback = PAN_ARCH < 9,   /* iter13: JM-class only for now */
+       .EXT_sampler_filter_minmax = PAN_ARCH >= 10,
+       .EXT_scalar_block_layout = true,
+       .EXT_separate_stencil_usage = true,
+@@ -495,6 +496,10 @@
+       .robustImageAccess2 = false,
+       .nullDescriptor = true,
+ 
++      /* VK_EXT_transform_feedback (iter13) */
++      .transformFeedback = PAN_ARCH < 9,
++      .geometryStreams = false,
++
+       /* VK_KHR_shader_clock */
+       .shaderSubgroupClock = device->kmod.dev->props.gpu_can_query_timestamp,
+       .shaderDeviceClock = device->kmod.dev->props.timestamp_device_coherent,
+@@ -1020,6 +1025,18 @@
+       .robustStorageBufferAccessSizeAlignment = 1,
+       .robustUniformBufferAccessSizeAlignment = 1,
+ 
++      /* VK_EXT_transform_feedback (iter13) */
++      .maxTransformFeedbackStreams = 1,
++      .maxTransformFeedbackBuffers = 4,
++      .maxTransformFeedbackBufferSize = UINT32_MAX,
++      .maxTransformFeedbackStreamDataSize = 512,
++      .maxTransformFeedbackBufferDataSize = 512,
++      .maxTransformFeedbackBufferDataStride = 2048,
++      .transformFeedbackQueries = false,
++      .transformFeedbackStreamsLinesTriangles = false,
++      .transformFeedbackRasterizationStreamSelect = false,
++      .transformFeedbackDraw = false,
++
+       /* VK_EXT_shader_object */
+       /* We do not currently support VK_EXT_shader_object but this is used
+        * internally by vk_shader
+--- a/src/panfrost/vulkan/panvk_vX_shader.c	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/panvk_vX_shader.c	2026-05-20 18:52:56.556745611 +0200
+@@ -21,6 +21,7 @@
+ #include "panvk_physical_device.h"
+ #include "panvk_sampler.h"
+ #include "panvk_shader.h"
++#include "pan_nir.h"   /* iter13: pan_nir_lower_xfb */
+ 
+ #include "spirv/nir_spirv.h"
+ #include "util/memstream.h"
+@@ -100,6 +101,20 @@
+    case nir_intrinsic_load_raw_vertex_offset_pan:
+       val = load_sysval(b, graphics, bit_size, vs.raw_vertex_offset);
+       break;
++   case nir_intrinsic_load_num_vertices:    /* iter13: XFB index calc */
++      val = load_sysval(b, graphics, bit_size, vs.num_vertices);
++      break;
++   case nir_intrinsic_load_xfb_address: {   /* iter13: XFB buffer N base address */
++      unsigned idx = nir_intrinsic_base(intr);
++      switch (idx) {
++      case 0: val = load_sysval(b, graphics, bit_size, vs.xfb_address[0]); break;
++      case 1: val = load_sysval(b, graphics, bit_size, vs.xfb_address[1]); break;
++      case 2: val = load_sysval(b, graphics, bit_size, vs.xfb_address[2]); break;
++      case 3: val = load_sysval(b, graphics, bit_size, vs.xfb_address[3]); break;
++      default: return false;
++      }
++      break;
++   }
+    case nir_intrinsic_load_layer_id:
+       assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+       val = load_sysval(b, graphics, bit_size, layer_id);
+@@ -457,6 +472,7 @@
+             core_max_id);
+ 
+    pan_preprocess_nir(nir, pdev->kmod.dev->props.gpu_id);
++
+ }
+ 
+ static void
+@@ -870,6 +886,18 @@
+             nir_var_shader_in | nir_var_shader_out, UINT32_MAX);
+    NIR_PASS(_, nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
+             glsl_type_size, nir_lower_io_use_interpolated_input_intrinsics);
++
++#if PAN_ARCH < 9
++   /* iter13: VK_EXT_transform_feedback — runs AFTER nir_lower_io so that
++    * shader outputs are now store_output intrinsics that pan_nir_lower_xfb
++    * can rewrite to nir_store_global+nir_load_xfb_address. */
++   if (nir->info.stage == MESA_SHADER_VERTEX &&
++       nir->info.has_transform_feedback_varyings) {
++      NIR_PASS(_, nir, nir_opt_constant_folding);
++      NIR_PASS(_, nir, nir_io_add_intrinsic_xfb_info);
++      NIR_PASS(_, nir, pan_nir_lower_xfb);
++   }
++#endif
+ }
+ 
+ static VkResult
+@@ -1288,6 +1316,9 @@
+       .view_mask = (state && state->rp) ? state->rp->view_mask : 0,
+       .robust2_modes = robust2_modes,
+       .robust_descriptors = dev->vk.enabled_features.nullDescriptor,
++      /* iter13: XFB shaders must disable IDVS (matches Panfrost-Gallium). */
++      .no_idvs = (info->stage == MESA_SHADER_VERTEX) &&
++                 info->nir->info.has_transform_feedback_varyings,
+    };
+ 
+    switch (info->stage) {
+--- a/src/panfrost/vulkan/panvk_cmd_draw.h	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/panvk_cmd_draw.h	2026-05-20 18:52:57.748763011 +0200
+@@ -135,6 +135,19 @@
+    struct panvk_graphics_sysvals sysvals;
+ 
+ #if PAN_ARCH < 9
++   /* iter13: VK_EXT_transform_feedback state (JM-class only for now). */
++   struct {
++      bool active;
++      uint32_t buffer_count;
++      struct {
++         uint64_t addr;
++         uint64_t offset;
++         uint64_t size;
++      } buffers[4];
++   } xfb;
++#endif
++
++#if PAN_ARCH < 9
+    struct panvk_shader_link link;
+ #endif
+ 
+--- a/src/panfrost/vulkan/panvk_vX_cmd_draw.c	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/panvk_vX_cmd_draw.c	2026-05-20 19:10:23.031919662 +0200
+@@ -10,6 +10,7 @@
+ #include "panvk_entrypoints.h"
+ 
+ #include "pan_desc.h"
++#include "pan_compiler.h"   /* PAN_SHADER_OOB_ADDRESS */
+ #include "pan_util.h"
+ 
+ static void
+@@ -722,6 +723,35 @@
+    set_gfx_sysval(cmdbuf, dirty_sysvals, vs.raw_vertex_offset,
+                   info->vertex.raw_offset);
+    set_gfx_sysval(cmdbuf, dirty_sysvals, layer_id, info->layer_id);
++
++   /* iter13: VK_EXT_transform_feedback sysvals — always set (per draw),
++    * reflect bound XFB state. set_gfx_sysval is a no-op if value unchanged. */
++   set_gfx_sysval(cmdbuf, dirty_sysvals, vs.num_vertices, info->vertex.count);
++   {
++      const struct panvk_cmd_graphics_state *_gfx = &cmdbuf->state.gfx;
++      /* iter13: default each XFB buffer address to PAN_SHADER_OOB_ADDRESS
++       * (= 1<<63). This is the Panfrost-Gallium memory-sink idiom — the
++       * Bifrost MMU silently discards stores to this address, so a pipeline
++       * with XFB outputs used in a non-XFB draw (or in an XFB draw with
++       * fewer bound buffers than the shader declares) is safe instead of
++       * faulting. See gallium/drivers/panfrost/pan_cmdstream.c PAN_SYSVAL_XFB. */
++      uint64_t _xa0 = PAN_SHADER_OOB_ADDRESS, _xa1 = PAN_SHADER_OOB_ADDRESS,
++               _xa2 = PAN_SHADER_OOB_ADDRESS, _xa3 = PAN_SHADER_OOB_ADDRESS;
++      if (_gfx->xfb.active) {
++         if (_gfx->xfb.buffer_count > 0 && _gfx->xfb.buffers[0].addr)
++            _xa0 = _gfx->xfb.buffers[0].addr + _gfx->xfb.buffers[0].offset;
++         if (_gfx->xfb.buffer_count > 1 && _gfx->xfb.buffers[1].addr)
++            _xa1 = _gfx->xfb.buffers[1].addr + _gfx->xfb.buffers[1].offset;
++         if (_gfx->xfb.buffer_count > 2 && _gfx->xfb.buffers[2].addr)
++            _xa2 = _gfx->xfb.buffers[2].addr + _gfx->xfb.buffers[2].offset;
++         if (_gfx->xfb.buffer_count > 3 && _gfx->xfb.buffers[3].addr)
++            _xa3 = _gfx->xfb.buffers[3].addr + _gfx->xfb.buffers[3].offset;
++      }
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[0], _xa0);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[1], _xa1);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[2], _xa2);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[3], _xa3);
++   }
+ #endif
+ 
+    if (dyn_gfx_state_dirty(cmdbuf, CB_BLEND_CONSTANTS)) {
+--- a/src/panfrost/vulkan/meson.build	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/meson.build	2026-05-20 18:53:04.484861338 +0200
+@@ -73,6 +73,7 @@
+ jm_inc_dir = ['jm']
+ jm_files = [
+   'jm/panvk_vX_bind_queue.c',
++  'jm/panvk_vX_cmd_xfb.c',   # iter13
+   'jm/panvk_vX_cmd_buffer.c',
+   'jm/panvk_vX_cmd_dispatch.c',
+   'jm/panvk_vX_cmd_draw.c',
+--- a/src/panfrost/vulkan/jm/panvk_vX_cmd_buffer.c	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/jm/panvk_vX_cmd_buffer.c	2026-05-20 19:10:26.163965149 +0200
+@@ -473,5 +473,12 @@
+ 
+    vk_command_buffer_begin(&cmdbuf->vk, pBeginInfo);
+ 
++#if PAN_ARCH < 9
++   /* iter13: clear XFB state on Begin so a reused command buffer does not
++    * inherit stale xfb.buffer_count / xfb.active / xfb.buffers[] from a
++    * prior recording. */
++   memset(&cmdbuf->state.gfx.xfb, 0, sizeof(cmdbuf->state.gfx.xfb));
++#endif
++
+    return VK_SUCCESS;
+ }
+--- a/src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c	2026-05-18 12:50:53.067999996 +0200
++++ b/src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c	2026-05-20 19:10:27.175979847 +0200
+@@ -0,0 +1,111 @@
++/*
++ * Copyright © 2026 mfritsche / claude-noether
++ * SPDX-License-Identifier: MIT
++ *
++ * iter13: VK_EXT_transform_feedback command handlers for the JM
++ * architecture path (Bifrost v6/v7 + Valhall-JM v9).
++ *
++ * The runtime contract:
++ *   - vkCmdBindTransformFeedbackBuffersEXT: stash (gpu_addr, offset, size)
++ *     for each slot into cmdbuf->state.gfx.xfb.buffers[].
++ *   - vkCmdBeginTransformFeedbackEXT: set cmdbuf->state.gfx.xfb.active = true.
++ *     Mark sysvals dirty so the next draw re-emits vs.xfb_address[].
++ *   - vkCmdEndTransformFeedbackEXT: set active = false.
++ *
++ * Counter buffers (firstCounterBuffer/counterBufferCount/pCounterBuffers/
++ * pCounterBufferOffsets) are accepted by API but ignored — v1 doesn't
++ * support pause/resume. transformFeedbackDraw is advertised as false.
++ *
++ * Per-draw integration: jm/panvk_vX_cmd_draw.c reads cmdbuf->state.gfx.xfb
++ * and populates vs.xfb_address[i] for shader use. The pan_nir_lower_xfb
++ * pass in panvk_vX_shader.c emits nir_load_xfb_address(i) which lowers
++ * (via panvk_vX_shader.c sysval handler) to a load from the per-draw
++ * sysval push area.
++ */
++
++#include "vk_log.h"
++#include "util/log.h"
++
++#include "panvk_cmd_buffer.h"
++#include "panvk_cmd_draw.h"
++#include "panvk_buffer.h"
++#include "panvk_entrypoints.h"
++
++VKAPI_ATTR void VKAPI_CALL
++panvk_per_arch(CmdBindTransformFeedbackBuffersEXT)(
++   VkCommandBuffer commandBuffer,
++   uint32_t firstBinding,
++   uint32_t bindingCount,
++   const VkBuffer *pBuffers,
++   const VkDeviceSize *pOffsets,
++   const VkDeviceSize *pSizes)
++{
++   VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
++   struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
++
++   for (uint32_t i = 0; i < bindingCount; i++) {
++      uint32_t slot = firstBinding + i;
++      if (slot >= 4)
++         continue;
++
++      VK_FROM_HANDLE(panvk_buffer, buf, pBuffers[i]);
++      gfx->xfb.buffers[slot].addr = panvk_buffer_gpu_ptr(buf, 0);
++      gfx->xfb.buffers[slot].offset = pOffsets[i];
++      gfx->xfb.buffers[slot].size =
++         (pSizes != NULL && pSizes[i] != VK_WHOLE_SIZE)
++            ? pSizes[i]
++            : (buf->vk.size - pOffsets[i]);
++   }
++
++   if (firstBinding + bindingCount > gfx->xfb.buffer_count)
++      gfx->xfb.buffer_count = firstBinding + bindingCount;
++}
++
++VKAPI_ATTR void VKAPI_CALL
++panvk_per_arch(CmdBeginTransformFeedbackEXT)(
++   VkCommandBuffer commandBuffer,
++   uint32_t firstCounterBuffer,
++   uint32_t counterBufferCount,
++   const VkBuffer *pCounterBuffers,
++   const VkDeviceSize *pCounterBufferOffsets)
++{
++   VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
++   struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
++
++   /* Counter buffers ignored in v1 — see VkPhysicalDeviceTransformFeedback
++    * PropertiesEXT.transformFeedbackDraw = false in panvk_vX_physical_device.c.
++    * App is spec-compliant if it does not pass counter buffers (which our
++    * features advertisement allows), but warn loudly if it does so we do not
++    * silently produce wrong capture state. */
++   (void)firstCounterBuffer;
++   (void)pCounterBufferOffsets;
++   if (counterBufferCount > 0 && pCounterBuffers != NULL) {
++      mesa_logw("panvk: CmdBeginTransformFeedbackEXT: counter buffers not "
++                "implemented (transformFeedbackDraw=false); XFB resume will "
++                "restart at buffer offset 0");
++   }
++
++   gfx->xfb.active = true;
++   /* Per-draw set_gfx_sysval picks up the change automatically — no
++    * explicit dirty marking required (set_gfx_sysval uses memcmp +
++    * BITSET to detect state diffs and re-emit sysvals). */
++}
++
++VKAPI_ATTR void VKAPI_CALL
++panvk_per_arch(CmdEndTransformFeedbackEXT)(
++   VkCommandBuffer commandBuffer,
++   uint32_t firstCounterBuffer,
++   uint32_t counterBufferCount,
++   const VkBuffer *pCounterBuffers,
++   const VkDeviceSize *pCounterBufferOffsets)
++{
++   VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
++   struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
++
++   (void)firstCounterBuffer;
++   (void)counterBufferCount;
++   (void)pCounterBuffers;
++   (void)pCounterBufferOffsets;
++
++   gfx->xfb.active = false;
++}

arch/mesa-panvk-bifrost-video/0004-panvk-bifrost-xfb-primitive-decomposition.patch

@@ -0,0 +1,629 @@
+diff -urN a/src/panfrost/vulkan/meson.build b/src/panfrost/vulkan/meson.build
+--- a/src/panfrost/vulkan/meson.build	2026-05-21 14:04:02.529474145 +0200
++++ b/src/panfrost/vulkan/meson.build	2026-05-21 14:04:04.106755486 +0200
+@@ -123,6 +123,7 @@
+   'panvk_vX_nir_lower_input_attachment_loads.c',
+   'panvk_vX_sampler.c',
+   'panvk_vX_shader.c',
++  'panvk_vX_xfb_lower.c',
+   sha1_h,
+ ]
+ 
+diff -urN a/src/panfrost/vulkan/panvk_shader.h b/src/panfrost/vulkan/panvk_shader.h
+--- a/src/panfrost/vulkan/panvk_shader.h	2026-05-21 14:04:02.525251986 +0200
++++ b/src/panfrost/vulkan/panvk_shader.h	2026-05-21 14:04:04.084251800 +0200
+@@ -154,6 +154,8 @@
+       /* aligned_u64 attribute below inserts the 4-byte alignment gap
+        * after num_vertices automatically — no explicit pad needed. */
+       aligned_u64 xfb_address[4];  /* iter13: 4 transform feedback buffer base addresses */
++      uint32_t xfb_topology;       /* iter17: panvk_xfb_topology enum value */
++      uint32_t xfb_output_count;   /* iter17: per-instance output verts after decomp */
+ #endif
+       int32_t first_vertex;
+       int32_t base_instance;
+@@ -569,4 +571,76 @@
+    struct pan_compute_dim local_size, const void *bin_ptr, size_t bin_size,
+    struct panvk_shader **shader_out);
+ 
++
++#if PAN_ARCH < 9
++/* iter17: encoding for vs.xfb_topology sysval. Maps VkPrimitiveTopology values
++ * we need to distinguish at shader runtime for XFB capture. LIST topologies
++ * use the iter13 single-store fast path; non-LIST need per-vertex decomposition. */
++enum panvk_xfb_topology {
++   PANVK_XFB_TOPO_LIST            = 0,
++   PANVK_XFB_TOPO_LINE_STRIP      = 1,
++   PANVK_XFB_TOPO_TRI_STRIP       = 2,
++   PANVK_XFB_TOPO_TRI_FAN         = 3,
++   PANVK_XFB_TOPO_LINE_LIST_ADJ   = 4,
++   PANVK_XFB_TOPO_LINE_STRIP_ADJ  = 5,
++   PANVK_XFB_TOPO_TRI_LIST_ADJ    = 6,
++   PANVK_XFB_TOPO_TRI_STRIP_ADJ   = 7,
++};
++
++#include "panvk_macros.h"
++struct nir_shader;
++bool panvk_per_arch(nir_lower_xfb)(struct nir_shader *nir);
++
++/* Map VkPrimitiveTopology to panvk_xfb_topology enum (driver-side helper). */
++static inline uint32_t
++panvk_vk_topology_to_xfb_enum(VkPrimitiveTopology topo)
++{
++   switch (topo) {
++   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
++      return PANVK_XFB_TOPO_LINE_STRIP;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
++      return PANVK_XFB_TOPO_TRI_STRIP;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
++      return PANVK_XFB_TOPO_TRI_FAN;
++   case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
++      return PANVK_XFB_TOPO_LINE_LIST_ADJ;
++   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
++      return PANVK_XFB_TOPO_LINE_STRIP_ADJ;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
++      return PANVK_XFB_TOPO_TRI_LIST_ADJ;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
++      return PANVK_XFB_TOPO_TRI_STRIP_ADJ;
++   case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
++   case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
++   default:
++      return PANVK_XFB_TOPO_LIST;
++   }
++}
++
++/* Compute the per-instance output vertex count for a given (topology, input count). */
++static inline uint32_t
++panvk_xfb_output_count(VkPrimitiveTopology topo, uint32_t input_count)
++{
++   switch (topo) {
++   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
++      return input_count >= 1 ? 2u * (input_count - 1u) : 0u;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
++      return input_count >= 2 ? 3u * (input_count - 2u) : 0u;
++   case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
++      return (input_count / 4u) * 2u;
++   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
++      return input_count >= 3 ? 2u * (input_count - 3u) : 0u;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
++      return (input_count / 6u) * 3u;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
++      return input_count >= 6 ? 3u * (input_count / 2u - 2u) : 0u;
++   default:
++      return input_count;  /* LIST topologies: 1:1 mapping */
++   }
++}
++#endif
++
++
+ #endif
+diff -urN a/src/panfrost/vulkan/panvk_vX_cmd_draw.c b/src/panfrost/vulkan/panvk_vX_cmd_draw.c
+--- a/src/panfrost/vulkan/panvk_vX_cmd_draw.c	2026-05-21 14:04:02.528576354 +0200
++++ b/src/panfrost/vulkan/panvk_vX_cmd_draw.c	2026-05-21 14:04:04.091357598 +0200
+@@ -727,6 +727,20 @@
+    /* iter13: VK_EXT_transform_feedback sysvals — always set (per draw),
+     * reflect bound XFB state. set_gfx_sysval is a no-op if value unchanged. */
+    set_gfx_sysval(cmdbuf, dirty_sysvals, vs.num_vertices, info->vertex.count);
++
++   /* iter17: XFB primitive-decomposition sysvals.
++    * xfb_topology = enum value for the current bound topology.
++    * xfb_output_count = per-instance output vertex count after decomposition.
++    * For LIST topologies, output_count == input vertex count and the shader
++    * takes the iter13 single-store fast path. */
++   {
++      VkPrimitiveTopology vk_topo =
++         cmdbuf->vk.dynamic_graphics_state.ia.primitive_topology;
++      uint32_t topo_enum = panvk_vk_topology_to_xfb_enum(vk_topo);
++      uint32_t out_count = panvk_xfb_output_count(vk_topo, info->vertex.count);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_topology, topo_enum);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_output_count, out_count);
++   }
+    {
+       const struct panvk_cmd_graphics_state *_gfx = &cmdbuf->state.gfx;
+       /* iter13: default each XFB buffer address to PAN_SHADER_OOB_ADDRESS
+diff -urN a/src/panfrost/vulkan/panvk_vX_shader.c b/src/panfrost/vulkan/panvk_vX_shader.c
+--- a/src/panfrost/vulkan/panvk_vX_shader.c	2026-05-21 14:04:02.527576494 +0200
++++ b/src/panfrost/vulkan/panvk_vX_shader.c	2026-05-21 14:04:04.098356619 +0200
+@@ -895,7 +895,10 @@
+        nir->info.has_transform_feedback_varyings) {
+       NIR_PASS(_, nir, nir_opt_constant_folding);
+       NIR_PASS(_, nir, nir_io_add_intrinsic_xfb_info);
+-      NIR_PASS(_, nir, pan_nir_lower_xfb);
++      /* iter17: panvk-specific replacement for pan_nir_lower_xfb that handles
++       * primitive decomposition for non-LIST topologies. Single-store LIST
++       * fast path matches iter13 behavior. */
++      NIR_PASS(_, nir, panvk_per_arch(nir_lower_xfb));
+    }
+ #endif
+ }
+diff -urN a/src/panfrost/vulkan/panvk_vX_xfb_lower.c b/src/panfrost/vulkan/panvk_vX_xfb_lower.c
+--- a/src/panfrost/vulkan/panvk_vX_xfb_lower.c	1970-01-01 01:00:00.000000000 +0100
++++ b/src/panfrost/vulkan/panvk_vX_xfb_lower.c	2026-05-21 14:04:04.115354242 +0200
+@@ -0,0 +1,486 @@
++/*
++ * Copyright © 2026 mfritsche / claude-noether
++ * SPDX-License-Identifier: MIT
++ *
++ * iter17: panvk-specific replacement for pan_nir_lower_xfb that handles
++ * primitive decomposition for transform_feedback on non-LIST topologies
++ * (TRIANGLE_STRIP/FAN, LINE_STRIP, *_WITH_ADJACENCY).
++ *
++ * Approach: emit a topology dispatch at the start of each store_output
++ * lowering. The shader reads vs.xfb_topology sysval at runtime and branches
++ * into per-topology emission logic. For each affected topology, the lowered
++ * code emits guarded conditional stores — one per primitive this vertex
++ * contributes to, computing the output buffer position via primitive index
++ * and slot within the decomposed primitive.
++ *
++ * For LIST topologies (POINT/LINE/TRIANGLE LIST), takes a fast path that
++ * matches iter13's single-store behavior.
++ *
++ * For TRIANGLE_FAN, the central vertex (v=0) contributes to ALL primitives
++ * as slot 2 — handled via a NIR loop bounded by num_vertices.
++ *
++ * See ~/src/panvk-bifrost/iter17/phase{0,1,2}_*.md for full design context.
++ */
++
++#include "panvk_macros.h"
++
++#if PAN_ARCH < 9
++
++#include "panvk_shader.h"
++
++#include "compiler/nir/nir_builder.h"
++#include "pan_nir.h"
++
++#include <vulkan/vulkan_core.h>
++
++/* ----- Address arithmetic ----- */
++
++static nir_def *
++xfb_store_addr(nir_builder *b, nir_def *buf, nir_def *out_idx,
++               uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *byte_off = nir_iadd_imm(b,
++      nir_imul_imm(b, out_idx, stride), offset_bytes);
++   return nir_iadd(b, buf, nir_u2u64(b, byte_off));
++}
++
++static void
++emit_list_store(nir_builder *b, nir_def *buf, nir_def *output_count,
++                nir_def *instance_id, nir_def *raw_vid, nir_def *value,
++                uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *out_idx = nir_iadd(b,
++      nir_imul(b, instance_id, output_count), raw_vid);
++   nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
++   nir_store_global(b, value, addr);
++}
++
++static void
++emit_prim_store(nir_builder *b, nir_def *buf, nir_def *output_count,
++                nir_def *instance_id, nir_def *eligible,
++                nir_def *prim_idx, nir_def *slot,
++                uint32_t verts_per_prim,
++                nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_push_if(b, eligible);
++   {
++      nir_def *out_idx = nir_iadd(b,
++         nir_imul(b, instance_id, output_count),
++         nir_iadd(b, nir_imul_imm(b, prim_idx, verts_per_prim), slot));
++      nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
++      nir_store_global(b, value, addr);
++   }
++   nir_pop_if(b, NULL);
++}
++
++/* ----- Per-topology emission ----- */
++
++/* TRIANGLE_STRIP: vertex v contributes to prims v, v-1, v-2 (per eligibility). */
++static void
++emit_tri_strip(nir_builder *b, nir_def *v, nir_def *N,
++               nir_def *buf, nir_def *output_count, nir_def *instance_id,
++               nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *Nm2 = nir_iadd_imm(b, N, -2);
++   nir_def *Nm1 = nir_iadd_imm(b, N, -1);
++
++   /* Prim v, slot 0: v < N-2 */
++   emit_prim_store(b, buf, output_count, instance_id,
++      nir_ult(b, v, Nm2),
++      v, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
++
++   /* Prim v-1, slot = 1 if prim even else 2: 1 <= v < N-1 */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -1);
++      nir_def *parity = nir_iand_imm(b, prim, 1u);
++      nir_def *slot = nir_iadd_imm(b, parity, 1);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 1)),
++         nir_ult(b, v, Nm1));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, slot, 3, value, stride, offset_bytes);
++   }
++
++   /* Prim v-2, slot = 2 if prim even else 1: 2 <= v < N */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -2);
++      nir_def *parity = nir_iand_imm(b, prim, 1u);
++      nir_def *slot = nir_isub(b, nir_imm_int(b, 2), parity);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 2)),
++         nir_ult(b, v, N));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, slot, 3, value, stride, offset_bytes);
++   }
++}
++
++/* LINE_STRIP: vertex v contributes to prim v slot 0 + prim v-1 slot 1. */
++static void
++emit_line_strip(nir_builder *b, nir_def *v, nir_def *N,
++                nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *Nm1 = nir_iadd_imm(b, N, -1);
++
++   /* Prim v, slot 0: v < N-1 */
++   emit_prim_store(b, buf, output_count, instance_id,
++      nir_ult(b, v, Nm1),
++      v, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
++
++   /* Prim v-1, slot 1: 1 <= v < N */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -1);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 1)),
++         nir_ult(b, v, N));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
++   }
++}
++
++/* TRIANGLE_FAN: prim p emits {p+1, p+2, 0}.
++ *   vertex v=0: contributes to ALL prims as slot 2 (loop required)
++ *   vertex v>=1: contributes to prim v-1 as slot 0 (if 1 <= v <= N-2)
++ *   vertex v>=2: contributes to prim v-2 as slot 1 (if 2 <= v <= N-1)
++ */
++static void
++emit_tri_fan(nir_builder *b, nir_def *v, nir_def *N,
++             nir_def *buf, nir_def *output_count, nir_def *instance_id,
++             nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *Nm1 = nir_iadd_imm(b, N, -1);
++   nir_def *Nm2 = nir_iadd_imm(b, N, -2);
++
++   /* Prim v-1, slot 0: 1 <= v < N-1 */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -1);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 1)),
++         nir_ult(b, v, Nm1));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
++   }
++
++   /* Prim v-2, slot 1: 2 <= v < N */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -2);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 2)),
++         nir_ult(b, v, N));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 1), 3, value, stride, offset_bytes);
++   }
++
++   /* Central vertex (v == 0): loop over all prims, write to slot 2. */
++   nir_push_if(b, nir_ieq_imm(b, v, 0));
++   {
++      nir_variable *p_var = nir_local_variable_create(b->impl,
++         glsl_uint_type(), "fan_p");
++      nir_store_var(b, p_var, nir_imm_int(b, 0), 0x1);
++      nir_push_loop(b);
++      {
++         nir_def *p = nir_load_var(b, p_var);
++         nir_push_if(b, nir_uge(b, p, Nm2));
++         {
++            nir_jump(b, nir_jump_break);
++         }
++         nir_pop_if(b, NULL);
++
++         nir_def *out_idx = nir_iadd(b,
++            nir_imul(b, instance_id, output_count),
++            nir_iadd_imm(b, nir_imul_imm(b, p, 3), 2));
++         nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
++         nir_store_global(b, value, addr);
++
++         nir_store_var(b, p_var, nir_iadd_imm(b, p, 1), 0x1);
++      }
++      nir_pop_loop(b, NULL);
++   }
++   nir_pop_if(b, NULL);
++}
++
++/* LINE_LIST_WITH_ADJACENCY: 4-vertex groups [4i..4i+3]; output {4i+1, 4i+2}.
++ *   v contributes if v%4 == 1: prim v/4 slot 0
++ *   v contributes if v%4 == 2: prim v/4 slot 1
++ */
++static void
++emit_line_list_adj(nir_builder *b, nir_def *v, nir_def *N,
++                   nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                   nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   (void)N; /* eligibility is mod-based, not range-based */
++   nir_def *vmod4 = nir_iand_imm(b, v, 3u);
++   nir_def *prim = nir_ushr_imm(b, v, 2);  /* v / 4 */
++
++   emit_prim_store(b, buf, output_count, instance_id,
++      nir_ieq_imm(b, vmod4, 1),
++      prim, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
++
++   emit_prim_store(b, buf, output_count, instance_id,
++      nir_ieq_imm(b, vmod4, 2),
++      prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
++}
++
++/* LINE_STRIP_WITH_ADJACENCY: prim p emits {p+1, p+2}.
++ *   v contributes to prim v-1 slot 0 (1 <= v <= N-2)
++ *   v contributes to prim v-2 slot 1 (2 <= v <= N-1)
++ */
++static void
++emit_line_strip_adj(nir_builder *b, nir_def *v, nir_def *N,
++                    nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                    nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *Nm1 = nir_iadd_imm(b, N, -1);
++   nir_def *Nm2 = nir_iadd_imm(b, N, -2);
++
++   /* Prim v-1, slot 0: 1 <= v <= N-2 ⇔ v >= 1 AND v <= N-2 ⇔ v >= 1 AND v < N-1 */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -1);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 1)),
++         nir_ult(b, v, Nm1));
++      (void)Nm2;
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
++   }
++
++   /* Prim v-2, slot 1: 2 <= v <= N-1 ⇔ v >= 2 AND v < N */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -2);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 2)),
++         nir_ult(b, v, N));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
++   }
++}
++
++/* TRIANGLE_LIST_WITH_ADJACENCY: 6-vertex groups; output {6i, 6i+2, 6i+4}.
++ *   v contributes if v%6 == 0: prim v/6 slot 0
++ *   v contributes if v%6 == 2: prim v/6 slot 1
++ *   v contributes if v%6 == 4: prim v/6 slot 2
++ */
++static void
++emit_tri_list_adj(nir_builder *b, nir_def *v, nir_def *N,
++                  nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                  nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   (void)N;
++   nir_def *vmod6 = nir_umod_imm(b, v, 6);
++   nir_def *prim = nir_udiv_imm(b, v, 6);
++
++   for (uint32_t slot = 0; slot < 3; slot++) {
++      emit_prim_store(b, buf, output_count, instance_id,
++         nir_ieq_imm(b, vmod6, slot * 2),
++         prim, nir_imm_int(b, slot), 3, value, stride, offset_bytes);
++   }
++}
++
++/* TRIANGLE_STRIP_WITH_ADJACENCY: prim i emits:
++ *   even i: {2i, 2i+2, 2i+4}    (slots 0, 1, 2 ← input indices 2i, 2i+2, 2i+4)
++ *   odd  i: {2i, 2i+4, 2i+2}    (slots 0, 1, 2 ← input indices 2i, 2i+4, 2i+2)
++ *
++ * Only EVEN input vertices contribute (since all output indices are 2*something).
++ * For even input v:
++ *   prim v/2 slot 0 (always, if v/2 < N/2-2)
++ *   prim (v-2)/2 slot 1 if (v-2)/2 even, slot 2 if odd   (when v >= 2)
++ *   prim (v-4)/2 slot 2 if (v-4)/2 even, slot 1 if odd   (when v >= 4)
++ */
++static void
++emit_tri_strip_adj(nir_builder *b, nir_def *v, nir_def *N,
++                   nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                   nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   /* Bail for odd input vertices — they never contribute. */
++   nir_def *v_is_even = nir_ieq_imm(b, nir_iand_imm(b, v, 1u), 0);
++   nir_push_if(b, v_is_even);
++   {
++      nir_def *N_half = nir_ushr_imm(b, N, 1);
++      nir_def *max_prim = nir_iadd_imm(b, N_half, -2);  /* N/2 - 2 */
++      nir_def *v_half = nir_ushr_imm(b, v, 1);
++
++      /* Prim v/2 slot 0: v/2 < N/2 - 2 */
++      emit_prim_store(b, buf, output_count, instance_id,
++         nir_ult(b, v_half, max_prim),
++         v_half, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
++
++      /* Prim (v-2)/2 = v/2 - 1: v >= 2 AND prim < N/2-2 */
++      {
++         nir_def *prim = nir_iadd_imm(b, v_half, -1);
++         nir_def *parity = nir_iand_imm(b, prim, 1u);
++         nir_def *slot = nir_iadd_imm(b, parity, 1);  /* even→1, odd→2 */
++         nir_def *eligible = nir_iand(b,
++            nir_uge(b, v, nir_imm_int(b, 2)),
++            nir_ult(b, prim, max_prim));
++         emit_prim_store(b, buf, output_count, instance_id, eligible,
++                         prim, slot, 3, value, stride, offset_bytes);
++      }
++
++      /* Prim (v-4)/2 = v/2 - 2: v >= 4 AND prim < N/2-2 */
++      {
++         nir_def *prim = nir_iadd_imm(b, v_half, -2);
++         nir_def *parity = nir_iand_imm(b, prim, 1u);
++         nir_def *slot = nir_isub(b, nir_imm_int(b, 2), parity);  /* even→2, odd→1 */
++         nir_def *eligible = nir_iand(b,
++            nir_uge(b, v, nir_imm_int(b, 4)),
++            nir_ult(b, prim, max_prim));
++         emit_prim_store(b, buf, output_count, instance_id, eligible,
++                         prim, slot, 3, value, stride, offset_bytes);
++      }
++   }
++   nir_pop_if(b, NULL);
++}
++
++/* ----- Main lowering: per store_output XFB channel ----- */
++
++static void
++lower_xfb_output_iter17(nir_builder *b, nir_intrinsic_instr *intr,
++                        unsigned channel_idx, unsigned num_components,
++                        unsigned buffer, unsigned offset_words)
++{
++   assert(buffer < MAX_XFB_BUFFERS);
++   assert(nir_intrinsic_component(intr) == 0);
++
++   uint16_t stride = b->shader->info.xfb_stride[buffer] * 4;
++   assert(stride != 0);
++   uint16_t offset_bytes = offset_words * 4;
++
++   BITSET_SET(b->shader->info.system_values_read, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
++   BITSET_SET(b->shader->info.system_values_read, SYSTEM_VALUE_INSTANCE_ID);
++
++   nir_def *topology = load_sysval(b, graphics, 32, vs.xfb_topology);
++   nir_def *out_count = load_sysval(b, graphics, 32, vs.xfb_output_count);
++   nir_def *N = nir_load_num_vertices(b);
++   nir_def *v = nir_load_raw_vertex_id_pan(b);
++   nir_def *instance = nir_load_instance_id(b);
++   nir_def *buf = nir_load_xfb_address(b, 64, .base = buffer);
++
++   nir_def *src = intr->src[0].ssa;
++   nir_component_mask_t mask = nir_component_mask(num_components);
++   nir_def *value = nir_channels(b, src, mask << channel_idx);
++
++   /* Topology dispatch ladder. LIST first (fast path). */
++   nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LIST));
++   {
++      emit_list_store(b, buf, out_count, instance, v, value,
++                      stride, offset_bytes);
++   }
++   nir_push_else(b, NULL);
++   {
++      /* iter17 Janet Finding 3: gate all non-LIST emission on
++       * output_count > 0. For degenerate input counts (N < min required
++       * for the topology), output_count is 0 and we must emit NO stores
++       * — otherwise N-2 / N-3 / etc. arithmetic underflows in the
++       * eligibility predicates and we falsely fire stores. */
++      nir_push_if(b, nir_ult(b, nir_imm_int(b, 0), out_count));
++      {
++      nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_STRIP));
++      {
++         emit_tri_strip(b, v, N, buf, out_count, instance, value,
++                        stride, offset_bytes);
++      }
++      nir_push_else(b, NULL);
++      {
++         nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_STRIP));
++         {
++            emit_line_strip(b, v, N, buf, out_count, instance, value,
++                            stride, offset_bytes);
++         }
++         nir_push_else(b, NULL);
++         {
++            nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_FAN));
++            {
++               emit_tri_fan(b, v, N, buf, out_count, instance, value,
++                            stride, offset_bytes);
++            }
++            nir_push_else(b, NULL);
++            {
++               nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_LIST_ADJ));
++               {
++                  emit_line_list_adj(b, v, N, buf, out_count, instance, value,
++                                     stride, offset_bytes);
++               }
++               nir_push_else(b, NULL);
++               {
++                  nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_STRIP_ADJ));
++                  {
++                     emit_line_strip_adj(b, v, N, buf, out_count, instance, value,
++                                         stride, offset_bytes);
++                  }
++                  nir_push_else(b, NULL);
++                  {
++                     nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_LIST_ADJ));
++                     {
++                        emit_tri_list_adj(b, v, N, buf, out_count, instance, value,
++                                          stride, offset_bytes);
++                     }
++                     nir_push_else(b, NULL);
++                     {
++                        /* TRI_STRIP_ADJ — last case */
++                        emit_tri_strip_adj(b, v, N, buf, out_count, instance, value,
++                                           stride, offset_bytes);
++                     }
++                     nir_pop_if(b, NULL);
++                  }
++                  nir_pop_if(b, NULL);
++               }
++               nir_pop_if(b, NULL);
++            }
++            nir_pop_if(b, NULL);
++         }
++         nir_pop_if(b, NULL);
++      }
++      nir_pop_if(b, NULL);
++      }
++      nir_pop_if(b, NULL);  /* Janet Finding 3: close output_count > 0 guard */
++   }
++   nir_pop_if(b, NULL);
++}
++
++/* Mirror of pan_nir_lower_xfb's lower_xfb: load_vertex_id rewrite +
++ * dispatch store_output through our topology-aware emission. */
++static bool
++lower_xfb_iter17(nir_builder *b, nir_intrinsic_instr *intr,
++                 UNUSED void *data)
++{
++   if (intr->intrinsic == nir_intrinsic_load_vertex_id) {
++      b->cursor = nir_instr_remove(&intr->instr);
++      nir_def *repl = nir_iadd(b, nir_load_raw_vertex_id_pan(b),
++                               nir_load_raw_vertex_offset_pan(b));
++      nir_def_rewrite_uses(&intr->def, repl);
++      return true;
++   }
++
++   if (intr->intrinsic != nir_intrinsic_store_output)
++      return false;
++
++   bool progress = false;
++   b->cursor = nir_before_instr(&intr->instr);
++
++   /* io_xfb has only out[0,1]; the other 2 channels are in io_xfb2.
++    * Outer loop selects which annotation; inner picks which channel. */
++   for (unsigned i = 0; i < 2; ++i) {
++      nir_io_xfb xfb = i ? nir_intrinsic_io_xfb2(intr)
++                         : nir_intrinsic_io_xfb(intr);
++      for (unsigned j = 0; j < 2; ++j) {
++         if (!xfb.out[j].num_components)
++            continue;
++         lower_xfb_output_iter17(b, intr, i * 2 + j, xfb.out[j].num_components,
++                                 xfb.out[j].buffer, xfb.out[j].offset);
++         progress = true;
++      }
++   }
++
++   if (progress)
++      nir_instr_remove(&intr->instr);
++   return progress;
++}
++
++bool
++panvk_per_arch(nir_lower_xfb)(nir_shader *nir)
++{
++   return nir_shader_intrinsics_pass(
++      nir, lower_xfb_iter17, nir_metadata_control_flow, NULL);
++}
++
++#endif /* PAN_ARCH < 9 */

arch/mesa-panvk-bifrost-video/PKGBUILD

@@ -0,0 +1,181 @@
+# Maintainer: Markus Fritsche <fritsche.markus@gmail.com>
+#
+# mesa-panvk-bifrost-video — sibling of mesa-panvk-bifrost (r4) that adds
+# VK_KHR_video_decode_h264 on Mali Bifrost SBCs (PAN_ARCH 6/7) backed by
+# the SoC's V4L2-stateless hantro VPU (RK3566/RK3568).
+#
+# Campaign: ~/src/panvk-bifrost-video/ — Phase 4 byte-exact validated
+# 2026-05-21 (48/48 BBB display frames match ffmpeg+libva-v4l2-request-
+# fourier byte-for-byte on the same hantro). Phase 5 second-model review
+# completed; load-bearing findings (output_map OOB, static counter,
+# session_init unwind, probe_hantro gate) all applied.
+#
+# What it does (on top of r4):
+#   - 0001..0004: inherited from mesa-panvk-bifrost (robustness2/null-
+#     descriptor, vk1.1/1.2 advertisement, EXT_transform_feedback, XFB
+#     primitive decomposition) — symlinked from the r4 package directory
+#     so the patches don't drift between siblings.
+#   - 0005: VK_KHR_video_queue + VK_KHR_video_decode_queue +
+#     VK_KHR_video_decode_h264 backed by V4L2-stateless hantro.
+#     Touches 14 files in src/panfrost/vulkan/; full diff in
+#     0005-panvk-bifrost-video-KHR-video-decode-h264.patch.
+#
+# Co-existence:
+#   - Installs to /usr/lib/panvk-bifrost-video/ (parallel to r4's
+#     /usr/lib/panvk-bifrost/). Pick at runtime via VK_ICD_FILENAMES.
+#   - r4 stays the recommended default for the Chromium-GPU-process
+#     consumer (no video needed there). Use this package when the
+#     consumer wants Vulkan video decode (mpv-fourier, ffmpeg-vulkan,
+#     future Chromium-VulkanVideoDecoder).
+#
+# Phase 1 limitations to know about (documented in source comments):
+#   - Single video session per device (active_video singleton)
+#   - Synchronous decode at record time — no pipelining yet
+#   - Hardcoded /dev/video1 + /dev/media0 (matches RK3566/68, blocks
+#     other SoCs without a topology-walk port)
+#   - Bitstream source buffer assumed HOST_VISIBLE (true on panvk-
+#     bifrost, would need fallback on other backends)
+#
+# Build target: arch-aarch64 runner via marfrit-packages Gitea Actions.
+# Mesa build is slow (~30-60min on Cortex-A55).
+
+pkgname=mesa-panvk-bifrost-video
+_mesaver=26.0.6
+pkgver=26.0.6.r5.video1
+pkgrel=1
+pkgdesc="Patched Mesa libvulkan_panfrost.so adding VK_KHR_video_decode_h264 on Bifrost SBCs (sibling of mesa-panvk-bifrost-r4)"
+arch=('aarch64')
+url="https://git.reauktion.de/marfrit/panvk-bifrost"
+license=('MIT')
+
+depends=(
+    'mesa'              # for shared mesa runtime libs
+    'libdrm'
+    'wayland'
+    'libxcb'
+    'libx11'
+    'libxshmfence'
+    'zlib'
+    'zstd'
+    'libelf'
+    'libffi'
+    'expat'
+    'llvm-libs'
+    'lm_sensors'
+)
+makedepends=(
+    'meson'
+    'ninja'
+    'glslang'
+    'python-mako'
+    'python-packaging'
+    'wayland-protocols'
+    'libxrandr'
+    'xorgproto'
+    'libdrm'
+    'llvm'
+    'libclc'
+    'spirv-llvm-translator'
+    'spirv-tools'
+    'rust-bindgen'
+    'patch'
+)
+
+source=(
+    "https://archive.mesa3d.org/mesa-${_mesaver}.tar.xz"
+    "0001-panvk-expose-robustness2-nullDescriptor-bifrost.patch"
+    "0002-panvk-expose-vulkan-1.1-1.2-on-bifrost.patch"
+    "0003-panvk-bifrost-vk-ext-transform-feedback.patch"
+    "0004-panvk-bifrost-xfb-primitive-decomposition.patch"
+    "0005-panvk-bifrost-video-KHR-video-decode-h264.patch"
+    "icd.json"
+)
+# Mesa tarball checksum matches the sibling r4 package — same upstream version.
+sha256sums=(
+    'SKIP'  # mesa tarball — co-trust w/ r4 sibling
+    'SKIP'  # patches are local
+    'SKIP'
+    'SKIP'
+    'SKIP'
+    'SKIP'
+    'SKIP'  # icd.json
+)
+
+prepare() {
+    cd "mesa-${_mesaver}"
+
+    # r1+r2: small sed-based edits inherited from r4 (verbatim from the
+    # sibling PKGBUILD — keep in sync).
+    sed -i 's|\.KHR_robustness2 = PAN_ARCH >= 10,|.KHR_robustness2 = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
+    sed -i 's|\.EXT_robustness2 = PAN_ARCH >= 10,|.EXT_robustness2 = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
+    sed -i 's|\.nullDescriptor = PAN_ARCH >= 10,|.nullDescriptor = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
+    sed -i 's|bool has_vk1_1 = PAN_ARCH >= 10;|bool has_vk1_1 = true;|' src/panfrost/vulkan/panvk_vX_physical_device.c
+    sed -i 's|bool has_vk1_2 = PAN_ARCH >= 10;|bool has_vk1_2 = true;|' src/panfrost/vulkan/panvk_vX_physical_device.c
+
+    # r3: EXT_transform_feedback for Bifrost.
+    patch -p1 < "${srcdir}/0003-panvk-bifrost-vk-ext-transform-feedback.patch"
+
+    # r4: XFB primitive decomposition NIR pass.
+    patch -p1 < "${srcdir}/0004-panvk-bifrost-xfb-primitive-decomposition.patch"
+
+    # video: VK_KHR_video_decode_h264 via V4L2-hantro.
+    patch -p1 < "${srcdir}/0005-panvk-bifrost-video-KHR-video-decode-h264.patch"
+
+    # Sanity-check r1..r4 (inherited).
+    grep -q "KHR_robustness2 = true," src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "EXT_robustness2 = true," src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "nullDescriptor = true," src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "has_vk1_1 = true;" src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "has_vk1_2 = true;" src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "EXT_transform_feedback = PAN_ARCH < 9," src/panfrost/vulkan/panvk_vX_physical_device.c
+    test -f src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c
+    grep -q "panvk_per_arch(nir_lower_xfb)" src/panfrost/vulkan/panvk_vX_shader.c
+    test -f src/panfrost/vulkan/panvk_vX_xfb_lower.c
+
+    # Sanity-check video patch landed.
+    grep -q "KHR_video_queue = PAN_ARCH < 9 && panvk_v4l2_probe_hantro()" \
+        src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "PANVK_QUEUE_FAMILY_VIDEO_DECODE" src/panfrost/vulkan/panvk_device.h
+    test -f src/panfrost/vulkan/panvk_video_decode.c
+    test -f src/panfrost/vulkan/panvk_video_decode.h
+    test -f src/panfrost/vulkan/panvk_v4l2.c
+    test -f src/panfrost/vulkan/panvk_v4l2_h264.c
+    test -f src/panfrost/vulkan/panvk_v4l2_h264_slice_header.c
+    test -f src/panfrost/vulkan/panvk_v4l2_h264_slice_header.h
+    grep -q "panvk_v4l2_h264_slice_header.c" src/panfrost/vulkan/meson.build
+    grep -q "panvk_video_queue_submit_noop" src/panfrost/vulkan/panvk_vX_device.c
+}
+
+build() {
+    cd "mesa-${_mesaver}"
+    # Mirror r4's narrow build profile.
+    meson setup build/ \
+        --prefix=/usr \
+        --libdir=lib \
+        --buildtype=release \
+        -Dvulkan-drivers=panfrost \
+        -Dgallium-drivers= \
+        -Dplatforms=wayland,x11 \
+        -Dglx=disabled \
+        -Degl=disabled \
+        -Dgles1=disabled \
+        -Dgles2=disabled \
+        -Dvulkan-layers= \
+        -Dtools= \
+        -Dgallium-rusticl=false \
+        -Dmicrosoft-clc=disabled
+    meson compile -C build
+}
+
+package() {
+    cd "${srcdir}/mesa-${_mesaver}"
+
+    # Co-install path — parallel to r4's /usr/lib/panvk-bifrost/.
+    install -Dm755 build/src/panfrost/vulkan/libvulkan_panfrost.so \
+        "$pkgdir/usr/lib/panvk-bifrost-video/libvulkan_panfrost.so"
+
+    # ICD JSON pointing at the video build. Opt-in via VK_ICD_FILENAMES;
+    # NOT in /usr/share/vulkan/icd.d/ so it doesn't override stock or r4.
+    install -Dm644 "$srcdir/icd.json" \
+        "$pkgdir/usr/lib/panvk-bifrost-video/icd.json"
+}

arch/mesa-panvk-bifrost-video/README.md

@@ -0,0 +1,40 @@
+# mesa-panvk-bifrost-video
+
+Patched Mesa `libvulkan_panfrost.so` that **adds `VK_KHR_video_decode_h264`** on Mali Bifrost SBCs (PAN_ARCH 6/7, RK3566/RK3568 class hardware), backed by the SoC's V4L2-stateless **hantro** VPU.
+
+This is a **sibling** of [mesa-panvk-bifrost](../mesa-panvk-bifrost/) (the r4 package that exposes Bifrost to Chromium's Vulkan compositor). Pick this one when the consumer wants Vulkan **video decode** in addition; pick r4 for compositor-only.
+
+## Status
+
+Phase 4 byte-exact validated 2026-05-21: 48/48 unique BBB display frames decoded by this package are byte-identical to `ffmpeg+libva-v4l2-request-fourier` running on the same hantro hardware. Phase 5 second-model review completed; all load-bearing findings addressed. First publish via marfrit-packages CI 2026-05-22 (PR #79 merge did not auto-fire Actions; this re-trigger restores the standard build/sign/publish path).
+
+## How to use
+
+```sh
+# Co-installs alongside r4 and stock mesa.
+sudo pacman -S mesa-panvk-bifrost-video
+
+# Opt in (not on the default loader search path).
+export VK_ICD_FILENAMES=/usr/lib/panvk-bifrost-video/icd.json
+export PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1   # mesa-upstream gate
+
+# Run a Vulkan video consumer.
+vulkan-video-dec-simple-test -i your.h264 --codec h264 --noPresent --maxFrameCount 50
+# or
+ffmpeg -hwaccel vulkan -i your.mp4 ...
+```
+
+## Phase 1 limitations
+
+Documented in source comments and worth knowing before relying on this in production:
+
+- **Single video session per device.** Concurrent `VkVideoSessionKHR` on the same device clobber each other (`active_video` singleton). Sufficient for current single-stream consumers.
+- **Synchronous decode at record time.** The full V4L2 ioctl dance runs to completion inside `vkCmdDecodeVideoKHR`. No pipelining. Throughput is bounded by hantro's ~1.16× realtime on 1080p H.264.
+- **Hardcoded `/dev/video1` + `/dev/media0`.** Matches RK3566/68 but won't work on other SoCs without a topology-walk port (see `libva-v4l2-request-fourier` for the full version).
+- **Bitstream source buffer assumed HOST_VISIBLE.** True on panvk-bifrost (no DEVICE_LOCAL-only memory types exist), but the code silently skips decode if the app bound the buffer to non-host-visible memory.
+
+## Co-existence
+
+- Installs to `/usr/lib/panvk-bifrost-video/` — parallel to r4's `/usr/lib/panvk-bifrost/` and stock `/usr/lib/`.
+- Opt-in via `VK_ICD_FILENAMES`; does NOT register itself in `/usr/share/vulkan/icd.d/`.
+- Three drivers coexist without conflict; the user picks at runtime which to use.

arch/mesa-panvk-bifrost-video/icd.json

@@ -0,0 +1,7 @@
+{
+    "ICD": {
+        "api_version": "1.4.335",
+        "library_path": "/usr/lib/panvk-bifrost-video/libvulkan_panfrost.so"
+    },
+    "file_format_version": "1.0.1"
+}

arch/mesa-panvk-bifrost/0001-panvk-expose-robustness2-nullDescriptor-bifrost.patch

@@ -0,0 +1,57 @@
+From: claude-noether (on behalf of mfritsche)
+Date: 2026-05-19
+Subject: panvk: expose VK_KHR/EXT_robustness2 + nullDescriptor on Bifrost (PAN_ARCH 6/7)
+
+Without this, Mesa's Zink driver refuses to use PanVk-Bifrost as its Vulkan
+backend, falling back silently to llvmpipe (software rasterizer) for all
+GL-via-Zink on Bifrost SBCs. That defeats the entire purpose of having a
+Vulkan driver on Bifrost — GL acceleration via Zink is the most natural
+near-term consumer.
+
+panvk_vX_nir_lower_descriptors.c:1309 and panvk_vX_shader.c:1355 already
+plumb dev->vk.enabled_features.nullDescriptor arch-agnostically — the gate
+at panvk_vX_physical_device.c was set conservatively when Bifrost was
+unmaintained, not because of hardware incapability.
+
+iter1–7 of the panvk-bifrost campaign proved fundamental driver functions
+on Mali-G52 r1 MC1 (PAN_ARCH=7). This patch is the iter8 follow-up.
+
+robustBufferAccess2 and robustImageAccess2 are NOT flipped — they're
+independent rb2 features Zink doesn't require, gated differently
+(robustBufferAccess2 = PAN_ARCH >= 11, robustImageAccess2 = false), and
+out of scope for iter8.
+
+---
+ src/panfrost/vulkan/panvk_vX_physical_device.c | 6 +++---
+ 1 file changed, 3 insertions(+), 3 deletions(-)
+
+diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
+--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
+@@ -91,7 +91,7 @@ get_device_extensions(const struct panvk_physical_device *device,
+       .KHR_pipeline_binary = true,
+       .KHR_pipeline_executable_properties = true,
+       .KHR_pipeline_library = true,
+-      .KHR_robustness2 = PAN_ARCH >= 10,
++      .KHR_robustness2 = true,
+       .KHR_sampler_mirror_clamp_to_edge = true,
+       .KHR_sampler_ycbcr_conversion = true,
+       .KHR_separate_depth_stencil_layouts = true,
+@@ -168,7 +168,7 @@ get_device_extensions(const struct panvk_physical_device *device,
+       .EXT_queue_family_foreign = true,
+       .EXT_robustness = pan_arch(device->kmod.dev->props.gpu_id) >= 9,
+       .EXT_image_robustness = true,
+-      .EXT_robustness2 = PAN_ARCH >= 10,
++      .EXT_robustness2 = true,
+       .EXT_sampler_filter_minmax = PAN_ARCH >= 10,
+       .EXT_scalar_block_layout = true,
+       .EXT_separate_stencil_usage = true,
+@@ -493,7 +493,7 @@ get_device_features(const struct panvk_physical_device *device,
+       /* VK_KHR_robustness2 */
+       .robustBufferAccess2 = PAN_ARCH >= 11,
+       .robustImageAccess2 = false,
+-      .nullDescriptor = PAN_ARCH >= 10,
++      .nullDescriptor = true,
+
+       /* VK_KHR_shader_clock */
+       .shaderSubgroupClock = device->kmod.dev->props.gpu_can_query_timestamp,

arch/mesa-panvk-bifrost/0002-panvk-expose-vulkan-1.1-1.2-on-bifrost.patch

@@ -0,0 +1,47 @@
+From: claude-noether (on behalf of mfritsche)
+Date: 2026-05-20
+Subject: panvk: expose Vulkan 1.1 + 1.2 on Bifrost (PAN_ARCH 6/7)
+
+ANGLE (Chromium's GL stack) requires apiVersion >= 1.1 to initialize. Without
+this, Brave / Chromium's GPU process fails at GL info collection:
+
+  vk_renderer.cpp:2659 (initialize): ANGLE Requires a minimum Vulkan device
+                                     version of 1.1
+  Display::initialize error 0: Internal Vulkan error (-9): The requested
+                               version of Vulkan is not supported by the driver
+
+Stack-up with iter8's robustness2 patch enables ANGLE → PanVk-Bifrost →
+Skia (via --enable-features=Vulkan) on Bifrost SBCs.
+
+PanVk-Bifrost already supports the bulk of 1.1-promoted features as extensions
+(multiview, maintenance1-3, descriptor update template, 16-bit storage,
+descriptor update template, sampler ycbcr, variable pointers, etc. — all
+visible in iter0 vulkaninfo). The version bump primarily bundles them.
+
+Risk: Vulkan 1.1 has features beyond what iter1–7 exercised (protected memory,
+full subgroup ops). Specific app failures will be characterizable.
+
+1.2 is also flipped — Brave's Vulkan path may want descriptor indexing,
+buffer device address, etc. (all listed in iter0 vulkaninfo as supported
+extensions, just gated as 1.0-with-extensions, not 1.2-core).
+
+---
+ src/panfrost/vulkan/panvk_vX_physical_device.c | 4 ++--
+ 1 file changed, 2 insertions(+), 2 deletions(-)
+
+diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
+--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
+@@ -38,8 +38,8 @@ get_device_extensions(const struct panvk_physical_device *device,
+                       struct vk_device_extension_table *ext)
+ {
+    *ext = (struct vk_device_extension_table){
+-      .KHR_8bit_storage = true,
+-      .KHR_16bit_storage = true,
+-      bool has_vk1_1 = PAN_ARCH >= 10;
+-      bool has_vk1_2 = PAN_ARCH >= 10;
++      .KHR_8bit_storage = true,
++      .KHR_16bit_storage = true,
++      bool has_vk1_1 = true;
++      bool has_vk1_2 = true;
+       *ext = (struct vk_device_extension_table){

arch/mesa-panvk-bifrost/0003-panvk-bifrost-vk-ext-transform-feedback.patch

@@ -0,0 +1,328 @@
+--- a/src/panfrost/vulkan/panvk_shader.h	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/panvk_shader.h	2026-05-20 18:52:53.312698258 +0200
+@@ -150,6 +150,10 @@
+    struct {
+ #if PAN_ARCH < 9
+       int32_t raw_vertex_offset;
++      uint32_t num_vertices;       /* iter13: XFB needs per-draw vertex count */
++      /* aligned_u64 attribute below inserts the 4-byte alignment gap
++       * after num_vertices automatically — no explicit pad needed. */
++      aligned_u64 xfb_address[4];  /* iter13: 4 transform feedback buffer base addresses */
+ #endif
+       int32_t first_vertex;
+       int32_t base_instance;
+--- a/src/panfrost/vulkan/panvk_vX_physical_device.c	2026-05-20 19:09:29.711145446 +0200
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c	2026-05-20 18:52:54.832720445 +0200
+@@ -169,6 +169,7 @@
+       .EXT_provoking_vertex = true,
+       .EXT_queue_family_foreign = true,
+       .EXT_robustness2 = true,
++      .EXT_transform_feedback = PAN_ARCH < 9,   /* iter13: JM-class only for now */
+       .EXT_sampler_filter_minmax = PAN_ARCH >= 10,
+       .EXT_scalar_block_layout = true,
+       .EXT_separate_stencil_usage = true,
+@@ -495,6 +496,10 @@
+       .robustImageAccess2 = false,
+       .nullDescriptor = true,
+ 
++      /* VK_EXT_transform_feedback (iter13) */
++      .transformFeedback = PAN_ARCH < 9,
++      .geometryStreams = false,
++
+       /* VK_KHR_shader_clock */
+       .shaderSubgroupClock = device->kmod.dev->props.gpu_can_query_timestamp,
+       .shaderDeviceClock = device->kmod.dev->props.timestamp_device_coherent,
+@@ -1020,6 +1025,18 @@
+       .robustStorageBufferAccessSizeAlignment = 1,
+       .robustUniformBufferAccessSizeAlignment = 1,
+ 
++      /* VK_EXT_transform_feedback (iter13) */
++      .maxTransformFeedbackStreams = 1,
++      .maxTransformFeedbackBuffers = 4,
++      .maxTransformFeedbackBufferSize = UINT32_MAX,
++      .maxTransformFeedbackStreamDataSize = 512,
++      .maxTransformFeedbackBufferDataSize = 512,
++      .maxTransformFeedbackBufferDataStride = 2048,
++      .transformFeedbackQueries = false,
++      .transformFeedbackStreamsLinesTriangles = false,
++      .transformFeedbackRasterizationStreamSelect = false,
++      .transformFeedbackDraw = false,
++
+       /* VK_EXT_shader_object */
+       /* We do not currently support VK_EXT_shader_object but this is used
+        * internally by vk_shader
+--- a/src/panfrost/vulkan/panvk_vX_shader.c	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/panvk_vX_shader.c	2026-05-20 18:52:56.556745611 +0200
+@@ -21,6 +21,7 @@
+ #include "panvk_physical_device.h"
+ #include "panvk_sampler.h"
+ #include "panvk_shader.h"
++#include "pan_nir.h"   /* iter13: pan_nir_lower_xfb */
+ 
+ #include "spirv/nir_spirv.h"
+ #include "util/memstream.h"
+@@ -100,6 +101,20 @@
+    case nir_intrinsic_load_raw_vertex_offset_pan:
+       val = load_sysval(b, graphics, bit_size, vs.raw_vertex_offset);
+       break;
++   case nir_intrinsic_load_num_vertices:    /* iter13: XFB index calc */
++      val = load_sysval(b, graphics, bit_size, vs.num_vertices);
++      break;
++   case nir_intrinsic_load_xfb_address: {   /* iter13: XFB buffer N base address */
++      unsigned idx = nir_intrinsic_base(intr);
++      switch (idx) {
++      case 0: val = load_sysval(b, graphics, bit_size, vs.xfb_address[0]); break;
++      case 1: val = load_sysval(b, graphics, bit_size, vs.xfb_address[1]); break;
++      case 2: val = load_sysval(b, graphics, bit_size, vs.xfb_address[2]); break;
++      case 3: val = load_sysval(b, graphics, bit_size, vs.xfb_address[3]); break;
++      default: return false;
++      }
++      break;
++   }
+    case nir_intrinsic_load_layer_id:
+       assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+       val = load_sysval(b, graphics, bit_size, layer_id);
+@@ -457,6 +472,7 @@
+             core_max_id);
+ 
+    pan_preprocess_nir(nir, pdev->kmod.dev->props.gpu_id);
++
+ }
+ 
+ static void
+@@ -870,6 +886,18 @@
+             nir_var_shader_in | nir_var_shader_out, UINT32_MAX);
+    NIR_PASS(_, nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
+             glsl_type_size, nir_lower_io_use_interpolated_input_intrinsics);
++
++#if PAN_ARCH < 9
++   /* iter13: VK_EXT_transform_feedback — runs AFTER nir_lower_io so that
++    * shader outputs are now store_output intrinsics that pan_nir_lower_xfb
++    * can rewrite to nir_store_global+nir_load_xfb_address. */
++   if (nir->info.stage == MESA_SHADER_VERTEX &&
++       nir->info.has_transform_feedback_varyings) {
++      NIR_PASS(_, nir, nir_opt_constant_folding);
++      NIR_PASS(_, nir, nir_io_add_intrinsic_xfb_info);
++      NIR_PASS(_, nir, pan_nir_lower_xfb);
++   }
++#endif
+ }
+ 
+ static VkResult
+@@ -1288,6 +1316,9 @@
+       .view_mask = (state && state->rp) ? state->rp->view_mask : 0,
+       .robust2_modes = robust2_modes,
+       .robust_descriptors = dev->vk.enabled_features.nullDescriptor,
++      /* iter13: XFB shaders must disable IDVS (matches Panfrost-Gallium). */
++      .no_idvs = (info->stage == MESA_SHADER_VERTEX) &&
++                 info->nir->info.has_transform_feedback_varyings,
+    };
+ 
+    switch (info->stage) {
+--- a/src/panfrost/vulkan/panvk_cmd_draw.h	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/panvk_cmd_draw.h	2026-05-20 18:52:57.748763011 +0200
+@@ -135,6 +135,19 @@
+    struct panvk_graphics_sysvals sysvals;
+ 
+ #if PAN_ARCH < 9
++   /* iter13: VK_EXT_transform_feedback state (JM-class only for now). */
++   struct {
++      bool active;
++      uint32_t buffer_count;
++      struct {
++         uint64_t addr;
++         uint64_t offset;
++         uint64_t size;
++      } buffers[4];
++   } xfb;
++#endif
++
++#if PAN_ARCH < 9
+    struct panvk_shader_link link;
+ #endif
+ 
+--- a/src/panfrost/vulkan/panvk_vX_cmd_draw.c	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/panvk_vX_cmd_draw.c	2026-05-20 19:10:23.031919662 +0200
+@@ -10,6 +10,7 @@
+ #include "panvk_entrypoints.h"
+ 
+ #include "pan_desc.h"
++#include "pan_compiler.h"   /* PAN_SHADER_OOB_ADDRESS */
+ #include "pan_util.h"
+ 
+ static void
+@@ -722,6 +723,35 @@
+    set_gfx_sysval(cmdbuf, dirty_sysvals, vs.raw_vertex_offset,
+                   info->vertex.raw_offset);
+    set_gfx_sysval(cmdbuf, dirty_sysvals, layer_id, info->layer_id);
++
++   /* iter13: VK_EXT_transform_feedback sysvals — always set (per draw),
++    * reflect bound XFB state. set_gfx_sysval is a no-op if value unchanged. */
++   set_gfx_sysval(cmdbuf, dirty_sysvals, vs.num_vertices, info->vertex.count);
++   {
++      const struct panvk_cmd_graphics_state *_gfx = &cmdbuf->state.gfx;
++      /* iter13: default each XFB buffer address to PAN_SHADER_OOB_ADDRESS
++       * (= 1<<63). This is the Panfrost-Gallium memory-sink idiom — the
++       * Bifrost MMU silently discards stores to this address, so a pipeline
++       * with XFB outputs used in a non-XFB draw (or in an XFB draw with
++       * fewer bound buffers than the shader declares) is safe instead of
++       * faulting. See gallium/drivers/panfrost/pan_cmdstream.c PAN_SYSVAL_XFB. */
++      uint64_t _xa0 = PAN_SHADER_OOB_ADDRESS, _xa1 = PAN_SHADER_OOB_ADDRESS,
++               _xa2 = PAN_SHADER_OOB_ADDRESS, _xa3 = PAN_SHADER_OOB_ADDRESS;
++      if (_gfx->xfb.active) {
++         if (_gfx->xfb.buffer_count > 0 && _gfx->xfb.buffers[0].addr)
++            _xa0 = _gfx->xfb.buffers[0].addr + _gfx->xfb.buffers[0].offset;
++         if (_gfx->xfb.buffer_count > 1 && _gfx->xfb.buffers[1].addr)
++            _xa1 = _gfx->xfb.buffers[1].addr + _gfx->xfb.buffers[1].offset;
++         if (_gfx->xfb.buffer_count > 2 && _gfx->xfb.buffers[2].addr)
++            _xa2 = _gfx->xfb.buffers[2].addr + _gfx->xfb.buffers[2].offset;
++         if (_gfx->xfb.buffer_count > 3 && _gfx->xfb.buffers[3].addr)
++            _xa3 = _gfx->xfb.buffers[3].addr + _gfx->xfb.buffers[3].offset;
++      }
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[0], _xa0);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[1], _xa1);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[2], _xa2);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_address[3], _xa3);
++   }
+ #endif
+ 
+    if (dyn_gfx_state_dirty(cmdbuf, CB_BLEND_CONSTANTS)) {
+--- a/src/panfrost/vulkan/meson.build	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/meson.build	2026-05-20 18:53:04.484861338 +0200
+@@ -73,6 +73,7 @@
+ jm_inc_dir = ['jm']
+ jm_files = [
+   'jm/panvk_vX_bind_queue.c',
++  'jm/panvk_vX_cmd_xfb.c',   # iter13
+   'jm/panvk_vX_cmd_buffer.c',
+   'jm/panvk_vX_cmd_dispatch.c',
+   'jm/panvk_vX_cmd_draw.c',
+--- a/src/panfrost/vulkan/jm/panvk_vX_cmd_buffer.c	2026-04-29 22:19:00.000000000 +0200
++++ b/src/panfrost/vulkan/jm/panvk_vX_cmd_buffer.c	2026-05-20 19:10:26.163965149 +0200
+@@ -473,5 +473,12 @@
+ 
+    vk_command_buffer_begin(&cmdbuf->vk, pBeginInfo);
+ 
++#if PAN_ARCH < 9
++   /* iter13: clear XFB state on Begin so a reused command buffer does not
++    * inherit stale xfb.buffer_count / xfb.active / xfb.buffers[] from a
++    * prior recording. */
++   memset(&cmdbuf->state.gfx.xfb, 0, sizeof(cmdbuf->state.gfx.xfb));
++#endif
++
+    return VK_SUCCESS;
+ }
+--- a/src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c	2026-05-18 12:50:53.067999996 +0200
++++ b/src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c	2026-05-20 19:10:27.175979847 +0200
+@@ -0,0 +1,111 @@
++/*
++ * Copyright © 2026 mfritsche / claude-noether
++ * SPDX-License-Identifier: MIT
++ *
++ * iter13: VK_EXT_transform_feedback command handlers for the JM
++ * architecture path (Bifrost v6/v7 + Valhall-JM v9).
++ *
++ * The runtime contract:
++ *   - vkCmdBindTransformFeedbackBuffersEXT: stash (gpu_addr, offset, size)
++ *     for each slot into cmdbuf->state.gfx.xfb.buffers[].
++ *   - vkCmdBeginTransformFeedbackEXT: set cmdbuf->state.gfx.xfb.active = true.
++ *     Mark sysvals dirty so the next draw re-emits vs.xfb_address[].
++ *   - vkCmdEndTransformFeedbackEXT: set active = false.
++ *
++ * Counter buffers (firstCounterBuffer/counterBufferCount/pCounterBuffers/
++ * pCounterBufferOffsets) are accepted by API but ignored — v1 doesn't
++ * support pause/resume. transformFeedbackDraw is advertised as false.
++ *
++ * Per-draw integration: jm/panvk_vX_cmd_draw.c reads cmdbuf->state.gfx.xfb
++ * and populates vs.xfb_address[i] for shader use. The pan_nir_lower_xfb
++ * pass in panvk_vX_shader.c emits nir_load_xfb_address(i) which lowers
++ * (via panvk_vX_shader.c sysval handler) to a load from the per-draw
++ * sysval push area.
++ */
++
++#include "vk_log.h"
++#include "util/log.h"
++
++#include "panvk_cmd_buffer.h"
++#include "panvk_cmd_draw.h"
++#include "panvk_buffer.h"
++#include "panvk_entrypoints.h"
++
++VKAPI_ATTR void VKAPI_CALL
++panvk_per_arch(CmdBindTransformFeedbackBuffersEXT)(
++   VkCommandBuffer commandBuffer,
++   uint32_t firstBinding,
++   uint32_t bindingCount,
++   const VkBuffer *pBuffers,
++   const VkDeviceSize *pOffsets,
++   const VkDeviceSize *pSizes)
++{
++   VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
++   struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
++
++   for (uint32_t i = 0; i < bindingCount; i++) {
++      uint32_t slot = firstBinding + i;
++      if (slot >= 4)
++         continue;
++
++      VK_FROM_HANDLE(panvk_buffer, buf, pBuffers[i]);
++      gfx->xfb.buffers[slot].addr = panvk_buffer_gpu_ptr(buf, 0);
++      gfx->xfb.buffers[slot].offset = pOffsets[i];
++      gfx->xfb.buffers[slot].size =
++         (pSizes != NULL && pSizes[i] != VK_WHOLE_SIZE)
++            ? pSizes[i]
++            : (buf->vk.size - pOffsets[i]);
++   }
++
++   if (firstBinding + bindingCount > gfx->xfb.buffer_count)
++      gfx->xfb.buffer_count = firstBinding + bindingCount;
++}
++
++VKAPI_ATTR void VKAPI_CALL
++panvk_per_arch(CmdBeginTransformFeedbackEXT)(
++   VkCommandBuffer commandBuffer,
++   uint32_t firstCounterBuffer,
++   uint32_t counterBufferCount,
++   const VkBuffer *pCounterBuffers,
++   const VkDeviceSize *pCounterBufferOffsets)
++{
++   VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
++   struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
++
++   /* Counter buffers ignored in v1 — see VkPhysicalDeviceTransformFeedback
++    * PropertiesEXT.transformFeedbackDraw = false in panvk_vX_physical_device.c.
++    * App is spec-compliant if it does not pass counter buffers (which our
++    * features advertisement allows), but warn loudly if it does so we do not
++    * silently produce wrong capture state. */
++   (void)firstCounterBuffer;
++   (void)pCounterBufferOffsets;
++   if (counterBufferCount > 0 && pCounterBuffers != NULL) {
++      mesa_logw("panvk: CmdBeginTransformFeedbackEXT: counter buffers not "
++                "implemented (transformFeedbackDraw=false); XFB resume will "
++                "restart at buffer offset 0");
++   }
++
++   gfx->xfb.active = true;
++   /* Per-draw set_gfx_sysval picks up the change automatically — no
++    * explicit dirty marking required (set_gfx_sysval uses memcmp +
++    * BITSET to detect state diffs and re-emit sysvals). */
++}
++
++VKAPI_ATTR void VKAPI_CALL
++panvk_per_arch(CmdEndTransformFeedbackEXT)(
++   VkCommandBuffer commandBuffer,
++   uint32_t firstCounterBuffer,
++   uint32_t counterBufferCount,
++   const VkBuffer *pCounterBuffers,
++   const VkDeviceSize *pCounterBufferOffsets)
++{
++   VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
++   struct panvk_cmd_graphics_state *gfx = &cmdbuf->state.gfx;
++
++   (void)firstCounterBuffer;
++   (void)counterBufferCount;
++   (void)pCounterBuffers;
++   (void)pCounterBufferOffsets;
++
++   gfx->xfb.active = false;
++}

arch/mesa-panvk-bifrost/0004-panvk-bifrost-xfb-primitive-decomposition.patch

@@ -0,0 +1,629 @@
+diff -urN a/src/panfrost/vulkan/meson.build b/src/panfrost/vulkan/meson.build
+--- a/src/panfrost/vulkan/meson.build	2026-05-21 14:04:02.529474145 +0200
++++ b/src/panfrost/vulkan/meson.build	2026-05-21 14:04:04.106755486 +0200
+@@ -123,6 +123,7 @@
+   'panvk_vX_nir_lower_input_attachment_loads.c',
+   'panvk_vX_sampler.c',
+   'panvk_vX_shader.c',
++  'panvk_vX_xfb_lower.c',
+   sha1_h,
+ ]
+ 
+diff -urN a/src/panfrost/vulkan/panvk_shader.h b/src/panfrost/vulkan/panvk_shader.h
+--- a/src/panfrost/vulkan/panvk_shader.h	2026-05-21 14:04:02.525251986 +0200
++++ b/src/panfrost/vulkan/panvk_shader.h	2026-05-21 14:04:04.084251800 +0200
+@@ -154,6 +154,8 @@
+       /* aligned_u64 attribute below inserts the 4-byte alignment gap
+        * after num_vertices automatically — no explicit pad needed. */
+       aligned_u64 xfb_address[4];  /* iter13: 4 transform feedback buffer base addresses */
++      uint32_t xfb_topology;       /* iter17: panvk_xfb_topology enum value */
++      uint32_t xfb_output_count;   /* iter17: per-instance output verts after decomp */
+ #endif
+       int32_t first_vertex;
+       int32_t base_instance;
+@@ -569,4 +571,76 @@
+    struct pan_compute_dim local_size, const void *bin_ptr, size_t bin_size,
+    struct panvk_shader **shader_out);
+ 
++
++#if PAN_ARCH < 9
++/* iter17: encoding for vs.xfb_topology sysval. Maps VkPrimitiveTopology values
++ * we need to distinguish at shader runtime for XFB capture. LIST topologies
++ * use the iter13 single-store fast path; non-LIST need per-vertex decomposition. */
++enum panvk_xfb_topology {
++   PANVK_XFB_TOPO_LIST            = 0,
++   PANVK_XFB_TOPO_LINE_STRIP      = 1,
++   PANVK_XFB_TOPO_TRI_STRIP       = 2,
++   PANVK_XFB_TOPO_TRI_FAN         = 3,
++   PANVK_XFB_TOPO_LINE_LIST_ADJ   = 4,
++   PANVK_XFB_TOPO_LINE_STRIP_ADJ  = 5,
++   PANVK_XFB_TOPO_TRI_LIST_ADJ    = 6,
++   PANVK_XFB_TOPO_TRI_STRIP_ADJ   = 7,
++};
++
++#include "panvk_macros.h"
++struct nir_shader;
++bool panvk_per_arch(nir_lower_xfb)(struct nir_shader *nir);
++
++/* Map VkPrimitiveTopology to panvk_xfb_topology enum (driver-side helper). */
++static inline uint32_t
++panvk_vk_topology_to_xfb_enum(VkPrimitiveTopology topo)
++{
++   switch (topo) {
++   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
++      return PANVK_XFB_TOPO_LINE_STRIP;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
++      return PANVK_XFB_TOPO_TRI_STRIP;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
++      return PANVK_XFB_TOPO_TRI_FAN;
++   case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
++      return PANVK_XFB_TOPO_LINE_LIST_ADJ;
++   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
++      return PANVK_XFB_TOPO_LINE_STRIP_ADJ;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
++      return PANVK_XFB_TOPO_TRI_LIST_ADJ;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
++      return PANVK_XFB_TOPO_TRI_STRIP_ADJ;
++   case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
++   case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
++   default:
++      return PANVK_XFB_TOPO_LIST;
++   }
++}
++
++/* Compute the per-instance output vertex count for a given (topology, input count). */
++static inline uint32_t
++panvk_xfb_output_count(VkPrimitiveTopology topo, uint32_t input_count)
++{
++   switch (topo) {
++   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
++      return input_count >= 1 ? 2u * (input_count - 1u) : 0u;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
++      return input_count >= 2 ? 3u * (input_count - 2u) : 0u;
++   case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
++      return (input_count / 4u) * 2u;
++   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
++      return input_count >= 3 ? 2u * (input_count - 3u) : 0u;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
++      return (input_count / 6u) * 3u;
++   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
++      return input_count >= 6 ? 3u * (input_count / 2u - 2u) : 0u;
++   default:
++      return input_count;  /* LIST topologies: 1:1 mapping */
++   }
++}
++#endif
++
++
+ #endif
+diff -urN a/src/panfrost/vulkan/panvk_vX_cmd_draw.c b/src/panfrost/vulkan/panvk_vX_cmd_draw.c
+--- a/src/panfrost/vulkan/panvk_vX_cmd_draw.c	2026-05-21 14:04:02.528576354 +0200
++++ b/src/panfrost/vulkan/panvk_vX_cmd_draw.c	2026-05-21 14:04:04.091357598 +0200
+@@ -727,6 +727,20 @@
+    /* iter13: VK_EXT_transform_feedback sysvals — always set (per draw),
+     * reflect bound XFB state. set_gfx_sysval is a no-op if value unchanged. */
+    set_gfx_sysval(cmdbuf, dirty_sysvals, vs.num_vertices, info->vertex.count);
++
++   /* iter17: XFB primitive-decomposition sysvals.
++    * xfb_topology = enum value for the current bound topology.
++    * xfb_output_count = per-instance output vertex count after decomposition.
++    * For LIST topologies, output_count == input vertex count and the shader
++    * takes the iter13 single-store fast path. */
++   {
++      VkPrimitiveTopology vk_topo =
++         cmdbuf->vk.dynamic_graphics_state.ia.primitive_topology;
++      uint32_t topo_enum = panvk_vk_topology_to_xfb_enum(vk_topo);
++      uint32_t out_count = panvk_xfb_output_count(vk_topo, info->vertex.count);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_topology, topo_enum);
++      set_gfx_sysval(cmdbuf, dirty_sysvals, vs.xfb_output_count, out_count);
++   }
+    {
+       const struct panvk_cmd_graphics_state *_gfx = &cmdbuf->state.gfx;
+       /* iter13: default each XFB buffer address to PAN_SHADER_OOB_ADDRESS
+diff -urN a/src/panfrost/vulkan/panvk_vX_shader.c b/src/panfrost/vulkan/panvk_vX_shader.c
+--- a/src/panfrost/vulkan/panvk_vX_shader.c	2026-05-21 14:04:02.527576494 +0200
++++ b/src/panfrost/vulkan/panvk_vX_shader.c	2026-05-21 14:04:04.098356619 +0200
+@@ -895,7 +895,10 @@
+        nir->info.has_transform_feedback_varyings) {
+       NIR_PASS(_, nir, nir_opt_constant_folding);
+       NIR_PASS(_, nir, nir_io_add_intrinsic_xfb_info);
+-      NIR_PASS(_, nir, pan_nir_lower_xfb);
++      /* iter17: panvk-specific replacement for pan_nir_lower_xfb that handles
++       * primitive decomposition for non-LIST topologies. Single-store LIST
++       * fast path matches iter13 behavior. */
++      NIR_PASS(_, nir, panvk_per_arch(nir_lower_xfb));
+    }
+ #endif
+ }
+diff -urN a/src/panfrost/vulkan/panvk_vX_xfb_lower.c b/src/panfrost/vulkan/panvk_vX_xfb_lower.c
+--- a/src/panfrost/vulkan/panvk_vX_xfb_lower.c	1970-01-01 01:00:00.000000000 +0100
++++ b/src/panfrost/vulkan/panvk_vX_xfb_lower.c	2026-05-21 14:04:04.115354242 +0200
+@@ -0,0 +1,486 @@
++/*
++ * Copyright © 2026 mfritsche / claude-noether
++ * SPDX-License-Identifier: MIT
++ *
++ * iter17: panvk-specific replacement for pan_nir_lower_xfb that handles
++ * primitive decomposition for transform_feedback on non-LIST topologies
++ * (TRIANGLE_STRIP/FAN, LINE_STRIP, *_WITH_ADJACENCY).
++ *
++ * Approach: emit a topology dispatch at the start of each store_output
++ * lowering. The shader reads vs.xfb_topology sysval at runtime and branches
++ * into per-topology emission logic. For each affected topology, the lowered
++ * code emits guarded conditional stores — one per primitive this vertex
++ * contributes to, computing the output buffer position via primitive index
++ * and slot within the decomposed primitive.
++ *
++ * For LIST topologies (POINT/LINE/TRIANGLE LIST), takes a fast path that
++ * matches iter13's single-store behavior.
++ *
++ * For TRIANGLE_FAN, the central vertex (v=0) contributes to ALL primitives
++ * as slot 2 — handled via a NIR loop bounded by num_vertices.
++ *
++ * See ~/src/panvk-bifrost/iter17/phase{0,1,2}_*.md for full design context.
++ */
++
++#include "panvk_macros.h"
++
++#if PAN_ARCH < 9
++
++#include "panvk_shader.h"
++
++#include "compiler/nir/nir_builder.h"
++#include "pan_nir.h"
++
++#include <vulkan/vulkan_core.h>
++
++/* ----- Address arithmetic ----- */
++
++static nir_def *
++xfb_store_addr(nir_builder *b, nir_def *buf, nir_def *out_idx,
++               uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *byte_off = nir_iadd_imm(b,
++      nir_imul_imm(b, out_idx, stride), offset_bytes);
++   return nir_iadd(b, buf, nir_u2u64(b, byte_off));
++}
++
++static void
++emit_list_store(nir_builder *b, nir_def *buf, nir_def *output_count,
++                nir_def *instance_id, nir_def *raw_vid, nir_def *value,
++                uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *out_idx = nir_iadd(b,
++      nir_imul(b, instance_id, output_count), raw_vid);
++   nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
++   nir_store_global(b, value, addr);
++}
++
++static void
++emit_prim_store(nir_builder *b, nir_def *buf, nir_def *output_count,
++                nir_def *instance_id, nir_def *eligible,
++                nir_def *prim_idx, nir_def *slot,
++                uint32_t verts_per_prim,
++                nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_push_if(b, eligible);
++   {
++      nir_def *out_idx = nir_iadd(b,
++         nir_imul(b, instance_id, output_count),
++         nir_iadd(b, nir_imul_imm(b, prim_idx, verts_per_prim), slot));
++      nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
++      nir_store_global(b, value, addr);
++   }
++   nir_pop_if(b, NULL);
++}
++
++/* ----- Per-topology emission ----- */
++
++/* TRIANGLE_STRIP: vertex v contributes to prims v, v-1, v-2 (per eligibility). */
++static void
++emit_tri_strip(nir_builder *b, nir_def *v, nir_def *N,
++               nir_def *buf, nir_def *output_count, nir_def *instance_id,
++               nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *Nm2 = nir_iadd_imm(b, N, -2);
++   nir_def *Nm1 = nir_iadd_imm(b, N, -1);
++
++   /* Prim v, slot 0: v < N-2 */
++   emit_prim_store(b, buf, output_count, instance_id,
++      nir_ult(b, v, Nm2),
++      v, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
++
++   /* Prim v-1, slot = 1 if prim even else 2: 1 <= v < N-1 */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -1);
++      nir_def *parity = nir_iand_imm(b, prim, 1u);
++      nir_def *slot = nir_iadd_imm(b, parity, 1);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 1)),
++         nir_ult(b, v, Nm1));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, slot, 3, value, stride, offset_bytes);
++   }
++
++   /* Prim v-2, slot = 2 if prim even else 1: 2 <= v < N */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -2);
++      nir_def *parity = nir_iand_imm(b, prim, 1u);
++      nir_def *slot = nir_isub(b, nir_imm_int(b, 2), parity);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 2)),
++         nir_ult(b, v, N));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, slot, 3, value, stride, offset_bytes);
++   }
++}
++
++/* LINE_STRIP: vertex v contributes to prim v slot 0 + prim v-1 slot 1. */
++static void
++emit_line_strip(nir_builder *b, nir_def *v, nir_def *N,
++                nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *Nm1 = nir_iadd_imm(b, N, -1);
++
++   /* Prim v, slot 0: v < N-1 */
++   emit_prim_store(b, buf, output_count, instance_id,
++      nir_ult(b, v, Nm1),
++      v, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
++
++   /* Prim v-1, slot 1: 1 <= v < N */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -1);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 1)),
++         nir_ult(b, v, N));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
++   }
++}
++
++/* TRIANGLE_FAN: prim p emits {p+1, p+2, 0}.
++ *   vertex v=0: contributes to ALL prims as slot 2 (loop required)
++ *   vertex v>=1: contributes to prim v-1 as slot 0 (if 1 <= v <= N-2)
++ *   vertex v>=2: contributes to prim v-2 as slot 1 (if 2 <= v <= N-1)
++ */
++static void
++emit_tri_fan(nir_builder *b, nir_def *v, nir_def *N,
++             nir_def *buf, nir_def *output_count, nir_def *instance_id,
++             nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *Nm1 = nir_iadd_imm(b, N, -1);
++   nir_def *Nm2 = nir_iadd_imm(b, N, -2);
++
++   /* Prim v-1, slot 0: 1 <= v < N-1 */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -1);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 1)),
++         nir_ult(b, v, Nm1));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
++   }
++
++   /* Prim v-2, slot 1: 2 <= v < N */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -2);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 2)),
++         nir_ult(b, v, N));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 1), 3, value, stride, offset_bytes);
++   }
++
++   /* Central vertex (v == 0): loop over all prims, write to slot 2. */
++   nir_push_if(b, nir_ieq_imm(b, v, 0));
++   {
++      nir_variable *p_var = nir_local_variable_create(b->impl,
++         glsl_uint_type(), "fan_p");
++      nir_store_var(b, p_var, nir_imm_int(b, 0), 0x1);
++      nir_push_loop(b);
++      {
++         nir_def *p = nir_load_var(b, p_var);
++         nir_push_if(b, nir_uge(b, p, Nm2));
++         {
++            nir_jump(b, nir_jump_break);
++         }
++         nir_pop_if(b, NULL);
++
++         nir_def *out_idx = nir_iadd(b,
++            nir_imul(b, instance_id, output_count),
++            nir_iadd_imm(b, nir_imul_imm(b, p, 3), 2));
++         nir_def *addr = xfb_store_addr(b, buf, out_idx, stride, offset_bytes);
++         nir_store_global(b, value, addr);
++
++         nir_store_var(b, p_var, nir_iadd_imm(b, p, 1), 0x1);
++      }
++      nir_pop_loop(b, NULL);
++   }
++   nir_pop_if(b, NULL);
++}
++
++/* LINE_LIST_WITH_ADJACENCY: 4-vertex groups [4i..4i+3]; output {4i+1, 4i+2}.
++ *   v contributes if v%4 == 1: prim v/4 slot 0
++ *   v contributes if v%4 == 2: prim v/4 slot 1
++ */
++static void
++emit_line_list_adj(nir_builder *b, nir_def *v, nir_def *N,
++                   nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                   nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   (void)N; /* eligibility is mod-based, not range-based */
++   nir_def *vmod4 = nir_iand_imm(b, v, 3u);
++   nir_def *prim = nir_ushr_imm(b, v, 2);  /* v / 4 */
++
++   emit_prim_store(b, buf, output_count, instance_id,
++      nir_ieq_imm(b, vmod4, 1),
++      prim, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
++
++   emit_prim_store(b, buf, output_count, instance_id,
++      nir_ieq_imm(b, vmod4, 2),
++      prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
++}
++
++/* LINE_STRIP_WITH_ADJACENCY: prim p emits {p+1, p+2}.
++ *   v contributes to prim v-1 slot 0 (1 <= v <= N-2)
++ *   v contributes to prim v-2 slot 1 (2 <= v <= N-1)
++ */
++static void
++emit_line_strip_adj(nir_builder *b, nir_def *v, nir_def *N,
++                    nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                    nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   nir_def *Nm1 = nir_iadd_imm(b, N, -1);
++   nir_def *Nm2 = nir_iadd_imm(b, N, -2);
++
++   /* Prim v-1, slot 0: 1 <= v <= N-2 ⇔ v >= 1 AND v <= N-2 ⇔ v >= 1 AND v < N-1 */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -1);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 1)),
++         nir_ult(b, v, Nm1));
++      (void)Nm2;
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 0), 2, value, stride, offset_bytes);
++   }
++
++   /* Prim v-2, slot 1: 2 <= v <= N-1 ⇔ v >= 2 AND v < N */
++   {
++      nir_def *prim = nir_iadd_imm(b, v, -2);
++      nir_def *eligible = nir_iand(b,
++         nir_uge(b, v, nir_imm_int(b, 2)),
++         nir_ult(b, v, N));
++      emit_prim_store(b, buf, output_count, instance_id, eligible,
++                      prim, nir_imm_int(b, 1), 2, value, stride, offset_bytes);
++   }
++}
++
++/* TRIANGLE_LIST_WITH_ADJACENCY: 6-vertex groups; output {6i, 6i+2, 6i+4}.
++ *   v contributes if v%6 == 0: prim v/6 slot 0
++ *   v contributes if v%6 == 2: prim v/6 slot 1
++ *   v contributes if v%6 == 4: prim v/6 slot 2
++ */
++static void
++emit_tri_list_adj(nir_builder *b, nir_def *v, nir_def *N,
++                  nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                  nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   (void)N;
++   nir_def *vmod6 = nir_umod_imm(b, v, 6);
++   nir_def *prim = nir_udiv_imm(b, v, 6);
++
++   for (uint32_t slot = 0; slot < 3; slot++) {
++      emit_prim_store(b, buf, output_count, instance_id,
++         nir_ieq_imm(b, vmod6, slot * 2),
++         prim, nir_imm_int(b, slot), 3, value, stride, offset_bytes);
++   }
++}
++
++/* TRIANGLE_STRIP_WITH_ADJACENCY: prim i emits:
++ *   even i: {2i, 2i+2, 2i+4}    (slots 0, 1, 2 ← input indices 2i, 2i+2, 2i+4)
++ *   odd  i: {2i, 2i+4, 2i+2}    (slots 0, 1, 2 ← input indices 2i, 2i+4, 2i+2)
++ *
++ * Only EVEN input vertices contribute (since all output indices are 2*something).
++ * For even input v:
++ *   prim v/2 slot 0 (always, if v/2 < N/2-2)
++ *   prim (v-2)/2 slot 1 if (v-2)/2 even, slot 2 if odd   (when v >= 2)
++ *   prim (v-4)/2 slot 2 if (v-4)/2 even, slot 1 if odd   (when v >= 4)
++ */
++static void
++emit_tri_strip_adj(nir_builder *b, nir_def *v, nir_def *N,
++                   nir_def *buf, nir_def *output_count, nir_def *instance_id,
++                   nir_def *value, uint16_t stride, uint16_t offset_bytes)
++{
++   /* Bail for odd input vertices — they never contribute. */
++   nir_def *v_is_even = nir_ieq_imm(b, nir_iand_imm(b, v, 1u), 0);
++   nir_push_if(b, v_is_even);
++   {
++      nir_def *N_half = nir_ushr_imm(b, N, 1);
++      nir_def *max_prim = nir_iadd_imm(b, N_half, -2);  /* N/2 - 2 */
++      nir_def *v_half = nir_ushr_imm(b, v, 1);
++
++      /* Prim v/2 slot 0: v/2 < N/2 - 2 */
++      emit_prim_store(b, buf, output_count, instance_id,
++         nir_ult(b, v_half, max_prim),
++         v_half, nir_imm_int(b, 0), 3, value, stride, offset_bytes);
++
++      /* Prim (v-2)/2 = v/2 - 1: v >= 2 AND prim < N/2-2 */
++      {
++         nir_def *prim = nir_iadd_imm(b, v_half, -1);
++         nir_def *parity = nir_iand_imm(b, prim, 1u);
++         nir_def *slot = nir_iadd_imm(b, parity, 1);  /* even→1, odd→2 */
++         nir_def *eligible = nir_iand(b,
++            nir_uge(b, v, nir_imm_int(b, 2)),
++            nir_ult(b, prim, max_prim));
++         emit_prim_store(b, buf, output_count, instance_id, eligible,
++                         prim, slot, 3, value, stride, offset_bytes);
++      }
++
++      /* Prim (v-4)/2 = v/2 - 2: v >= 4 AND prim < N/2-2 */
++      {
++         nir_def *prim = nir_iadd_imm(b, v_half, -2);
++         nir_def *parity = nir_iand_imm(b, prim, 1u);
++         nir_def *slot = nir_isub(b, nir_imm_int(b, 2), parity);  /* even→2, odd→1 */
++         nir_def *eligible = nir_iand(b,
++            nir_uge(b, v, nir_imm_int(b, 4)),
++            nir_ult(b, prim, max_prim));
++         emit_prim_store(b, buf, output_count, instance_id, eligible,
++                         prim, slot, 3, value, stride, offset_bytes);
++      }
++   }
++   nir_pop_if(b, NULL);
++}
++
++/* ----- Main lowering: per store_output XFB channel ----- */
++
++static void
++lower_xfb_output_iter17(nir_builder *b, nir_intrinsic_instr *intr,
++                        unsigned channel_idx, unsigned num_components,
++                        unsigned buffer, unsigned offset_words)
++{
++   assert(buffer < MAX_XFB_BUFFERS);
++   assert(nir_intrinsic_component(intr) == 0);
++
++   uint16_t stride = b->shader->info.xfb_stride[buffer] * 4;
++   assert(stride != 0);
++   uint16_t offset_bytes = offset_words * 4;
++
++   BITSET_SET(b->shader->info.system_values_read, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
++   BITSET_SET(b->shader->info.system_values_read, SYSTEM_VALUE_INSTANCE_ID);
++
++   nir_def *topology = load_sysval(b, graphics, 32, vs.xfb_topology);
++   nir_def *out_count = load_sysval(b, graphics, 32, vs.xfb_output_count);
++   nir_def *N = nir_load_num_vertices(b);
++   nir_def *v = nir_load_raw_vertex_id_pan(b);
++   nir_def *instance = nir_load_instance_id(b);
++   nir_def *buf = nir_load_xfb_address(b, 64, .base = buffer);
++
++   nir_def *src = intr->src[0].ssa;
++   nir_component_mask_t mask = nir_component_mask(num_components);
++   nir_def *value = nir_channels(b, src, mask << channel_idx);
++
++   /* Topology dispatch ladder. LIST first (fast path). */
++   nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LIST));
++   {
++      emit_list_store(b, buf, out_count, instance, v, value,
++                      stride, offset_bytes);
++   }
++   nir_push_else(b, NULL);
++   {
++      /* iter17 Janet Finding 3: gate all non-LIST emission on
++       * output_count > 0. For degenerate input counts (N < min required
++       * for the topology), output_count is 0 and we must emit NO stores
++       * — otherwise N-2 / N-3 / etc. arithmetic underflows in the
++       * eligibility predicates and we falsely fire stores. */
++      nir_push_if(b, nir_ult(b, nir_imm_int(b, 0), out_count));
++      {
++      nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_STRIP));
++      {
++         emit_tri_strip(b, v, N, buf, out_count, instance, value,
++                        stride, offset_bytes);
++      }
++      nir_push_else(b, NULL);
++      {
++         nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_STRIP));
++         {
++            emit_line_strip(b, v, N, buf, out_count, instance, value,
++                            stride, offset_bytes);
++         }
++         nir_push_else(b, NULL);
++         {
++            nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_FAN));
++            {
++               emit_tri_fan(b, v, N, buf, out_count, instance, value,
++                            stride, offset_bytes);
++            }
++            nir_push_else(b, NULL);
++            {
++               nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_LIST_ADJ));
++               {
++                  emit_line_list_adj(b, v, N, buf, out_count, instance, value,
++                                     stride, offset_bytes);
++               }
++               nir_push_else(b, NULL);
++               {
++                  nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LINE_STRIP_ADJ));
++                  {
++                     emit_line_strip_adj(b, v, N, buf, out_count, instance, value,
++                                         stride, offset_bytes);
++                  }
++                  nir_push_else(b, NULL);
++                  {
++                     nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_TRI_LIST_ADJ));
++                     {
++                        emit_tri_list_adj(b, v, N, buf, out_count, instance, value,
++                                          stride, offset_bytes);
++                     }
++                     nir_push_else(b, NULL);
++                     {
++                        /* TRI_STRIP_ADJ — last case */
++                        emit_tri_strip_adj(b, v, N, buf, out_count, instance, value,
++                                           stride, offset_bytes);
++                     }
++                     nir_pop_if(b, NULL);
++                  }
++                  nir_pop_if(b, NULL);
++               }
++               nir_pop_if(b, NULL);
++            }
++            nir_pop_if(b, NULL);
++         }
++         nir_pop_if(b, NULL);
++      }
++      nir_pop_if(b, NULL);
++      }
++      nir_pop_if(b, NULL);  /* Janet Finding 3: close output_count > 0 guard */
++   }
++   nir_pop_if(b, NULL);
++}
++
++/* Mirror of pan_nir_lower_xfb's lower_xfb: load_vertex_id rewrite +
++ * dispatch store_output through our topology-aware emission. */
++static bool
++lower_xfb_iter17(nir_builder *b, nir_intrinsic_instr *intr,
++                 UNUSED void *data)
++{
++   if (intr->intrinsic == nir_intrinsic_load_vertex_id) {
++      b->cursor = nir_instr_remove(&intr->instr);
++      nir_def *repl = nir_iadd(b, nir_load_raw_vertex_id_pan(b),
++                               nir_load_raw_vertex_offset_pan(b));
++      nir_def_rewrite_uses(&intr->def, repl);
++      return true;
++   }
++
++   if (intr->intrinsic != nir_intrinsic_store_output)
++      return false;
++
++   bool progress = false;
++   b->cursor = nir_before_instr(&intr->instr);
++
++   /* io_xfb has only out[0,1]; the other 2 channels are in io_xfb2.
++    * Outer loop selects which annotation; inner picks which channel. */
++   for (unsigned i = 0; i < 2; ++i) {
++      nir_io_xfb xfb = i ? nir_intrinsic_io_xfb2(intr)
++                         : nir_intrinsic_io_xfb(intr);
++      for (unsigned j = 0; j < 2; ++j) {
++         if (!xfb.out[j].num_components)
++            continue;
++         lower_xfb_output_iter17(b, intr, i * 2 + j, xfb.out[j].num_components,
++                                 xfb.out[j].buffer, xfb.out[j].offset);
++         progress = true;
++      }
++   }
++
++   if (progress)
++      nir_instr_remove(&intr->instr);
++   return progress;
++}
++
++bool
++panvk_per_arch(nir_lower_xfb)(nir_shader *nir)
++{
++   return nir_shader_intrinsics_pass(
++      nir, lower_xfb_iter17, nir_metadata_control_flow, NULL);
++}
++
++#endif /* PAN_ARCH < 9 */

arch/mesa-panvk-bifrost/0005-panvk-bifrost-fragment-stores-atomics.patch

@@ -0,0 +1,50 @@
+From: marfrit-packages noether <claude-noether@reauktion.de>
+Subject: [PATCH] panvk: report fragmentStoresAndAtomics = true on Bifrost
+
+Backports Mesa main's unconditional advertisement of
+fragmentStoresAndAtomics for panvk (snapshot ref: src/panfrost/vulkan/
+panvk_vX_physical_device.c at commit-time 2026-05-06; the line reads
+`.fragmentStoresAndAtomics = true,` on main with no PAN_ARCH gate).
+
+Motivation: Chromium Dawn's WebGPU initializer in
+third_party/dawn/src/dawn/native/vulkan/PhysicalDeviceVk.cpp:250
+unconditionally rejects any Vulkan adapter that doesn't advertise this
+feature, causing Dawn to fall back to the SwiftShader CPU adapter
+on PineTab2 / RK3566 / Mali-G52 r1 MC1 (PAN_ARCH 7). With this patch the
+device advertises true, satisfying Dawn's gate. Tracked at
+https://git.reauktion.de/marfrit/panvk-bifrost/issues/2.
+
+The disjunction with `instance->force_enable_shader_atomics` is
+preserved as a kill-switch: in compiler terms it's dead code
+(`true || X == true`), but it leaves the DRI option
+`pan_force_enable_shader_atomics` semantically wired so future
+rebases or downstream debugging can see the link to the runtime knob.
+
+Caveat: the existing DRI option's description in src/util/driconf.h
+still labels this as "may not work reliably and is for debug purposes
+only". Mesa main's choice to ship it as default-on for all panvk
+architectures (including Bifrost, which is non-conformant per the
+PAN_I_WANT_A_BROKEN_VULKAN_DRIVER gate) reflects an upstream judgment
+that the practical risk is acceptable. Verify-before-ship for this
+package: dEQP-VK.glsl.atomic_operations.* + dEQP-VK.image.store.*
+deltas vs the r4 baseline must show no new fails. Pass counts may rise
+(tests that previously NotSupported now run); the load-bearing line is
+the Failed column staying at zero.
+
+---
+ src/panfrost/vulkan/panvk_vX_physical_device.c | 3 +--
+ 1 file changed, 1 insertion(+), 2 deletions(-)
+
+diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
+--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
+@@ -280,8 +280,7 @@
+       .vertexPipelineStoresAndAtomics =
+          (PAN_ARCH >= 13 && instance->enable_vertex_pipeline_stores_atomics) ||
+          instance->force_enable_shader_atomics,
+-      .fragmentStoresAndAtomics =
+-         (PAN_ARCH >= 10) || instance->force_enable_shader_atomics,
++      .fragmentStoresAndAtomics = true || instance->force_enable_shader_atomics,
+       .shaderTessellationAndGeometryPointSize = false,
+       .shaderImageGatherExtended = true,
+       .shaderStorageImageExtendedFormats = true,

arch/mesa-panvk-bifrost/0006-panvk-bifrost-legacy-dithering.patch

@@ -0,0 +1,51 @@
+From: marfrit-packages noether <claude-noether@reauktion.de>
+Subject: [PATCH] panvk: advertise VK_EXT_legacy_dithering on Bifrost
+
+Backports Mesa main's flip — vanilla 26.0.6 doesn't have the extension
+in the panvk advertisement list; main does (line 172 / 647 on snapshot
+617da94, 2026-05-06).
+
+VK_EXT_legacy_dithering exposes the classic OpenGL-style dithering
+behavior to Vulkan apps. Pure-software composition; no new HW path.
+ARM's own libmali driver release r51p0 (BXODROIDN2PL, Aug 2024) lists
+this extension in its Vulkan implementation for ODROID-N2 boards
+using the same Mali-G52 architecture family — confirms ARM ships it
+for Mali-G52-class hardware.
+
+Consumer benefit: dithering matters for low-bit-depth framebuffers
+(RGB565 / RGB5A1 — common on portable / battery-saving renders)
+where banding is visible. DXVK / vkd3d-proton both opt in when
+available.
+
+Verify-before-ship: vulkaninfo lists the extension and
+VkPhysicalDeviceLegacyDitheringFeaturesEXT.legacyDithering == true.
+
+Cross-refs:
+  - marfrit/panvk-bifrost research/r6_r7_mali_g52_feature_audit_2026-05-24.md
+  - ARM blob r51p0 strings dump (in-blob extension confirmed)
+
+---
+ src/panfrost/vulkan/panvk_vX_physical_device.c | 5 +++++
+ 1 file changed, 5 insertions(+)
+
+diff --git a/src/panfrost/vulkan/panvk_vX_physical_device.c b/src/panfrost/vulkan/panvk_vX_physical_device.c
+--- a/src/panfrost/vulkan/panvk_vX_physical_device.c
++++ b/src/panfrost/vulkan/panvk_vX_physical_device.c
+@@ -156,6 +156,7 @@
+       .EXT_image_drm_format_modifier = true,
+       .EXT_image_robustness = true,
+       .EXT_index_type_uint8 = true,
++      .EXT_legacy_dithering = true,
+       .EXT_line_rasterization = true,
+       .EXT_load_store_op_none = true,
+       .EXT_non_seamless_cube_map = true,
+@@ -552,6 +553,9 @@
+
+       /* VK_EXT_multisampled_render_to_single_sampled */
+       .multisampledRenderToSingleSampled = true,
++
++      /* VK_EXT_legacy_dithering */
++      .legacyDithering = true,
+    };
+ }
+

arch/mesa-panvk-bifrost/0007-panvk-bifrost-xfb-component-base-fix.patch

@@ -0,0 +1,103 @@
+From: marfrit-packages noether <claude-noether@reauktion.de>
+Subject: [PATCH] panvk-bifrost: fix XFB store channel-extract for packed varyings
+
+iter19 — fixes a reliable SIGSEGV during vkCreateGraphicsPipeline on any
+shader that uses XFB-bound varyings declared with non-zero `layout
+(component=N)` qualifiers. Surfaced by
+dEQP-VK.transform_feedback.simple.holes_vert; backtrace lands 11 frames
+into libvulkan_panfrost.so called from `vkt::TransformFeedback::
+TransformFeedbackHolesInstance::iterate`.
+
+Root cause: `lower_xfb_output_iter17` (and upstream `lower_xfb_output`,
+which carries a `// TODO` on the same assertion) computes the source-
+channel mask as `mask << channel_idx`, where `channel_idx` is the
+varying-location component (0..3) but `src` only contains channels for
+the source-side range starting at `nir_intrinsic_component(intr)`. For
+`flat out float vegeta` declared with `component=2`, NIR emits
+`store_output src=<vec1>, component=2`, and the lowering computes
+`mask << 2` against a single-component src — out-of-range; the
+resulting malformed nir_def then segfaults inside downstream NIR
+constant-folding (`nir_constant_expressions.c::evaluate_*`).
+
+The assertion `assert(nir_intrinsic_component(intr) == 0)` was inherited
+from upstream `pan_nir_lower_xfb.c` as a documented `// TODO`; release
+builds (-DNDEBUG) elide it. The fix translates `channel_idx` to the
+source-channel space by subtracting `nir_intrinsic_component(intr)`
+before shifting the mask, and replaces the elided asserts with explicit
+release-mode guards (the patch closes the same release-mode-elision
+class as the original bug).
+
+Verified on PineTab2 (Mali-G52 r1 MC1, PAN_ARCH 7) against vulkan-cts
+1.3.10.0:
+  - holes_vert / holes_extra_draw_vert no longer SIGSEGV (now Fail on
+    color-check; that is a separate iter20 finding — the rasterized
+    varying gets removed alongside the XFB-bound one).
+  - basic_*: 36/36 Pass. depth_clip_*: 1 Pass + 4 NotSupported.
+    lines_or_triangles*: 16 NotSupported. 0 Fail across the full set.
+  - holes_geom / holes_extra_draw_geom remain NotSupported
+    (geometryShader not on G52) — unchanged.
+
+Caveat: max_output_components_64/_128/_256 were never reached on the
+r5 sweep (watchdog killed transform_feedback after the holes_vert
+crash). With this fix in place, those tests now run and surface
+*their own pre-existing* coredumps — confirmed on shipped r6 baseline
+too. They are NOT regressions from this patch; they are latent crashes
+unmasked by it. iter20+ territory.
+
+Phase 5 (2nd-model) review: APPROVE WITH CHANGES (non-blocking).
+Changes applied: release-mode defensive guards on both preconditions
+plus a dispatcher-side comment clarifying the i*2+j semantics.
+
+Cross-refs:
+  - iter19/phase{0,1,2,3}_holes_vert*.md in panvk-bifrost repo
+
+---
+ src/panfrost/vulkan/panvk_vX_xfb_lower.c | 24 +++++++++++++++++++++---
+ 1 file changed, 21 insertions(+), 3 deletions(-)
+
+diff --git a/src/panfrost/vulkan/panvk_vX_xfb_lower.c b/src/panfrost/vulkan/panvk_vX_xfb_lower.c
+@@ -339,7 +339,20 @@
+                         unsigned buffer, unsigned offset_words)
+ {
+    assert(buffer < MAX_XFB_BUFFERS);
+-   assert(nir_intrinsic_component(intr) == 0);
++
++   /* iter19: nir_intrinsic_component(intr) is the source-channel base —
++    * for a packed varying like `layout (location=0, component=2) flat out
++    * float vegeta`, NIR emits store_output with component=2 and a single-
++    * component src. The XFB iteration index `channel_idx` (0..3) is the
++    * varying-location component, not the source channel. Translate by
++    * subtracting the base before shifting the mask. Fixes the long-
++    * standing `assert(nir_intrinsic_component(intr) == 0) // TODO` in
++    * upstream pan_nir_lower_xfb that surfaces on holes_vert. */
++   const unsigned base_comp = nir_intrinsic_component(intr);
++   /* Defensive against release-build elision: this is precisely the
++    * bug class the patch is fixing, so don't re-introduce it. */
++   if (channel_idx < base_comp)
++      return;
+ 
+    uint16_t stride = b->shader->info.xfb_stride[buffer] * 4;
+    assert(stride != 0);
+@@ -357,7 +370,11 @@
+ 
+    nir_def *src = intr->src[0].ssa;
+    nir_component_mask_t mask = nir_component_mask(num_components);
+-   nir_def *value = nir_channels(b, src, mask << channel_idx);
++   const unsigned src_channel = channel_idx - base_comp;
++   /* Same defensive class as the channel_idx >= base_comp guard above. */
++   if (src_channel + num_components > src->num_components)
++      return;
++   nir_def *value = nir_channels(b, src, mask << src_channel);
+ 
+    /* Topology dispatch ladder. LIST first (fast path). */
+    nir_push_if(b, nir_ieq_imm(b, topology, PANVK_XFB_TOPO_LIST));
+@@ -465,6 +482,9 @@
+       for (unsigned j = 0; j < 2; ++j) {
+          if (!xfb.out[j].num_components)
+             continue;
++         /* `i*2+j` is the varying-location component (0..3) — io_xfb covers
++          * slots 0..1, io_xfb2 covers 2..3. The leaf translates this into
++          * a source-channel index by subtracting nir_intrinsic_component(intr). */
+          lower_xfb_output_iter17(b, intr, i * 2 + j, xfb.out[j].num_components,
+                                  xfb.out[j].buffer, xfb.out[j].offset);
+          progress = true;

arch/mesa-panvk-bifrost/PKGBUILD

@@ -0,0 +1,229 @@
+# Maintainer: Markus Fritsche <fritsche.markus@gmail.com>
+#
+# mesa-panvk-bifrost — patched Mesa libvulkan_panfrost.so that exposes
+# Bifrost-gen Mali (Mali-G31/G52/G72/G76, PAN_ARCH 6/7) to Chromium-family
+# browsers' Vulkan compositor.
+#
+# Campaign: ~/src/panvk-bifrost/ — iter9 close (technical milestone
+# 2026-05-20, operator-confirmed Brave window). Goal close requires
+# this package built + published + installable per the
+# feedback_package_done_means_installable.md 3-point check.
+#
+# What it does:
+#   - Builds Mesa 26.0.6 (matching ohm's stock mesa pkg) with two
+#     patches against src/panfrost/vulkan/panvk_vX_physical_device.c:
+#     (1) expose VK_KHR/EXT_robustness2 + nullDescriptor on PAN_ARCH 6/7
+#     (2) set has_vk1_1/has_vk1_2 to true on Bifrost
+#   - Installs the patched libvulkan_panfrost.so to /usr/lib/panvk-bifrost/
+#     (NOT /usr/lib — co-installs alongside stock mesa).
+#   - Registers a custom ICD JSON at a non-default path so the system
+#     Vulkan loader only picks up our driver when VK_ICD_FILENAMES is set.
+#   - Ships /usr/bin/brave-vulkan launcher that wires up env vars + flags.
+#
+# Co-existence: stock /usr/lib/libvulkan_panfrost.so is untouched. Stock
+# /usr/share/vulkan/icd.d/panfrost_icd.json is untouched. Users opt in
+# via brave-vulkan (or by setting VK_ICD_FILENAMES manually).
+#
+# Build target: arch-aarch64 runner via marfrit-packages Gitea Actions.
+# Mesa build is slow (~30-60min on Cortex-A55). Build deps installed
+# inside the runner via pacman -S in the workflow.
+
+pkgname=mesa-panvk-bifrost
+_mesaver=26.0.6
+pkgver=26.0.6.r7
+pkgrel=1
+pkgdesc="Patched Mesa libvulkan_panfrost.so exposing Bifrost-gen Mali to Vulkan apps (panvk-bifrost campaign)"
+arch=('aarch64')
+url="https://git.reauktion.de/marfrit/panvk-bifrost"
+license=('MIT')
+
+# We co-install at /usr/lib/panvk-bifrost/ so no conflicts with stock mesa.
+# We DO provide a script that requires brave to be installed.
+depends=(
+    'mesa'              # for shared mesa runtime libs (libgallium-mesa etc.)
+    'libdrm'
+    'wayland'
+    'libxcb'
+    'libx11'
+    'libxshmfence'
+    'zlib'
+    'zstd'
+    'libdisplay-info'
+    'expat'
+    'systemd-libs'      # libudev
+    'spirv-tools'
+)
+optdepends=(
+    'brave-bin: for the brave-vulkan launcher script'
+)
+makedepends=(
+    'meson'
+    'ninja'
+    'pkgconf'
+    'bison'
+    'flex'
+    'python'
+    'python-yaml'
+    'python-mako'
+    'glslang'
+    'spirv-tools'
+    'llvm'
+    'clang'
+    'libclc'
+    'spirv-llvm-translator'
+    'vulkan-headers'
+    'wayland-protocols'
+)
+
+source=(
+    "https://archive.mesa3d.org/mesa-${_mesaver}.tar.xz"
+    "0001-panvk-expose-robustness2-nullDescriptor-bifrost.patch"
+    "0002-panvk-expose-vulkan-1.1-1.2-on-bifrost.patch"
+    "0003-panvk-bifrost-vk-ext-transform-feedback.patch"
+    "0004-panvk-bifrost-xfb-primitive-decomposition.patch"
+    "0005-panvk-bifrost-fragment-stores-atomics.patch"
+    "0006-panvk-bifrost-legacy-dithering.patch"
+    "0007-panvk-bifrost-xfb-component-base-fix.patch"
+    "brave-vulkan"
+    "icd.json"
+)
+sha256sums=(
+    'SKIP'  # TODO: pin once we know the upstream tarball is stable. archive.mesa3d.org tarballs are stable, so we can hash-pin in iter10.
+    'SKIP'
+    'SKIP'
+    'SKIP'
+    'SKIP'
+    'SKIP'
+    'SKIP'
+    'SKIP'
+    'SKIP'
+    'SKIP'
+)
+
+prepare() {
+    cd "mesa-${_mesaver}"
+    # iter8 patch: robustness2 + nullDescriptor exposure on Bifrost.
+    # Three hunks in panvk_vX_physical_device.c. Apply via sed since
+    # the upstream context drifts between Mesa releases and unified-diff
+    # patching is brittle (we hit hunk-rejects during manual application).
+    sed -i 's|\.KHR_robustness2 = PAN_ARCH >= 10,|.KHR_robustness2 = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
+    sed -i 's|\.EXT_robustness2 = PAN_ARCH >= 10,|.EXT_robustness2 = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
+    sed -i 's|\.nullDescriptor = PAN_ARCH >= 10,|.nullDescriptor = true,|' src/panfrost/vulkan/panvk_vX_physical_device.c
+    # iter9 patch: bump api version to 1.1/1.2 for Bifrost. NOTE: this only
+    # affects extension exposure flags, NOT the reported apiVersion (which
+    # is set by get_api_version() further down and gated separately).
+    # The brave-vulkan launcher sets MESA_VK_VERSION_OVERRIDE=1.2 at runtime
+    # to deal with the latter — no source change needed for that path.
+    sed -i 's|bool has_vk1_1 = PAN_ARCH >= 10;|bool has_vk1_1 = true;|' src/panfrost/vulkan/panvk_vX_physical_device.c
+    sed -i 's|bool has_vk1_2 = PAN_ARCH >= 10;|bool has_vk1_2 = true;|' src/panfrost/vulkan/panvk_vX_physical_device.c
+
+    # iter13: VK_EXT_transform_feedback implementation for Bifrost (PAN_ARCH<9).
+    # Applied as a real unified-diff patch — the change is too large for sed.
+    # Phase-doc context: ~/src/panvk-bifrost/phase{4,5,6}_iter13_close.md.
+    # Unlocks ANGLE-Vulkan → GLES3 → WebGL2 / WebGPU on Brave (chrome://gpu
+    # reports "Hardware accelerated" across the board for the affected paths).
+    patch -p1 < "${srcdir}/0003-panvk-bifrost-vk-ext-transform-feedback.patch"
+
+    # iter17: XFB primitive decomposition for non-LIST topologies (TRI_STRIP,
+    # TRI_FAN, LINE_STRIP, *_WITH_ADJACENCY). Replacement panvk-specific
+    # NIR pass (panvk_per_arch(nir_lower_xfb)) substituted for upstream
+    # pan_nir_lower_xfb. Closes the 162 dEQP-VK winding_* failures from
+    # iter15 (958 P / 81 F / 0 Crash on full XFB CTS — remaining 81 fails
+    # are by-design resume_* tests, transformFeedbackDraw=false).
+    # Phase-doc context: ~/src/panvk-bifrost/iter17/phase{0,1,2,4,5,6,8}_*.md.
+    patch -p1 < "${srcdir}/0004-panvk-bifrost-xfb-primitive-decomposition.patch"
+
+    # r5 (2026-05-23): advertise .fragmentStoresAndAtomics = true on Bifrost
+    # to satisfy Chromium Dawn's WebGPU init gate
+    # (third_party/dawn/src/dawn/native/vulkan/PhysicalDeviceVk.cpp:250).
+    # Backports Mesa main's unconditional flip (same line as on main as of
+    # 2026-05-06). Disjunction with instance->force_enable_shader_atomics
+    # is preserved as a documented kill-switch even though the compiler
+    # folds it away. Closes marfrit/panvk-bifrost#2.
+    # Verify-before-ship: dEQP-VK.glsl.atomic_operations.* and
+    # dEQP-VK.image.store.* show no new Failed vs r4 baseline.
+    patch -p1 < "${srcdir}/0005-panvk-bifrost-fragment-stores-atomics.patch"
+
+    # r6 (2026-05-25): advertise VK_EXT_legacy_dithering. Backports Mesa
+    # main's unconditional flip. Pure-software composition; vk_render_pass
+    # already gates on enabled_features.legacyDithering and panvk_vX_blend
+    # + pan_format already plumb the dithered BLEND descriptor (BFMT2 table
+    # has MALI_BLEND_AU encodings for RGB565/RGB5A1/RGBA4/RGB10A2 on
+    # PAN_ARCH 7). Closes the EXT_legacy_dithering gap surfaced by
+    # marfrit/panvk-bifrost research/r6_r7_*. ARM blob r51p0 confirms the
+    # extension as Mali-G52-architecture supported.
+    patch -p1 < "${srcdir}/0006-panvk-bifrost-legacy-dithering.patch"
+
+    # r7 (2026-05-25): XFB store channel-extract fix for packed varyings.
+    # Eliminates a reliable SIGSEGV in vkCreateGraphicsPipeline whenever
+    # an XFB-bound vertex output is declared with non-zero
+    # `layout (component=N)`. Surfaced by dEQP-VK.transform_feedback.
+    # simple.holes_vert (now Fails on color-check rather than crashing;
+    # the color-check residual is a separate iter20 finding).
+    # Phase-doc context: ~/src/panvk-bifrost/iter19/phase{0,1,2,3}_*.md.
+    # Phase 5 reviewed; release-mode-elision defensive guards applied.
+    patch -p1 < "${srcdir}/0007-panvk-bifrost-xfb-component-base-fix.patch"
+
+    # Sanity-check the patches landed.
+    grep -q "KHR_robustness2 = true," src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "EXT_robustness2 = true," src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "nullDescriptor = true," src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "has_vk1_1 = true;" src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "has_vk1_2 = true;" src/panfrost/vulkan/panvk_vX_physical_device.c
+    # iter13 sanity:
+    grep -q "EXT_transform_feedback = PAN_ARCH < 9," src/panfrost/vulkan/panvk_vX_physical_device.c
+    test -f src/panfrost/vulkan/jm/panvk_vX_cmd_xfb.c
+    # iter17 sanity: pan_nir_lower_xfb call site has been replaced; new file present.
+    grep -q "panvk_per_arch(nir_lower_xfb)" src/panfrost/vulkan/panvk_vX_shader.c
+    # r5 sanity: fragmentStoresAndAtomics = true patch landed
+    grep -q "fragmentStoresAndAtomics = true ||" src/panfrost/vulkan/panvk_vX_physical_device.c
+    # r6 sanity: VK_EXT_legacy_dithering advertised
+    grep -q '\.EXT_legacy_dithering = true,' src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q '\.legacyDithering = true,' src/panfrost/vulkan/panvk_vX_physical_device.c
+    grep -q "xfb_topology" src/panfrost/vulkan/panvk_shader.h
+    grep -q "panvk_xfb_topology" src/panfrost/vulkan/panvk_shader.h
+    test -f src/panfrost/vulkan/panvk_vX_xfb_lower.c
+    # r7 sanity: XFB channel-base correction landed
+    grep -q "iter19: nir_intrinsic_component(intr) is the source-channel base" src/panfrost/vulkan/panvk_vX_xfb_lower.c
+    grep -q "mask << src_channel" src/panfrost/vulkan/panvk_vX_xfb_lower.c
+}
+
+build() {
+    cd "mesa-${_mesaver}"
+    # Minimal Mesa build — only the panvk Vulkan driver. No GLES/GLX/EGL/
+    # gallium drivers — keeps the build narrower and avoids pulling in
+    # the entire Mesa runtime in this package's dep graph.
+    meson setup build/ \
+        --prefix=/usr \
+        --libdir=lib \
+        --buildtype=release \
+        -Dvulkan-drivers=panfrost \
+        -Dgallium-drivers= \
+        -Dplatforms=wayland,x11 \
+        -Dglx=disabled \
+        -Degl=disabled \
+        -Dgles1=disabled \
+        -Dgles2=disabled \
+        -Dvulkan-layers= \
+        -Dtools= \
+        -Dgallium-rusticl=false \
+        -Dmicrosoft-clc=disabled
+    meson compile -C build
+}
+
+package() {
+    cd "${srcdir}/mesa-${_mesaver}"
+
+    # Patched lib — co-install path, NOT /usr/lib (to avoid clashing
+    # with stock mesa's libvulkan_panfrost.so).
+    install -Dm755 build/src/panfrost/vulkan/libvulkan_panfrost.so \
+        "$pkgdir/usr/lib/panvk-bifrost/libvulkan_panfrost.so"
+
+    # Custom ICD JSON. NOT under /usr/share/vulkan/icd.d/ (the default
+    # loader search path) — the user has to opt in via VK_ICD_FILENAMES.
+    install -Dm644 "$srcdir/icd.json" \
+        "$pkgdir/usr/lib/panvk-bifrost/icd.json"
+
+    # The brave-vulkan launcher wires up env + flags.
+    install -Dm755 "$srcdir/brave-vulkan" "$pkgdir/usr/bin/brave-vulkan"
+}

arch/mesa-panvk-bifrost/README.md

@@ -0,0 +1,83 @@
+# mesa-panvk-bifrost
+
+Patched Mesa `libvulkan_panfrost.so` exposing Bifrost-gen Mali GPUs
+(Mali-G31/G52/G72/G76, PAN_ARCH 6/7) to Chromium-family browsers'
+Vulkan compositor.
+
+Result of the **panvk-bifrost** campaign ([`~/src/panvk-bifrost/`](../../../panvk-bifrost/)),
+iter9 close 2026-05-20 (technical milestone — operator-confirmed Brave
+window) followed by this package as the actual delivery.
+
+## What it solves
+
+Stock Chromium / Brave on Bifrost SBCs (PineTab2, etc.) currently dies
+at GL bindings init:
+
+```
+ERROR: ui/gl/gl_context_egl.cc:120  GLES3 is unsupported and ES version fallback is disabled
+ERROR: ui/gl/init/gl_factory.cc:111  Requested GL implementation not found
+ERROR: components/viz/service/main/viz_main_impl.cc:189  Exiting GPU process due to errors during initialization
+```
+
+This package makes Brave's **Vulkan compositor path** work on Bifrost,
+side-stepping the failing GL stack. Browser chrome and standard page
+rendering work.
+
+## Known limitations (not addressed)
+
+- **WebGL / WebGL2** in-page: blocked. ANGLE needs `VK_EXT_transform_feedback`
+  to expose GLES3 contexts; PanVk-Bifrost doesn't currently support that.
+  Sites using WebGL will degrade or refuse.
+- **VAAPI hardware video decode**: unrelated to this package — see the
+  `libva-v4l2-request-fourier` package for that path.
+
+## Install
+
+```sh
+# Ensure [marfrit] is in /etc/pacman.conf, then:
+sudo pacman -Sy mesa-panvk-bifrost
+```
+
+## Use
+
+```sh
+brave-vulkan                                    # launches Brave with Vulkan
+brave-vulkan https://www.example.com
+brave-vulkan --your-flags-here                  # extra args passed through
+```
+
+The launcher sets:
+
+- `VK_ICD_FILENAMES=/usr/lib/panvk-bifrost/icd.json` (the patched driver)
+- `PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1` (Mesa upstream gate)
+- `MESA_VK_VERSION_OVERRIDE=1.2` (apiVersion bump for ANGLE)
+- Brave flags: `--use-gl=disabled --enable-features=Vulkan --use-vulkan=native --ozone-platform=x11 --no-sandbox --disable-gpu-sandbox --ignore-gpu-blocklist`
+
+## What's in the package
+
+- `/usr/lib/panvk-bifrost/libvulkan_panfrost.so` — patched Mesa Vulkan driver (Mesa 26.0.6 + 2 sed-applied patches)
+- `/usr/lib/panvk-bifrost/icd.json` — Vulkan ICD JSON pointing at the patched .so (NOT auto-loaded; only via `VK_ICD_FILENAMES`)
+- `/usr/bin/brave-vulkan` — launcher script
+
+System Mesa is untouched. The stock `/usr/lib/libvulkan_panfrost.so` and
+`/usr/share/vulkan/icd.d/panfrost_icd.json` continue to work for any
+other Vulkan app.
+
+## Co-existence
+
+Both stock Mesa and this package can be installed. Stock Vulkan apps
+(`vulkaninfo`, `vkcube`, etc.) use the stock driver by default. Only
+apps started via `brave-vulkan` (or with `VK_ICD_FILENAMES` manually
+set to our path) use the patched driver.
+
+## Campaign close criterion
+
+Per [`feedback_package_done_means_installable`](file:///home/mfritsche/.claude/projects/-home-mfritsche-src/memory/feedback_package_done_means_installable.md)
+three-point check:
+
+1. PR merged to `marfrit-packages`
+2. CI green AND `packages.reauktion.de/arch/aarch64/mesa-panvk-bifrost-*.pkg.tar.zst` exists
+3. `pacman -Ss mesa-panvk-bifrost` on a fresh consumer host (e.g. ohm
+   after `pacman -Syu` from clean state) returns the package + brave-vulkan launches
+
+When all three pass, panvk-bifrost iter9 closes.

arch/mesa-panvk-bifrost/brave-vulkan

@@ -0,0 +1,61 @@
+#!/bin/bash
+# brave-vulkan — launch Brave with the PanVk-Bifrost Vulkan compositor on
+# Bifrost SBCs (PineTab2 / Mali-G52 r1 MC1 and similar).
+#
+# Side-steps the GL stack failures stock Brave hits on Bifrost
+# (the README "Consumer-side benefit" path of the panvk-bifrost campaign).
+#
+# Provided by the mesa-panvk-bifrost package. See:
+#   /usr/share/doc/mesa-panvk-bifrost/README
+#   ~/src/panvk-bifrost/phase8_iteration9_close.md (campaign close)
+#
+# Usage: brave-vulkan [brave args...]
+# Equivalent to: brave [VULKAN_FLAGS] [your args]
+
+set -e
+
+# Patched Vulkan driver (from this package) — must point at the custom path
+# so we don't clash with the stock /usr/share/vulkan/icd.d/panfrost_icd.json
+export VK_ICD_FILENAMES=/usr/lib/panvk-bifrost/icd.json
+
+# PanVk's "I know it's not conformant" gate — the patched driver still
+# refuses to enumerate Bifrost without this env var (Mesa upstream choice,
+# kept for compatibility).
+export PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1
+
+# Override apiVersion to 1.2 — ANGLE (Chromium's GL stack) requires
+# device.apiVersion >= 1.1. The patched libvulkan_panfrost.so still has
+# a PAN_ARCH>=10 gate inside get_api_version(); easier to override at
+# runtime via this Mesa env var than to add a third patch.
+export MESA_VK_VERSION_OVERRIDE=1.2
+
+# Find the live Plasma session's Xauthority. On a fresh boot the suffix
+# is randomized; pgrep the Xwayland args to find the current one.
+if [ -z "${XAUTHORITY:-}" ]; then
+    XAUTHF=$(pgrep -fa Xwayland 2>/dev/null | grep -oE "/run/user/$(id -u)/xauth_[A-Za-z0-9]+" | head -1)
+    if [ -n "$XAUTHF" ]; then
+        export XAUTHORITY="$XAUTHF"
+    fi
+fi
+
+# Standard session env if not already set
+: "${XDG_RUNTIME_DIR:=/run/user/$(id -u)}"
+export XDG_RUNTIME_DIR
+
+# Default to the active Plasma Wayland session if WAYLAND_DISPLAY unset
+: "${WAYLAND_DISPLAY:=wayland-0}"
+export WAYLAND_DISPLAY
+
+# Default to the XWayland :1 unless DISPLAY is set
+: "${DISPLAY:=:1}"
+export DISPLAY
+
+exec brave \
+    --use-gl=disabled \
+    --enable-features=Vulkan \
+    --use-vulkan=native \
+    --ozone-platform=x11 \
+    --no-sandbox \
+    --disable-gpu-sandbox \
+    --ignore-gpu-blocklist \
+    "$@"

arch/mesa-panvk-bifrost/icd.json

@@ -0,0 +1,7 @@
+{
+    "ICD": {
+        "api_version": "1.4.335",
+        "library_path": "/usr/lib/panvk-bifrost/libvulkan_panfrost.so"
+    },
+    "file_format_version": "1.0.1"
+}

arch/qt6-base-fourier/PKGBUILD

@@ -21,14 +21,14 @@
 # pkgbase stays as qt6-base so $_pkgfn (= ${pkgbase/6-/} = "qtbase")
 # resolves correctly. The "-fourier" suffix lives only in the
 # directory name and the commit history; epoch=1 gives our local
-# build strict precedence over upstream pkgrel=2 until upstream lands
+# build strict precedence over upstream pkgrel=N until upstream lands
 # the GL_R8/ES3 fix and we can drop the epoch.
 pkgbase=qt6-base
 pkgname=(qt6-base-fourier
          qt6-xcb-private-headers-fourier)
-_pkgver=6.11.0
+_pkgver=6.11.1
 pkgver=${_pkgver/-/}
-pkgrel=3
+pkgrel=2
 epoch=1
 arch=(aarch64 x86_64)
 url='https://www.qt.io'
@@ -42,7 +42,6 @@ depends=(brotli
          double-conversion
          fontconfig
          freetype2
-         gcc-libs
          glib2
          glibc
          harfbuzz
@@ -51,6 +50,7 @@ depends=(brotli
          libb2
          libcups
          libdrm
+         libgcc
          libgl
          libice
          libinput
@@ -58,6 +58,7 @@ depends=(brotli
          libpng
          libproxy
          libsm
+         libstdc++
          liburing
          libx11
          libxcb
@@ -115,7 +116,7 @@ source=(git+https://code.qt.io/qt/$_pkgfn#tag=v$_pkgver
         0001-qopengltextureglyphcache-pick-GL_R8-on-ES3.patch
         0002-qrhigles2-RED_OR_ALPHA8-pick-GL_R8-on-ES3.patch
         0003-qopengltextureuploader-pick-GL_R8-on-ES3.patch)
-sha256sums=('2223c075e95d86f8dbf6395b025a74d996c418f094453c903290e3c2663fbed2'
+sha256sums=('2eafe504fae873d20f206b5661e2e10506879455cb2d370f42c5bb72ccf7a8a1'
             '5411edbe215c24b30448fac69bd0ba7c882f545e8cf05027b2b6e2227abc5e78'
             '4b93f6a79039e676a56f9d6990a324a64a36f143916065973ded89adc621e094'
             'SKIP'
@@ -126,9 +127,8 @@ prepare() {
   patch -d $_pkgfn -p1 < qt6-base-cflags.patch # Use system CFLAGS
   patch -d $_pkgfn -p1 < qt6-base-nostrip.patch # Don't strip binaries with qmake
 
-  # cherry-pick needs git author identity; git-cli refuses without it.
-  git -C $_pkgfn -c user.email=fourier@build -c user.name='qt6-base-fourier build' \
-      cherry-pick -n 8b54513cdcf62047376a5d27d784ad68a8f235bf # Fix qdbus crashes
+  # 8b54513cdcf6 (qdbus crash fix) cherry-pick removed: landed upstream
+  # in 6.11.1. Re-add if qdbus regressions re-surface.
 
   # qt6-base-fourier — three small runtime-checks that pick GL_R8 over
   # GL_ALPHA when the live GL context is ES 3 or newer. See the
@@ -165,7 +165,7 @@ build() {
     -DFEATURE_system_sqlite=ON \
     -DFEATURE_system_xcb_xinput=ON \
     -DFEATURE_no_direct_extern_access=$_no_direct_extern_access \
-    -DFEATURE_sql_ibase=OFF \
+    -DFEATURE_mimetype_database=OFF \
     -DCMAKE_INTERPROCEDURAL_OPTIMIZATION=ON \
     -DCMAKE_MESSAGE_LOG_LEVEL=STATUS
   cmake --build build
@@ -188,7 +188,7 @@ package_qt6-xcb-private-headers-fourier() {
   conflicts=(qt6-xcb-private-headers)
   replaces=(qt6-xcb-private-headers)
 
-  depends=("qt6-base-fourier=$pkgver")
+  depends=("qt6-base-fourier=$epoch:$pkgver")
   optdepends=()
   groups=()
 

debian/aish/build-deb.sh

@@ -0,0 +1,85 @@
+#!/bin/bash
+# Build aish_<ver>_all.deb from this directory using dpkg-deb directly.
+# Run from inside the runner container, which has dpkg installed.
+#
+# Matches the lmcp build-deb.sh pattern: no dh/debhelper, no Build-Depends
+# beyond `dpkg`, structurally a normal apt package (Architecture: all).
+set -euo pipefail
+
+PKGVER=0.1.0
+UPSTREAM_TAG=v0.1.0
+PKGREL=1
+AISH_TARBALL_SHA256=9ebc3939e028832e39391ae33efacb5ec9bcd99d123cbc8ca1cd6ca9a640b5b5
+HERE=$(dirname "$(readlink -f "$0")")
+
+# Reproducible build: pin all file mtimes + ar member timestamps to a fixed
+# epoch tied to this packaging release (aish v0.1.0 — 2026-05-25 00:00 UTC).
+# Without this, repeat builds produce different byte streams and reprepro
+# refuses re-includes with "size expected: X, got: Y".
+export SOURCE_DATE_EPOCH=1779667200
+
+work=$(mktemp -d)
+trap "rm -rf $work" EXIT
+
+cd "$work"
+curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo aish.tar.gz \
+    "https://git.reauktion.de/marfrit/aish/archive/${UPSTREAM_TAG}.tar.gz"
+echo "$AISH_TARBALL_SHA256  aish.tar.gz" | sha256sum -c
+tar xzf aish.tar.gz
+
+ROOT="$work/pkgroot"
+LIBDIR="$ROOT/usr/share/lua/5.1/aish"
+mkdir -p "$ROOT/DEBIAN" \
+         "$LIBDIR/ffi" \
+         "$LIBDIR/vendor" \
+         "$ROOT/usr/bin" \
+         "$ROOT/usr/share/doc/aish/examples"
+
+# Top-level modules
+for m in main broker context executor history mcp renderer repl router safety secrets; do
+    cp "aish/${m}.lua" "$LIBDIR/${m}.lua"
+done
+
+# FFI bindings
+for m in curl libc pty readline; do
+    cp "aish/ffi/${m}.lua" "$LIBDIR/ffi/${m}.lua"
+done
+
+# Vendored dependencies
+cp aish/vendor/dkjson.lua "$LIBDIR/vendor/dkjson.lua"
+
+# Launch wrapper
+install -m 755 aish/bin/aish "$ROOT/usr/bin/aish"
+
+# Documentation + example config
+cp aish/README.md          "$ROOT/usr/share/doc/aish/"
+cp aish/LICENSE            "$ROOT/usr/share/doc/aish/"
+cp aish/examples/config.lua "$ROOT/usr/share/doc/aish/examples/"
+cp "$HERE/debian/copyright" "$ROOT/usr/share/doc/aish/copyright"
+cp "$HERE/debian/changelog" "$ROOT/usr/share/doc/aish/changelog.Debian"
+gzip -9 -n "$ROOT/usr/share/doc/aish/changelog.Debian"
+
+cat > "$ROOT/DEBIAN/control" <<EOF
+Package: aish
+Version: ${PKGVER}-${PKGREL}
+Section: shells
+Priority: optional
+Architecture: all
+Depends: luajit, libreadline8t64 | libreadline8, libcurl4t64 | libcurl4
+Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+Homepage: https://git.reauktion.de/marfrit/aish
+Description: AI-augmented conversational shell (LuaJIT, FFI-only)
+ aish is an interactive REPL that interleaves shell execution and
+ language-model conversation against llama.cpp HTTP brokers. Pure
+ LuaJIT 2.x with FFI bindings to libcurl, GNU readline, and libc.
+ .
+ Modules install under /usr/share/lua/5.1/aish/. The launcher is
+ /usr/bin/aish. Example configuration is at
+ /usr/share/doc/aish/examples/config.lua (copy to
+ ~/.config/aish/config.lua and adapt).
+EOF
+
+# Build the .deb. Output to current dir of the caller.
+DEB_OUT=aish_${PKGVER}-${PKGREL}_all.deb
+dpkg-deb --root-owner-group --build "$ROOT" "$HERE/$DEB_OUT"
+echo "built: $HERE/$DEB_OUT"

debian/aish/debian/changelog

@@ -0,0 +1,14 @@
+aish (0.1.0-1) bookworm trixie; urgency=medium
+
+  * Initial release packaged for marfrit overlay repo. Phases 0-10
+    complete (102 closed issues): local llama.cpp + cloud broker
+    routing via hossenfelder, MCP tool calls with confirm-gate and
+    per-tool auto_approve, Chuck Norris autonomous mode with
+    destructive-op heuristic, cross-session memory.jsonl, multi-model
+    routing + GBNF grammar passthrough, project file-tree context,
+    cost/usage observability, /tokenize endpoint integration, project
+    overlay (.aish.lua + sha256-pinned trust ledger), cloud preplanner
+    → local executor split.
+  * Source-of-truth: git.reauktion.de/marfrit/aish, tagged v0.1.0.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Mon, 25 May 2026 00:00:00 +0000

debian/aish/debian/control

@@ -0,0 +1,20 @@
+Source: aish
+Section: shells
+Priority: optional
+Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+Standards-Version: 4.6.2
+Homepage: https://git.reauktion.de/marfrit/aish
+
+Package: aish
+Architecture: all
+Depends: ${misc:Depends}, luajit, libreadline8t64 | libreadline8, libcurl4t64 | libcurl4
+Description: AI-augmented conversational shell (LuaJIT, FFI-only)
+ aish is an interactive REPL that interleaves shell execution and language-
+ model conversation against llama.cpp HTTP brokers. Implementation is pure
+ LuaJIT 2.x with FFI bindings to libcurl, GNU readline, and libc — no C
+ extensions, no build step.
+ .
+ Modules install under /usr/share/lua/5.1/aish/. The launcher is
+ /usr/bin/aish. Example configuration is at
+ /usr/share/doc/aish/examples/config.lua (copy to ~/.config/aish/config.lua
+ and adapt).

debian/aish/debian/copyright

@@ -0,0 +1,30 @@
+Format: https://www.debian.org/doc/packaging-manuals/copyright-format/1.0/
+Upstream-Name: aish
+Source: https://git.reauktion.de/marfrit/aish
+
+Files: *
+Copyright: 2026 Markus Fritsche <mfritsche@reauktion.de>
+License: MIT
+ Permission is hereby granted, free of charge, to any person obtaining a copy
+ of this software and associated documentation files (the "Software"), to deal
+ in the Software without restriction, including without limitation the rights
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ copies of the Software, and to permit persons to whom the Software is
+ furnished to do so, subject to the following conditions:
+ .
+ The above copyright notice and this permission notice shall be included in
+ all copies or substantial portions of the Software.
+ .
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.
+
+Files: vendor/dkjson.lua
+Copyright: 2010-2014 David Heiko Kolf
+License: MIT
+ Permission is hereby granted, free of charge, to any person obtaining a copy
+ of this software and associated documentation files (the "Software"), to deal
+ in the Software without restriction, including the rights to use, copy,
+ modify, merge, publish, distribute, sublicense, and/or sell copies of the
+ Software, and to permit persons to whom the Software is furnished to do so,
+ subject to the following conditions: the above copyright notice and this
+ permission notice shall be included in all copies or substantial portions of
+ the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.

debian/chromium-fourier/build-deb.sh

@@ -0,0 +1,150 @@
+#!/bin/bash
+# Package pre-built chromium-fourier artifacts into a .deb.
+#
+# Chromium can't be compiled natively on any available aarch64 runner
+# (clang version wall — chromium requires its internal clang fork).
+# The build is cross-compiled on CT 220 (data, x86_64 Ryzen 7).
+# This script expects the build artifacts to exist at BUILD_DIR
+# (default: fetched from CT 220 via SSH).
+#
+# Sibling Arch package: ../../arch/chromium-fourier/PKGBUILD
+set -euo pipefail
+
+PKGVER=148.0.7778.178
+EPOCH=1
+PKGREL=1
+ARCH=arm64
+
+HERE=$(dirname "$(readlink -f "$0")")
+export SOURCE_DATE_EPOCH=1779854400  # 2026-05-24 09:00 UTC
+
+BUILD_DIR="${BUILD_DIR:-}"
+
+work=$(mktemp -d)
+trap "rm -rf $work" EXIT
+
+if [ -z "$BUILD_DIR" ]; then
+    echo "BUILD_DIR not set — fetching artifacts from CT 220 on data..."
+    BUILD_DIR="$work/artifacts"
+    mkdir -p "$BUILD_DIR"
+    ssh root@data "pct exec 220 -- tar -cf - -C /build/chromium/src/out/Default \
+        chrome chrome_crashpad_handler \
+        libEGL.so libGLESv2.so libvk_swiftshader.so libvulkan.so.1 \
+        vk_swiftshader_icd.json \
+        chrome_100_percent.pak chrome_200_percent.pak resources.pak \
+        v8_context_snapshot.bin snapshot_blob.bin icudtl.dat \
+        locales/" | tar -xf - -C "$BUILD_DIR"
+fi
+
+ROOT="$work/pkgroot"
+
+install -Dm755 "$BUILD_DIR/chrome"                   "$ROOT/usr/lib/chromium/chromium"
+install -Dm755 "$BUILD_DIR/chrome_crashpad_handler"  "$ROOT/usr/lib/chromium/chrome_crashpad_handler"
+
+for so in libEGL.so libGLESv2.so libvk_swiftshader.so libvulkan.so.1; do
+    [ -f "$BUILD_DIR/$so" ] && install -Dm755 "$BUILD_DIR/$so" "$ROOT/usr/lib/chromium/$so"
+done
+
+for icd in "$BUILD_DIR"/*_icd.json; do
+    [ -f "$icd" ] && install -Dm644 "$icd" "$ROOT/usr/lib/chromium/$(basename "$icd")"
+done
+
+for f in chrome_100_percent.pak chrome_200_percent.pak resources.pak \
+         v8_context_snapshot.bin snapshot_blob.bin icudtl.dat; do
+    [ -f "$BUILD_DIR/$f" ] && install -Dm644 "$BUILD_DIR/$f" "$ROOT/usr/lib/chromium/$f"
+done
+
+if [ -d "$BUILD_DIR/locales" ]; then
+    install -dm755 "$ROOT/usr/lib/chromium/locales"
+    cp -r "$BUILD_DIR/locales/"* "$ROOT/usr/lib/chromium/locales/"
+fi
+
+install -dm755 "$ROOT/usr/bin"
+cat > "$ROOT/usr/bin/chromium-fourier" <<'LAUNCHER'
+#!/bin/bash
+USER_HANDLES_VULKAN=0
+for arg in "$@"; do
+  case "$arg" in
+    --use-vulkan*|--enable-features=*Vulkan*|--disable-features=*Vulkan*|--use-angle=vulkan*)
+      USER_HANDLES_VULKAN=1
+      break
+      ;;
+  esac
+done
+
+vulkan_default=()
+if [ "$USER_HANDLES_VULKAN" = 0 ]; then
+  vulkan_default=(--disable-features=Vulkan)
+fi
+
+exec /usr/lib/chromium/chromium \
+  --ozone-platform=wayland \
+  --use-gl=angle --use-angle=gles \
+  --enable-features=AcceleratedVideoDecoder \
+  "${vulkan_default[@]}" \
+  "$@"
+LAUNCHER
+chmod 0755 "$ROOT/usr/bin/chromium-fourier"
+
+mkdir -p "$ROOT/usr/share/doc/chromium-fourier" "$ROOT/DEBIAN"
+install -Dm644 "$HERE/debian/copyright" \
+    "$ROOT/usr/share/doc/chromium-fourier/copyright"
+install -Dm644 "$HERE/debian/changelog" \
+    "$ROOT/usr/share/doc/chromium-fourier/changelog.Debian"
+gzip -9 -n "$ROOT/usr/share/doc/chromium-fourier/changelog.Debian"
+
+ISIZE=$(du -sk "$ROOT" | awk '{print $1}')
+cat > "$ROOT/DEBIAN/control" <<EOF
+Package: chromium-fourier
+Version: ${EPOCH}:${PKGVER}-${PKGREL}
+Section: web
+Priority: optional
+Architecture: ${ARCH}
+Installed-Size: ${ISIZE}
+Depends: libasound2,
+         libatk-bridge2.0-0,
+         libatk1.0-0,
+         libcairo2,
+         libcups2,
+         libdbus-1-3,
+         libdrm2,
+         libexpat1,
+         libfontconfig1,
+         libfreetype6,
+         libgbm1,
+         libglib2.0-0,
+         libgtk-3-0,
+         libnspr4,
+         libnss3,
+         libpango-1.0-0,
+         libpulse0,
+         libva2,
+         libwayland-client0,
+         libx11-6,
+         libxcb1,
+         libxkbcommon0,
+         libpipewire-0.3-0,
+         fonts-liberation,
+         v4l-utils
+Provides: www-browser
+Conflicts: chromium
+Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+Homepage: https://www.chromium.org/
+Description: Chromium with V4L2 HW video decode for Rockchip (Wayland + mainline)
+ Chromium ${PKGVER} with three patches enabling V4L2 hardware video
+ decoding on mainline Linux / Wayland for Rockchip SoCs (RK3566 hantro,
+ RK3588 VDPU381).
+ .
+ Cross-compiled from x86_64 using chromium's bundled clang (upstream
+ LLVM cannot compile chromium). Runtime target is aarch64.
+ .
+ Patches: enable-v4l2-decoder-default, wayland-allow-direct-egl-gles2,
+ nv12-external-oes-on-modifier-external-only.
+ .
+ Launcher at /usr/bin/chromium-fourier defaults to Wayland + ANGLE/GLES
+ with Vulkan disabled (panvk on RK3566 breaks V4L2 dispatch).
+EOF
+
+DEB_OUT="chromium-fourier_${EPOCH}%3a${PKGVER}-${PKGREL}_${ARCH}.deb"
+dpkg-deb --root-owner-group --build "$ROOT" "$HERE/$DEB_OUT"
+echo "built: $HERE/$DEB_OUT"

debian/chromium-fourier/debian/changelog

@@ -0,0 +1,8 @@
+chromium-fourier (1:148.0.7778.178-1) trixie; urgency=medium
+
+  * Chromium 148.0.7778.178 with V4L2 HW decode patches for Rockchip.
+  * Cross-compiled from x86_64 using chromium's bundled clang.
+  * Three fourier patches: enable-v4l2-decoder-default,
+    wayland-allow-direct-egl-gles2, nv12-external-oes-on-modifier-external-only.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Sat, 24 May 2026 09:00:00 +0200

debian/chromium-fourier/debian/copyright

@@ -0,0 +1,32 @@
+Format: https://www.debian.org/doc/packaging-manuals/copyright-format/1.0/
+Upstream-Name: Chromium
+Upstream-Contact: chromium-dev@chromium.org
+Source: https://www.chromium.org/
+
+Files: *
+Copyright: The Chromium Authors
+License: BSD-3-Clause
+
+Files: debian/*
+Copyright: 2026 Markus Fritsche <mfritsche@reauktion.de>
+License: BSD-3-Clause
+
+License: BSD-3-Clause
+ Redistribution and use in source and binary forms, with or without
+ modification, are permitted provided that the following conditions are met:
+ .
+ 1. Redistributions of source code must retain the above copyright notice,
+    this list of conditions and the following disclaimer.
+ .
+ 2. Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in the
+    documentation and/or other materials provided with the distribution.
+ .
+ 3. Neither the name of the copyright holder nor the names of its
+    contributors may be used to endorse or promote products derived from
+    this software without specific prior written permission.
+ .
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ ARE DISCLAIMED.

debian/claude-his-agent/build-deb.sh

@@ -16,7 +16,7 @@ work=$(mktemp -d)
 trap "rm -rf $work" EXIT
 
 cd "$work"
-curl -sSLfo his.tar.gz \
+curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo his.tar.gz \
     "https://git.reauktion.de/marfrit/claude-his-agent/archive/v${PKGVER}.tar.gz"
 echo "$HIS_TARBALL_SHA256  his.tar.gz" | sha256sum -c
 tar xzf his.tar.gz

debian/daedalus-v4l2-dkms/build-deb.sh

@@ -0,0 +1,200 @@
+#!/bin/bash
+# Build daedalus-v4l2-dkms_<ver>_all.deb (kernel module via DKMS).
+#
+# Installs kernel/ source tree to /usr/src/daedalus_v4l2-${PKGVER}/
+# plus a dkms.conf.  Postinst registers with DKMS (dkms add + build +
+# install).  Prerm deregisters.  Result: the daedalus_v4l2 module
+# auto-rebuilds against any installed kernel headers without users
+# needing to remember to dkms-add it.
+#
+# Architecture: all.  The kernel module itself is per-kernel-version,
+# but the SOURCE package is arch-independent.
+#
+# Sibling Arch package: ../../arch/daedalus-v4l2-dkms/PKGBUILD
+# Sibling userspace package: ../daedalus-v4l2/build-deb.sh
+set -euo pipefail
+
+UPSTREAM_COMMIT=872eec505eb91b561892d02a0526749348ddc121
+PKGVER=0.1.0+r45+g872eec5
+PKGREL=1  # reset for new upstream pin (872eec5 — PROTO_MAX_PAYLOAD 64 KiB -> 1 MiB, closes #19); lock-step with daedalus-v4l2 0.1.0+r45+g872eec5 REQUIRED
+MODULE_NAME=daedalus_v4l2
+
+HERE=$(dirname "$(readlink -f "$0")")
+
+# Reproducible build.  2026-05-18 23:00 UTC — Phase 8.13 close.
+export SOURCE_DATE_EPOCH=1779231600
+
+work=$(mktemp -d)
+trap "rm -rf $work" EXIT
+
+cd "$work"
+curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo daedalus-v4l2.tar.gz \
+    "https://git.reauktion.de/reauktion/daedalus-v4l2/archive/${UPSTREAM_COMMIT}.tar.gz"
+tar xzf daedalus-v4l2.tar.gz
+SRCDIR=daedalus-v4l2
+
+ROOT="$work/pkgroot"
+SRCROOT="$ROOT/usr/src/${MODULE_NAME}-${PKGVER}"
+
+mkdir -p "$SRCROOT/include" \
+         "$ROOT/DEBIAN" \
+         "$ROOT/usr/share/doc/daedalus-v4l2-dkms"
+
+# Copy kernel/ source files to the DKMS source dir.
+cp -r "$work/$SRCDIR/kernel/." "$SRCROOT/"
+
+# Embed the shared protocol header inline (rather than referencing
+# ../include/ which doesn't exist after DKMS extracts the tree).
+# Patch the Makefile to find it at $SRCROOT/include/ instead.
+install -m 644 "$work/$SRCDIR/include/daedalus_v4l2_proto.h" \
+    "$SRCROOT/include/daedalus_v4l2_proto.h"
+sed -i 's|-I\$(src)/\.\./include|-I$(src)/include|' "$SRCROOT/Makefile"
+
+# Generate dkms.conf with the actual version substituted.
+cat > "$SRCROOT/dkms.conf" <<EOF
+PACKAGE_NAME="${MODULE_NAME}"
+PACKAGE_VERSION="${PKGVER}"
+
+BUILT_MODULE_NAME[0]="${MODULE_NAME}"
+DEST_MODULE_LOCATION[0]="/updates"
+
+MAKE[0]="make KERNELDIR=/lib/modules/\${kernelver}/build all"
+CLEAN="make KERNELDIR=/lib/modules/\${kernelver}/build clean"
+
+AUTOINSTALL="yes"
+EOF
+
+# Doc
+install -m 644 "$work/$SRCDIR/README.md" \
+    "$ROOT/usr/share/doc/daedalus-v4l2-dkms/README.md"
+install -Dm644 "$HERE/debian/copyright" "$ROOT/usr/share/doc/daedalus-v4l2-dkms/copyright"
+install -Dm644 "$HERE/debian/changelog" "$ROOT/usr/share/doc/daedalus-v4l2-dkms/changelog.Debian"
+gzip -9 -n "$ROOT/usr/share/doc/daedalus-v4l2-dkms/changelog.Debian"
+
+# DKMS post-install / pre-remove hooks.
+cat > "$ROOT/DEBIAN/postinst" <<EOF
+#!/bin/sh
+set -e
+
+NAME=${MODULE_NAME}
+VERSION=${PKGVER}
+
+# Yellow + bold ANSI for the warning so it stands out in apt's
+# stream of "Setting up" lines.  Disable colour on non-TTY.
+if [ -t 1 ]; then
+    Y=\$(printf '\\033[1;33m'); R=\$(printf '\\033[0m')
+else
+    Y=''; R=''
+fi
+
+warn() {
+    printf '%s==> daedalus-v4l2-dkms: %s%s\\n' "\$Y" "\$1" "\$R" >&2
+}
+
+if [ "\$1" = "configure" ]; then
+    if ! command -v dkms >/dev/null 2>&1; then
+        warn "dkms not installed; module \$NAME/\$VERSION not registered."
+        warn "Install 'dkms' then run: dkms add \$NAME/\$VERSION && dkms autoinstall"
+        exit 0
+    fi
+
+    dkms add "\$NAME/\$VERSION" 2>/dev/null || true
+
+    # Enumerate every kernel whose headers are actually present
+    # (/lib/modules/<kver>/build resolves to a directory).  We iterate
+    # all of them — not just \$(uname -r) — so that installing this
+    # package after a kernel update covers the newly-installed kernel
+    # too, and so that a later kernel-headers install for a previously
+    # uncovered version gets picked up on dpkg-reconfigure.  Without
+    # this, autoinstall (which targets only the running kernel) leaves
+    # /dev/daedalus-v4l2 absent after a kernel switch + reboot
+    # (marfrit/marfrit-packages#64).
+    kvers=''
+    for d in /lib/modules/*/build; do
+        [ -d "\$d" ] || continue
+        k=\$(basename "\$(dirname "\$d")")
+        kvers="\$kvers \$k"
+    done
+
+    if [ -z "\$kvers" ]; then
+        warn ""
+        warn "No kernels with headers found under /lib/modules/*/build."
+        warn "Install kernel headers (e.g. linux-headers-rpi-2712 on Pi OS)"
+        warn "then finish with:"
+        warn "  sudo dkms autoinstall \$NAME/\$VERSION"
+        exit 0
+    fi
+
+    failed=''
+    for k in \$kvers; do
+        dkms autoinstall -k "\$k" "\$NAME/\$VERSION" >/dev/null 2>&1 || true
+        s=\$(dkms status -m "\$NAME" -v "\$VERSION" -k "\$k" 2>/dev/null || true)
+        if ! printf '%s\\n' "\$s" | grep -q -E 'installed|loaded'; then
+            failed="\$failed \$k"
+        fi
+    done
+
+    if [ -n "\$failed" ]; then
+        warn ""
+        warn "DKMS build did NOT land for kernel(s):\$failed"
+        warn ""
+        warn "Most likely cause: kernel headers missing for those versions."
+        warn "  Raspberry Pi OS / Pi 5:  apt install linux-headers-rpi-2712"
+        warn "  Debian generic:          apt install linux-headers-<version>"
+        warn ""
+        warn "After installing headers, finish with:"
+        for k in \$failed; do
+            warn "  sudo dkms autoinstall -k \$k \$NAME/\$VERSION"
+        done
+        warn "  sudo modprobe daedalus_v4l2  (after booting that kernel)"
+        warn ""
+        warn "Until then daedalus_v4l2 will NOT be loadable on those kernels"
+        warn "and the userspace daedalus-v4l2 daemon will have nothing to talk to."
+    fi
+fi
+
+#DEBHELPER#
+EOF
+chmod 755 "$ROOT/DEBIAN/postinst"
+
+cat > "$ROOT/DEBIAN/prerm" <<EOF
+#!/bin/sh
+set -e
+
+NAME=${MODULE_NAME}
+VERSION=${PKGVER}
+
+if [ "\$1" = "remove" ] && command -v dkms >/dev/null 2>&1; then
+    dkms remove "\$NAME/\$VERSION" --all || true
+fi
+
+#DEBHELPER#
+EOF
+chmod 755 "$ROOT/DEBIAN/prerm"
+
+cat > "$ROOT/DEBIAN/control" <<EOF
+Package: daedalus-v4l2-dkms
+Version: ${PKGVER}-${PKGREL}
+Section: kernel
+Priority: optional
+Architecture: all
+Depends: dkms (>= 2.1.0.0)
+Recommends: daedalus-v4l2, linux-headers-rpi-2712 | linux-headers-rpi | linux-headers-generic | linux-headers
+Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+Homepage: https://git.reauktion.de/reauktion/daedalus-v4l2
+Description: V4L2 stateless decoder shim kernel module (DKMS) — Pi 5 / CM5
+ Out-of-tree V4L2 m2m kernel module for the daedalus-v4l2 stack on
+ Raspberry Pi 5 / CM5.  Registers /dev/videoNN (V4L2 stateless m2m
+ decoder), /dev/mediaNN (media controller with request API), and
+ /dev/daedalus-v4l2 (chardev bridge to the userspace daemon).
+ .
+ The actual decode happens in the userspace daemon shipped by the
+ daedalus-v4l2 package — this module is just the kernel-side V4L2
+ plumbing.  Install both to actually serve VAAPI / V4L2 clients.
+ .
+ Built via DKMS against the running kernel's headers.
+EOF
+
+DEB_OUT="daedalus-v4l2-dkms_${PKGVER}-${PKGREL}_all.deb"
+dpkg-deb --root-owner-group --build "$ROOT" "$HERE/$DEB_OUT"
+echo "built: $HERE/$DEB_OUT"

debian/daedalus-v4l2-dkms/debian/changelog

@@ -0,0 +1,167 @@
+daedalus-v4l2-dkms (0.1.0+r45+g872eec5-1) bookworm trixie; urgency=medium
+
+  * Bump to 872eec5 — picks up daedalus-v4l2 PR #20 (closes #19).
+    Wire-protocol cap DAEDALUS_PROTO_MAX_PAYLOAD raised from 64 KiB
+    to 1 MiB in include/daedalus_v4l2_proto.h.  The kernel module
+    inherits the larger DAEDALUS_MAX_BITSTREAM via the same #define
+    and daedalus_fill_output_fmt now reports OUTPUT_MPLANE
+    sizeimage = ~1 MiB instead of 65484.
+  * Skips the r33 -> r45 commit range — between 5d8b436 and 872eec5
+    only one kernel/include change landed (the PROTO_MAX_PAYLOAD
+    bump above).  The intervening daemon-only bumps (r37 / r39 /
+    r41 / r43) didn't touch kernel/ or include/ at all.
+  * Effective wire cap is min(kernel, daemon) — lock-step install
+    WITH daedalus-v4l2 0.1.0+r45+g872eec5 REQUIRED.
+  * Allocations (kmemdup / kmalloc on payload, vb2 plane backing)
+    are dynamic and sized per-payload at runtime; the bump only
+    sets the ceiling.  KMALLOC_MAX_SIZE on aarch64 SLUB is several
+    MiB so 1 MiB is well within bounds.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Fri, 22 May 2026 21:00:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r33+g5d8b436-1) bookworm trixie; urgency=medium
+
+  * Bump to 5d8b436 — reverts daedalus-v4l2 PRs #7 + #8.  Kernel
+    module returns to the pre-#7 buf_done_and_job_finish completion
+    model: no src/dst lifecycle decoupling, no parked dst_bufs, no
+    1:1-contract violation against libva-v4l2-request-fourier
+    (closes daedalus-v4l2#9 + #10 as won't-fix at this layer; proper
+    fix tracked at daedalus-v4l2#11).
+  * Wire-protocol drops 1 → 0; lock-step install with daedalus-v4l2
+    0.1.0+r33+g5d8b436 REQUIRED.
+  * Carries forward the #64 multi-kernel postinst fix.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 14:50:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r30+g6ffe92b-1) bookworm trixie; urgency=medium
+
+  * Bump to 6ffe92b — fixes the kernel panic regression introduced
+    by 79256dc's split-completion design (closes daedalus-v4l2#8).
+    `device_run` now removes both src + dst from `m2m_ctx`'s
+    rdy_queue at pickup time, not at `buf_done` time.  Without
+    this, after `SRC_CONSUMED`'s `job_finish` released the m2m
+    scheduler, the NEXT `device_run` saw the still-queued parked
+    dst_buf and paired it with a fresh src — two inflight entries
+    referencing the same vb2_buffer, the later `HAS_PIXELS`
+    triggered list_del on an already-detached list_head, smashing
+    the rdy_queue → hard reboot on Pi CM5 during `mpv vaapi-copy`
+    playback of 720p H.264 (2026-05-21).
+  * Wire protocol unchanged — DAEDALUS_PROTO_VERSION stays at 1.
+    Daemon (userspace daedalus-v4l2 package) need NOT bump in
+    lockstep with this DKMS update; the existing
+    daedalus-v4l2 0.1.0+r28+g79256dc is wire-compatible with
+    daedalus-v4l2-dkms 0.1.0+r30+g6ffe92b.
+  * Carries forward the #64 multi-kernel postinst fix.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 14:00:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r28+g79256dc-1) bookworm trixie; urgency=medium
+
+  * Bump to 79256dc — H.264 B-frame display reorder fix (closes
+    daedalus-v4l2#6).  libavcodec's H.264 decoder reorders output to
+    display order before returning from avcodec_receive_frame; the
+    daemon was binding each REQ_DECODE's pixels to the cookie of the
+    bitstream that triggered the receive_frame call, not the cookie
+    of the bitstream that actually produced the picture.  For B-frame
+    sequences this paired cookie N's CAPTURE buffer with cookie N-2's
+    pixels and silently lost intermediate frames — visible as
+    "2 1 4 3 6 5" frame pairing in mpv / Firefox on Pi CM5.
+  * Wire-protocol bump (DAEDALUS_PROTO_VERSION 0 → 1): REQ_DECODE
+    gains __u64 src_pts; RESP_FRAME gains __u32 flags +
+    __u64 output_src_pts.  Kernel + daemon must install atomically
+    (this package + daedalus-v4l2 0.1.0+r28+g79256dc).
+  * Carries forward the #64 multi-kernel postinst fix from -2:
+    autoinstall for every /lib/modules/*/build that resolves to real
+    headers, not just $(uname -r).
+  * Closes #64.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 12:00:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r24+gf0d4186-2) bookworm trixie; urgency=medium
+
+  * postinst: autoinstall for every installed kernel with headers, not
+    just the running one.  Previously `dkms autoinstall $NAME/$VERSION`
+    built only against `$(uname -r)`, so installing the package on
+    kernel A and then rebooting into a separately-installed kernel B
+    left /lib/modules/B/updates/dkms/ empty — /dev/daedalus-v4l2 absent,
+    daedalus daemon nothing to talk to, browser/VAAPI silently falling
+    back to software with no obvious diagnostic.  Now we enumerate every
+    /lib/modules/*/build that resolves to a real directory and run
+    `dkms autoinstall -k <kver>` for each, reporting per-kernel failure
+    only when headers are missing.  Closes #64.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 09:30:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r24+gf0d4186-1) bookworm trixie; urgency=medium
+
+  * Bump to f0d4186 — per-ctx vb2 lock fix.  daedalus_queue_init now
+    uses ctx->vb_mutex instead of ctx->dev->m2m_lock for each
+    vb2_queue's lock, unblocking Firefox's multi-process VAAPI
+    clients (they were colliding on the device-wide mutex and one
+    would EBUSY-fail S_FMT while another was mid-streamon).
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Wed, 20 May 2026 23:00:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r22+g462aa4b-1) bookworm trixie; urgency=medium
+
+  * Bump to 462aa4b — kernel device_run() now calls
+    v4l2_ctrl_request_setup() before reading the H.264 stateless
+    control values from the bound media_request, so the values
+    daedalus ships to the userspace daemon match what the V4L2
+    client (libva-v4l2-request-fourier) actually set.  Closes the
+    libva→kernel control-binding gap that was causing decoded
+    frames to come back as best-effort zero garbage from libavcodec.
+  * Wire-ABI lockstep with daedalus-v4l2 0.1.0+r22+g462aa4b.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Wed, 20 May 2026 22:00:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r20+g3dd0eb0-1) bookworm trixie; urgency=medium
+
+  * Bump to 3dd0eb0 — DAEMON-PPS kernel-side changes.  device_run()
+    now reads the V4L2 H.264 stateless control values from the bound
+    media_request and ships them to the daemon inside REQ_DECODE
+    via the new struct daedalus_h264_meta block (gated on
+    DAEDALUS_REQ_FLAG_H264_META).  Required for H.264 decode to
+    work via the libva-v4l2-request -> daedalus daemon path; daemon
+    synthesises AnnexB SPS+PPS NAL units from the structs.
+  * Wire-ABI lockstep with daedalus-v4l2 0.1.0+r20+g3dd0eb0 — install
+    both packages together.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Wed, 20 May 2026 21:00:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r18+g481279c-1) bookworm trixie; urgency=medium
+
+  * Bump to 481279c in lockstep with the userspace daedalus-v4l2
+    package (which ships the new systemd unit + modules-load.d
+    drop-in).  No kernel-module behaviour change in this commit —
+    bump is purely to keep DKMS pkgver and userspace pkgver aligned
+    so /etc/modules-load.d/daedalus-v4l2.conf has a module to load.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Wed, 20 May 2026 16:00:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r17+gf0cd29a-1) bookworm trixie; urgency=medium
+
+  * Bump to f0cd29a — Linux 6.18 compatibility fix.  v4l2_fh_add /
+    v4l2_fh_del gained a `struct file *filp` second argument at
+    v6.18; module source now uses LINUX_VERSION_CODE conditional so
+    it builds against 6.12 (Pi 5 stock RPi-2712 kernel) AND 6.18+
+    (RPi-2712 / arm64).  Verified DKMS rebuild on higgs
+    (6.18.29+rpt-rpi-2712) succeeds + /dev/daedalus-v4l2 appears.
+  * Recommends widened so apt pulls the right kernel-headers
+    metapackage on RPi OS: linux-headers-rpi-2712 first, then
+    linux-headers-rpi, then linux-headers-generic, then
+    linux-headers.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Tue, 20 May 2026 10:30:00 +0000
+
+daedalus-v4l2-dkms (0.1.0+r16+gf55b2cd-1) bookworm trixie; urgency=medium
+
+  * Initial Debian DKMS packaging for the daedalus_v4l2 kernel module.
+  * Pinned to f55b2cd (Phase 8.13 close): kernel-side framework
+    integration (V4L2 m2m, dmabuf-export, media controller, request
+    API, NV12 single-plane + NV12M + P010 CAPTURE) that closes the
+    libva→/dev/video0→daemon round-trip with byte-exact pixels.
+  * Auto-builds via DKMS against the running kernel's headers.
+  * Companion userspace package: daedalus-v4l2 (daemon + tools).
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Mon, 18 May 2026 23:00:00 +0000

debian/daedalus-v4l2-dkms/debian/control

@@ -0,0 +1,24 @@
+Source: daedalus-v4l2-dkms
+Section: kernel
+Priority: optional
+Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+Build-Depends: debhelper-compat (= 13)
+Standards-Version: 4.6.2
+Homepage: https://git.reauktion.de/reauktion/daedalus-v4l2
+
+Package: daedalus-v4l2-dkms
+Architecture: all
+Depends: ${misc:Depends}, dkms (>= 2.1.0.0)
+Recommends: daedalus-v4l2,
+            linux-headers-rpi-2712 | linux-headers-rpi | linux-headers-generic | linux-headers
+Description: V4L2 stateless decoder shim kernel module (DKMS) — Pi 5 / CM5
+ Out-of-tree V4L2 m2m kernel module for the daedalus-v4l2 stack on
+ Raspberry Pi 5 / CM5. Registers /dev/videoNN (V4L2 stateless m2m
+ decoder), /dev/mediaNN (media controller with request API), and
+ /dev/daedalus-v4l2 (chardev bridge to the userspace daemon).
+ .
+ The actual decode happens in the userspace daemon shipped by the
+ daedalus-v4l2 package — this module is just the kernel-side V4L2
+ plumbing. Install both to actually serve VAAPI / V4L2 clients.
+ .
+ Built via DKMS against the running kernel's headers.

debian/daedalus-v4l2-dkms/debian/copyright

@@ -0,0 +1,21 @@
+Format: https://www.debian.org/doc/packaging-manuals/copyright-format/1.0/
+Upstream-Name: daedalus-v4l2
+Upstream-Contact: Markus Fritsche <fritsche.markus@gmail.com>
+Source: https://git.reauktion.de/reauktion/daedalus-v4l2
+
+Files: *
+Copyright: 2026 Markus Fritsche <fritsche.markus@gmail.com>
+License: GPL-2.0-or-later
+Comment:
+ Kernel module (loadable into the Linux kernel) — GPL-2.0-or-later.
+ The shared protocol header carries an additional Linux-syscall-note
+ exception so userspace inclusion is BSD-clean.
+
+License: GPL-2.0-or-later
+ This program is free software; you can redistribute it and/or modify
+ it under the terms of the GNU General Public License as published by
+ the Free Software Foundation; either version 2 of the License, or
+ (at your option) any later version.
+ .
+ On Debian systems, the complete text of the GNU General Public
+ License v2 can be found in `/usr/share/common-licenses/GPL-2'.

debian/daedalus-v4l2/build-deb.sh

@@ -0,0 +1,214 @@
+#!/bin/bash
+# Build daedalus-v4l2_<ver>_arm64.deb (userspace daemon + test tools).
+#
+# Mirrors arch/daedalus-v4l2 (Arch Linux build).  The companion DKMS
+# package (debian/daedalus-v4l2-dkms) carries the kernel module
+# separately so apt/dpkg can split kernel-version-tied and userspace
+# upgrade cadence.
+#
+# Sibling Arch package: ../../arch/daedalus-v4l2/PKGBUILD
+# Sibling DKMS package: ../daedalus-v4l2-dkms/build-deb.sh
+# Upstream repo: https://git.reauktion.de/reauktion/daedalus-v4l2
+set -euo pipefail
+
+# 6e6dfa1 = picks up daedalus-v4l2 PR #16 — daemon now dlopens
+# the Kwiboo fourier fork's libavcodec.so.62 / libavformat.so.62 /
+# libavutil.so.60 at /opt/fourier instead of Debian-stock soname
+# 61/61/59.  First step on the daedalus-fourier substitution arc
+# (daedalus-v4l2#11): routes the daemon through the libavcodec
+# source tree we own in marfrit-packages.  Headers + .pc files
+# come from ffmpeg-v4l2-request-fourier (installed by the CI
+# workflow before this script runs; see PKG_CONFIG_PATH below).
+UPSTREAM_COMMIT=872eec505eb91b561892d02a0526749348ddc121
+PKGVER=0.1.0+r45+g872eec5
+PKGREL=1  # reset for new upstream pin (872eec5 — PROTO_MAX_PAYLOAD 64 KiB -> 1 MiB, closes #19); lock-step with daedalus-v4l2-dkms 0.1.0+r45+g872eec5 REQUIRED
+
+# daedalus-fourier pin.  d87239d = marfrit/daedalus-fourier PR #1 merge
+# (install rules + pkg-config, enables this consumer to find_package
+# + link).  Bump in lockstep with the upstream daemon when daedalus-
+# fourier's API or installed shaders are changed by a new consumer.
+DAEDALUS_FOURIER_COMMIT=d87239d8172307d9a1b93c95cbed116d175b85cc
+
+HERE=$(dirname "$(readlink -f "$0")")
+
+# Reproducible build.  2026-05-18 23:00 UTC — Phase 8.13 close.
+export SOURCE_DATE_EPOCH=1779231600
+
+work=$(mktemp -d)
+trap "rm -rf $work" EXIT
+
+# --- daedalus-fourier: fetch + build + install to per-build prefix ---
+#
+# Static-linked into the daemon, so the temp prefix is only for the
+# duration of this build script.  Requires libvulkan-dev + glslang-tools
+# on the runner (already needed for the daedalus-fourier benches).
+FOURIER_PREFIX=$work/fourier-prefix
+mkdir -p "$FOURIER_PREFIX"
+
+cd "$work"
+curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo daedalus-fourier.tar.gz \
+    "https://git.reauktion.de/marfrit/daedalus-fourier/archive/${DAEDALUS_FOURIER_COMMIT}.tar.gz"
+tar xzf daedalus-fourier.tar.gz
+cd daedalus-fourier
+cmake -B build -G Ninja \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_INSTALL_PREFIX="$FOURIER_PREFIX"
+cmake --build build --target daedalus_core
+cmake --install build
+
+# --- daedalus-v4l2: fetch + build daemon against installed daedalus-fourier ---
+
+cd "$work"
+curl --connect-timeout 10 --max-time 600 --retry 3 --retry-delay 5 -sSLfo daedalus-v4l2.tar.gz \
+    "https://git.reauktion.de/reauktion/daedalus-v4l2/archive/${UPSTREAM_COMMIT}.tar.gz"
+tar xzf daedalus-v4l2.tar.gz
+SRCDIR=daedalus-v4l2
+
+# Build daemon (CMake) — point pkg-config at the daedalus-fourier
+# temp prefix so pkg_check_modules(DAEDALUS_FOURIER …) resolves to it.
+cd "$SRCDIR/daemon"
+PKG_CONFIG_PATH="$FOURIER_PREFIX/lib/pkgconfig:/opt/fourier/lib/pkgconfig" \
+cmake -B build -G Ninja \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_INSTALL_PREFIX=/usr
+cmake --build build
+
+# Build test tools (in-tree Makefile)
+cd "$work/$SRCDIR/tools"
+make
+
+# Stage
+ROOT="$work/pkgroot"
+mkdir -p "$ROOT/DEBIAN" \
+         "$ROOT/usr/bin" \
+         "$ROOT/usr/libexec/daedalus-v4l2" \
+         "$ROOT/usr/include" \
+         "$ROOT/usr/share/doc/daedalus-v4l2" \
+         "$ROOT/lib/systemd/system" \
+         "$ROOT/usr/lib/modules-load.d"
+
+install -m 755 "$work/$SRCDIR/daemon/build/daedalus_v4l2_daemon" \
+    "$ROOT/usr/bin/daedalus_v4l2_daemon"
+
+install -m 755 "$work/$SRCDIR/tools/test_chardev_pingpong" \
+    "$ROOT/usr/libexec/daedalus-v4l2/test_chardev_pingpong"
+install -m 755 "$work/$SRCDIR/tools/test_m2m_decode" \
+    "$ROOT/usr/libexec/daedalus-v4l2/test_m2m_decode"
+install -m 755 "$work/$SRCDIR/tools/test_m2m_stream" \
+    "$ROOT/usr/libexec/daedalus-v4l2/test_m2m_stream"
+
+install -m 644 "$work/$SRCDIR/include/daedalus_v4l2_proto.h" \
+    "$ROOT/usr/include/daedalus_v4l2_proto.h"
+
+# systemd unit + module autoload — without these the daemon never
+# starts and the libva/VAAPI consumer's REQ_DECODE has nobody on
+# the other end of /dev/daedalus-v4l2.
+install -m 644 "$work/$SRCDIR/packaging/systemd/daedalus-v4l2.service" \
+    "$ROOT/lib/systemd/system/daedalus-v4l2.service"
+install -m 644 "$work/$SRCDIR/packaging/systemd/daedalus-v4l2.modules-load" \
+    "$ROOT/usr/lib/modules-load.d/daedalus-v4l2.conf"
+
+install -m 644 "$work/$SRCDIR/README.md" \
+    "$ROOT/usr/share/doc/daedalus-v4l2/README.md"
+for d in "$work/$SRCDIR/docs/"*.md; do
+    install -m 644 "$d" "$ROOT/usr/share/doc/daedalus-v4l2/$(basename "$d")"
+done
+
+install -Dm644 "$HERE/debian/copyright" "$ROOT/usr/share/doc/daedalus-v4l2/copyright"
+install -Dm644 "$HERE/debian/changelog" "$ROOT/usr/share/doc/daedalus-v4l2/changelog.Debian"
+gzip -9 -n "$ROOT/usr/share/doc/daedalus-v4l2/changelog.Debian"
+
+# DEBIAN/postinst — enable service + reload modules-load.d so the
+# kernel module loads now if daedalus-v4l2-dkms is also installed.
+# Does NOT auto-start the service — that requires /dev/daedalus-v4l2
+# to already exist (the ConditionPathExists= in the .service file)
+# which may not be true on the very first install before the user
+# reboots or manually modprobes.  Operator decides when to start.
+cat > "$ROOT/DEBIAN/postinst" <<'POSTINST'
+#!/bin/sh
+set -e
+
+if [ "$1" = "configure" ]; then
+    # Reload systemd so the new unit file is visible.
+    if command -v systemctl >/dev/null 2>&1; then
+        systemctl daemon-reload >/dev/null 2>&1 || true
+        systemctl enable daedalus-v4l2.service >/dev/null 2>&1 || true
+    fi
+
+    # Trigger /usr/lib/modules-load.d/daedalus-v4l2.conf without a
+    # reboot.  Harmless if the module is already loaded; logs to
+    # journal if it can't load (most common cause: dkms hasn't built
+    # the module yet for the running kernel — see daedalus-v4l2-dkms
+    # postinst for the loud-warning path).
+    if command -v systemd-modules-load >/dev/null 2>&1; then
+        systemd-modules-load >/dev/null 2>&1 || true
+    fi
+
+    # Auto-start if /dev/daedalus-v4l2 came up (i.e. module loaded
+    # successfully).  ConditionPathExists in the unit file means
+    # `systemctl start` is a no-op if the device isn't there yet —
+    # avoids spurious failures during apt install on a host where
+    # daedalus-v4l2-dkms hasn't built yet.
+    if [ -e /dev/daedalus-v4l2 ] && command -v systemctl >/dev/null 2>&1; then
+        systemctl start daedalus-v4l2.service >/dev/null 2>&1 || true
+    fi
+fi
+
+#DEBHELPER#
+POSTINST
+chmod 755 "$ROOT/DEBIAN/postinst"
+
+cat > "$ROOT/DEBIAN/prerm" <<'PRERM'
+#!/bin/sh
+set -e
+
+if [ "$1" = "remove" ] && command -v systemctl >/dev/null 2>&1; then
+    systemctl stop    daedalus-v4l2.service >/dev/null 2>&1 || true
+    systemctl disable daedalus-v4l2.service >/dev/null 2>&1 || true
+fi
+
+#DEBHELPER#
+PRERM
+chmod 755 "$ROOT/DEBIAN/prerm"
+
+cat > "$ROOT/DEBIAN/postrm" <<'POSTRM'
+#!/bin/sh
+set -e
+
+if command -v systemctl >/dev/null 2>&1; then
+    systemctl daemon-reload >/dev/null 2>&1 || true
+fi
+
+#DEBHELPER#
+POSTRM
+chmod 755 "$ROOT/DEBIAN/postrm"
+
+cat > "$ROOT/DEBIAN/control" <<EOF
+Package: daedalus-v4l2
+Version: ${PKGVER}-${PKGREL}
+Section: video
+Priority: optional
+Architecture: arm64
+Depends: ffmpeg-v4l2-request-fourier (>= 2:8.1+rfourier), libdrm2
+Recommends: daedalus-v4l2-dkms
+Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+Homepage: https://git.reauktion.de/reauktion/daedalus-v4l2
+Description: Userspace daemon for the daedalus_v4l2 stateless decoder shim
+ daedalus-v4l2 ships the userspace daemon that backs the daedalus_v4l2
+ out-of-tree V4L2 kernel module on Raspberry Pi 5 / CM5.  Together they
+ expose /dev/videoNN + /dev/mediaNN as a V4L2 stateless decoder for
+ VP9, AV1, and H.264 — actual decoding happens in this single-threaded
+ daemon via dlopen'd FFmpeg, with decoded NV12 / P010 frames shipped
+ back through dmabuf.
+ .
+ Consumed end-to-end by libva-v4l2-request-fourier (>= 1.0.0+r376) so
+ that 'ffmpeg -hwaccel vaapi' against vp9_small.ivf produces a
+ byte-exact NV12 frame.
+ .
+ The kernel module ships separately in daedalus-v4l2-dkms; install
+ both to actually serve V4L2 clients.
+EOF
+
+DEB_OUT="daedalus-v4l2_${PKGVER}-${PKGREL}_arm64.deb"
+dpkg-deb --root-owner-group --build "$ROOT" "$HERE/$DEB_OUT"
+echo "built: $HERE/$DEB_OUT"

debian/daedalus-v4l2/debian/changelog

@@ -0,0 +1,250 @@
+daedalus-v4l2 (0.1.0+r45+g872eec5-1) bookworm trixie; urgency=medium
+
+  * Bump to 872eec5 — picks up daedalus-v4l2 PR #20 (closes #19).
+    Wire-protocol cap DAEDALUS_PROTO_MAX_PAYLOAD raised from 64 KiB
+    to 1 MiB.  DAEDALUS_MAX_BITSTREAM follows; daedalus_fill_output_fmt
+    now reports OUTPUT_MPLANE sizeimage = ~1 MiB instead of 65484.
+    libva-v4l2-request-fourier's S_FMT-driven OUTPUT-pool resize
+    finally succeeds; Firefox no longer falls off to libmozavcodec
+    SW when an H.264 slice exceeds 64 KiB (routine on any
+    720p+ stream).
+  * #define-only change in include/daedalus_v4l2_proto.h; struct
+    layout unchanged.  But effective cap is min(kernel, daemon) —
+    lock-step install of this package WITH
+    daedalus-v4l2-dkms 0.1.0+r45+g872eec5 REQUIRED.
+  * Daemon-side allocations are dynamic (malloc-on-payload), so
+    the practical growth is one ~1 MiB read buffer per daemon
+    process at startup.  Negligible on Pi 5 / 8 GB.
+  * Picks up the same r43 -> r45 transition as daedalus-v4l2-dkms
+    (which had been stuck at r33+g5d8b436 since the parking-design
+    revert because the kernel module didn't change in r37/r39/r41/r43).
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Fri, 22 May 2026 21:00:00 +0000
+
+daedalus-v4l2 (0.1.0+r43+g1d8f5af-1) bookworm trixie; urgency=medium
+
+  * Bump to 1d8f5af — picks up daedalus-v4l2 PR #18 (closes #17).
+    Daemon now drops degenerate (<4 byte) bitstreams at the REQ_DECODE
+    entry instead of letting avcodec_send_packet return
+    AVERROR_INVALIDDATA.  Reply RESP_FRAME with status=
+    DAEDALUS_DECODE_NO_FRAME so libva's V4L2 surface pool stays
+    healthy.
+  * Fixes the Firefox YouTube avc1 pause→resume regression observed
+    on higgs: libva-v4l2-request-fourier flushes a 3-byte stub
+    (presumably a bare NAL start code) into OUTPUT_MPLANE at the
+    pause boundary; the old INVALIDDATA error path made Firefox
+    fall off to libmozavcodec SW for the rest of the session.  With
+    this filter the daemon logs the sentinel as 'tiny bitstream 3
+    bytes — dropping as no-op' and the next real REQ_DECODE
+    proceeds normally.
+  * Wire protocol unchanged.  No daedalus-v4l2-dkms bump needed.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Fri, 22 May 2026 17:30:00 +0000
+
+daedalus-v4l2 (0.1.0+r41+g6e6dfa1-1) bookworm trixie; urgency=medium
+
+  * Bump to 6e6dfa1 — daedalus-v4l2 PR #16.  Daemon dlopens Kwiboo
+    fourier fork's libavcodec.so.62 / libavformat.so.62 /
+    libavutil.so.60 at /opt/fourier instead of Debian-stock
+    soname 61/61/59.  First step on the daedalus-fourier
+    substitution arc (daedalus-v4l2#11): the next PR series
+    layers daedalus_recipe_dispatch_h264_* substitution patches
+    into ffmpeg-v4l2-request-fourier's H264DSPContext NEON init,
+    reaching the daemon's production decode path.
+  * Build: PKG_CONFIG_PATH now includes /opt/fourier/lib/pkgconfig
+    so daemon's pkg_check_modules picks up the Kwiboo .pc files.
+  * CI workflow build-deps: libavcodec-dev / libavformat-dev /
+    libavutil-dev (Debian stock 7.1.3) → ffmpeg-v4l2-request-fourier
+    (provides /opt/fourier/include + .pc files).
+  * Wire protocol unchanged.  No daedalus-v4l2-dkms bump.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 21:30:00 +0000
+
+daedalus-v4l2 (0.1.0+r39+g3bc0da1-1) bookworm trixie; urgency=medium
+
+  * Bump to 3bc0da1 — picks up daedalus-v4l2 PR #15.  Per-frame
+    `decoder: OK ...` log line gains `decode_us=N` (libavcodec
+    send_packet + receive_frame wall-clock cost in microseconds).
+    New `decoder stats` summary line every 60 decoded frames with
+    codec, fps, avg decode_us, MBs/s throughput, B/MB bitrate.
+  * Pure observability — no decode-path behaviour change.
+    Establishes baseline metrics for the substitution work in
+    daedalus-v4l2#11 step 2 (replacing libavcodec primitives with
+    daedalus-fourier kernels one cycle at a time).
+  * On Pi CM5 / bbb 720p H.264 baseline: ~4 ms decode_us / 24 fps
+    / 90 K MBs/s — workload is well under 1 % of any single
+    daedalus-fourier kernel's NEON ceiling.
+  * Wire protocol unchanged.  No daedalus-v4l2-dkms bump needed.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 18:30:00 +0000
+
+daedalus-v4l2 (0.1.0+r37+g77e14e5-1) bookworm trixie; urgency=medium
+
+  * Bump to 77e14e5 — picks up daedalus-v4l2 PRs #12 + #13.
+  * #12 (LOW_DELAY half-measure): the daemon now sets
+    AV_CODEC_FLAG_LOW_DELAY on the H.264 AVCodecContext so libavcodec
+    emits frames in decode order ~99% of the time (a few stragglers
+    at GOP boundaries when the stream's SPS num_reorder_frames
+    overrides the flag).  Visible improvement vs the 2-1-4-3
+    pair-swap on Firefox YouTube + mpv playback; not a permanent
+    fix (see #11 for the architectural plan).
+  * #13 (daedalus-fourier linkage): the daemon now pkg-config-links
+    against the daedalus-fourier kernel library (marfrit/
+    daedalus-fourier) and logs substrate availability at startup.
+    No kernels dispatched yet — this is the build-time / link-time
+    foundation for the H.264 daemon-rewrite plan in #11
+    (substituting daedalus-fourier IDCT 4×4 / IDCT 8×8 / luma
+    deblock primitives for libavcodec's per-MB pixel math, one
+    cycle at a time, measuring CPU saved per substitution).
+  * Build-deb.sh now fetches + builds + installs daedalus-fourier
+    (pinned at d87239d, marfrit/daedalus-fourier PR #1) into a
+    per-build temp prefix, then builds the daemon with
+    PKG_CONFIG_PATH pointing at it.  daedalus-fourier is
+    statically linked into the daemon binary, so the resulting
+    .deb has no new runtime deps.  Requires libvulkan-dev +
+    glslang-tools on the CI runner (the daedalus-fourier benches
+    already needed those).
+  * Wire protocol unchanged — DAEDALUS_PROTO_VERSION stays at 0.
+    No daedalus-v4l2-dkms bump needed.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 16:30:00 +0000
+
+daedalus-v4l2 (0.1.0+r33+g5d8b436-1) bookworm trixie; urgency=medium
+
+  * Bump to 5d8b436 — reverts daedalus-v4l2 PRs #7 + #8 (the parking
+    design that broke libva-v4l2-request-fourier's 1:1 CAPTURE
+    contract; see daedalus-v4l2#9 + #10).  After daemon-r28+g79256dc
+    landed, mpv (--hwdec=vaapi-copy) failed pre-playing with
+    "Unable to dequeue buffer: Resource temporarily unavailable" /
+    "Failed to end picture decode" because the daemon parked CAPTURE
+    buffers waiting for libavcodec to release H.264 B-frames in
+    display order — violating the V4L2 stateless 1:1 contract.
+    Firefox tolerated the mess (visible "2 1 4 3" pair-swap); mpv
+    bailed.
+  * This bump restores f0d4186-equivalent behaviour, plus PR #4
+    (cosmetic H.264 DECODE_MODE / START_CODE menu controls).  PR #7
+    + PR #8 wire-protocol additions (src_pts / output_src_pts /
+    RESP_FRAME flags) are reverted — DAEDALUS_PROTO_VERSION drops
+    back from 1 → 0.  Lock-step install with daedalus-v4l2-dkms
+    0.1.0+r33+g5d8b436 REQUIRED.
+  * Visible regression: H.264 B-frame streams in Firefox revert to
+    the original "2 1 4 3 6 5" pair-swap visual.  The proper fix
+    (concurrent in-flight requests in daemon + display-order reorder
+    in libva-v4l2-request-fourier) is tracked at daedalus-v4l2#11.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 14:50:00 +0000
+
+daedalus-v4l2 (0.1.0+r28+g79256dc-1) bookworm trixie; urgency=medium
+
+  * Bump to 79256dc — H.264 B-frame display reorder fix (closes
+    daedalus-v4l2#6 + #4 menu controls).  Daemon side: the
+    avcodec_send_packet → receive_frame loop now stamps pkt->pts =
+    req->src_pts so libavcodec's display-ordered frame->pts identifies
+    which OUTPUT bitstream's pixels each drained frame belongs to.
+    chardev_client maintains a (src_pts → cookie) lookup table so the
+    daemon can ship pixels to the cookie of the *originating*
+    bitstream, not the cookie of whatever REQ triggered the
+    receive_frame call.  Multiple RESP_FRAME messages per REQ_DECODE
+    are now possible (one for the just-consumed src, one or more for
+    drained pixels).
+  * Wire-protocol bump (DAEDALUS_PROTO_VERSION 0 → 1): REQ_DECODE
+    gains __u64 src_pts; RESP_FRAME gains __u32 flags +
+    __u64 output_src_pts.  Daemon + kernel must install atomically
+    (this package + daedalus-v4l2-dkms 0.1.0+r28+g79256dc).
+  * Also subsumes 79256dc's predecessor 7ff2d89 — H.264 DECODE_MODE +
+    START_CODE menu-control registration that retires the
+    "Unable to set control(s) error_idx=2/2" warning libva-v4l2-
+    request emitted on every context init.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Thu, 21 May 2026 12:00:00 +0000
+
+daedalus-v4l2 (0.1.0+r24+gf0d4186-1) bookworm trixie; urgency=medium
+
+  * Bump to f0d4186 — kernel per-ctx vb2 lock fix.  daedalus_queue_init
+    was wiring src_vq->lock and dst_vq->lock to ctx->dev->m2m_lock (a
+    device-wide mutex), serialising every vb2 ioctl across all
+    concurrent clients of /dev/video0.  For Firefox (which spawns
+    separate content + RDD + GPU processes that each open the device
+    and run libva probe simultaneously), one libva session's
+    S_FMT(OUTPUT_MPLANE) hit EBUSY while another was mid-streamon —
+    Firefox VAAPI playback fell apart at startup.
+  * Fix gives each open() its own ctx->vb_mutex; vb2 ioctls run
+    independently per client.  Matches cedrus / rkvdec / hantro
+    pattern.
+  * Verified on higgs: Firefox YouTube playback engages VAAPI cleanly,
+    sustained ~230 fps decode at 640x368 through the daedalus daemon,
+    zero EBUSY in stderr or daemon journal.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Wed, 20 May 2026 23:00:00 +0000
+
+daedalus-v4l2 (0.1.0+r22+g462aa4b-1) bookworm trixie; urgency=medium
+
+  * Bump to 462aa4b — kernel-side fix for control-binding gap that
+    closes the libva→daemon SPS/PPS pipeline.  Kernel device_run now
+    calls v4l2_ctrl_request_setup() before reading ctrl->p_cur, so
+    the daemon's daedalus_h264_meta block actually carries THIS
+    request's V4L2 stateless H.264 control values instead of stale
+    /default ones.  Pairs with libva-v4l2-request-fourier r382+gc1bb444
+    (Fix 3 + Fix 4 from issue libva-v4l2-request-fourier#8).
+  * After-fix on higgs (Pi CM5): ffmpeg -hwaccel vaapi -i h264.mp4
+    produces unique decoded P-frames (per-frame fnv1a hashes differ)
+    and zero "error while decoding MB" / "reference frames exceeds
+    max" warnings.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Wed, 20 May 2026 22:00:00 +0000
+
+daedalus-v4l2 (0.1.0+r20+g3dd0eb0-1) bookworm trixie; urgency=medium
+
+  * Bump to 3dd0eb0 — DAEMON-PPS H.264 SPS/PPS NAL synthesiser.
+    Daemon now reconstructs AnnexB SPS+PPS NAL units from the V4L2
+    stateless H.264 control structs (forwarded by the kernel via
+    a new struct daedalus_h264_meta block in REQ_DECODE) and
+    prepends them to the slice bitstream before feeding libavcodec.
+    Without this, ffmpeg -hwaccel vaapi on H.264 sources failed
+    with "non-existing PPS 0 referenced" even after LIBVA-1/-2
+    routing correctly delivered the request.
+  * Wire protocol: new DAEDALUS_REQ_FLAG_H264_META bit + struct
+    daedalus_h264_meta; daemon and kernel must be installed in
+    lockstep (this package + daedalus-v4l2-dkms 0.1.0+r20+g3dd0eb0).
+  * VP9 / AV1 paths unchanged.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Wed, 20 May 2026 21:00:00 +0000
+
+daedalus-v4l2 (0.1.0+r18+g481279c-1) bookworm trixie; urgency=medium
+
+  * Bump to 481279c.  Upstream landed the systemd unit + modules-load.d
+    drop-in (packaging/systemd/daedalus-v4l2.{service,modules-load}).
+  * Package now ships /lib/systemd/system/daedalus-v4l2.service and
+    /usr/lib/modules-load.d/daedalus-v4l2.conf.
+  * postinst: daemon-reload, enable the service, trigger
+    systemd-modules-load, and start if /dev/daedalus-v4l2 is already
+    present.  Operator no longer needs to remember the modprobe +
+    systemctl-enable dance after install.
+  * prerm/postrm: stop + disable + daemon-reload.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Wed, 20 May 2026 16:00:00 +0000
+
+daedalus-v4l2 (0.1.0+r17+gf0cd29a-1) bookworm trixie; urgency=medium
+
+  * Bump to f0cd29a alongside daedalus-v4l2-dkms.  No userspace
+    change — the daemon binary in this release is bit-identical to
+    the f55b2cd one (the f0cd29a commit only touches kernel/).  Bump
+    keeps the userspace + DKMS pkgver lockstep so depmod /
+    LIBVA_DRIVER_NAME selection stays correct on hosts that
+    upgrade.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Tue, 20 May 2026 10:30:00 +0000
+
+daedalus-v4l2 (0.1.0+r16+gf55b2cd-1) bookworm trixie; urgency=medium
+
+  * Initial Debian packaging for the daedalus-v4l2 userspace daemon.
+  * Pinned to f55b2cd (Phase 8.13 close): first commit where the full
+    ffmpeg -hwaccel vaapi → libva-v4l2-request-fourier → /dev/video0
+    → daemon path lands a pixel-correct decoded NV12 frame back in
+    ffmpeg.
+  * Codecs: VP9, AV1, H.264 (all via dlopen'd FFmpeg 7.1.3).
+  * Capture formats: NV12M (2 plane), NV12 (1 plane, for libva),
+    P010 (10-bit single plane).
+  * Companion package: daedalus-v4l2-dkms (kernel module).
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Mon, 18 May 2026 23:00:00 +0000

debian/daedalus-v4l2/debian/control

@@ -0,0 +1,34 @@
+Source: daedalus-v4l2
+Section: video
+Priority: optional
+Maintainer: Markus Fritsche <mfritsche@reauktion.de>
+Build-Depends: debhelper-compat (= 13),
+               cmake,
+               ninja-build,
+               pkg-config,
+               libavformat-dev (>= 7.1),
+               libavcodec-dev (>= 7.1),
+               libavutil-dev (>= 7.1)
+Standards-Version: 4.6.2
+Homepage: https://git.reauktion.de/reauktion/daedalus-v4l2
+
+Package: daedalus-v4l2
+Architecture: arm64
+Depends: ${misc:Depends}, ${shlibs:Depends},
+         ffmpeg (>= 7.1),
+         libdrm2
+Recommends: daedalus-v4l2-dkms
+Description: Userspace daemon for the daedalus_v4l2 stateless decoder shim
+ daedalus-v4l2 ships the userspace daemon that backs the daedalus_v4l2
+ out-of-tree V4L2 kernel module on Raspberry Pi 5 / CM5. Together they
+ expose /dev/videoNN + /dev/mediaNN as a V4L2 stateless decoder for
+ VP9, AV1, and H.264 — actual decoding happens in this single-threaded
+ daemon via dlopen'd FFmpeg, with decoded NV12 / P010 frames shipped
+ back through dmabuf.
+ .
+ Consumed end-to-end by libva-v4l2-request-fourier (>= 1.0.0+r376) so
+ that 'ffmpeg -hwaccel vaapi' against vp9_small.ivf produces a
+ byte-exact NV12 frame.
+ .
+ The kernel module ships separately in daedalus-v4l2-dkms; install
+ both to actually serve V4L2 clients.

debian/daedalus-v4l2/debian/copyright

@@ -0,0 +1,40 @@
+Format: https://www.debian.org/doc/packaging-manuals/copyright-format/1.0/
+Upstream-Name: daedalus-v4l2
+Upstream-Contact: Markus Fritsche <fritsche.markus@gmail.com>
+Source: https://git.reauktion.de/reauktion/daedalus-v4l2
+
+Files: *
+Copyright: 2026 Markus Fritsche <fritsche.markus@gmail.com>
+License: BSD-2-Clause
+
+Files: include/daedalus_v4l2_proto.h
+Copyright: 2026 Markus Fritsche <fritsche.markus@gmail.com>
+License: GPL-2.0-or-later WITH Linux-syscall-note
+Comment:
+ Shared kernel↔daemon wire-protocol header. GPL-2.0-or-later (matches
+ the kernel module that includes it) with the standard
+ Linux-syscall-note exception so userspace inclusion is BSD-clean.
+
+License: BSD-2-Clause
+ Redistribution and use in source and binary forms, with or without
+ modification, are permitted provided that the following conditions
+ are met:
+ .
+ 1. Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+ 2. Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in
+    the documentation and/or other materials provided with the
+    distribution.
+ .
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES ARE DISCLAIMED.
+
+License: GPL-2.0-or-later
+ This program is free software; you can redistribute it and/or modify
+ it under the terms of the GNU General Public License as published by
+ the Free Software Foundation; either version 2 of the License, or
+ (at your option) any later version.
+ .
+ On Debian systems, the complete text of the GNU General Public
+ License v2 can be found in `/usr/share/common-licenses/GPL-2'.

debian/ffmpeg-v4l2-request-fourier/0001-libudev-bypass-fallback.patch

@@ -0,0 +1,166 @@
+--- a/libavutil/hwcontext_v4l2request.c
++++ b/libavutil/hwcontext_v4l2request.c
+@@ -19,12 +19,13 @@
+ #include "config.h"
+ 
+ #include <fcntl.h>
+ #include <linux/dma-buf.h>
+ #include <linux/media.h>
+ #include <sys/ioctl.h>
++#include <sys/stat.h>
+ #include <sys/mman.h>
+ #include <unistd.h>
+ 
+ #include <drm_fourcc.h>
+ #include <libudev.h>
+ 
+@@ -690,12 +691,125 @@
+     }
+ 
+     udev_enumerate_unref(enumerate);
+     return ret;
+ }
+ 
++/*
++ * Brute-force fallback used when libudev's scan fails (e.g. inside firefox's
++ * RDD sandbox where Mozilla's broker rejects fd-relative openat used by
++ * systemd's chase() symlink resolver). Iterates /dev/video[0..63], picks the
++ * one whose major/minor matches the requested devnum.
++ */
++static char *v4l2request_devnum_to_video_path_brute(dev_t devnum)
++{
++    char path[32];
++    struct stat st;
++    for (int i = 0; i < 64; i++) {
++        snprintf(path, sizeof(path), "/dev/video%d", i);
++        if (stat(path, &st) < 0)
++            continue;
++        if (st.st_rdev == devnum)
++            return av_strdup(path);
++    }
++    return NULL;
++}
++
++/* Brute-force version of v4l2request_probe_video_devices: replaces the
++ * udev_device_new_from_devnum + udev_device_get_devnode flow with
++ * stat()-based major/minor matching against /dev/video[0..63]. */
++static int v4l2request_probe_video_devices_brute(AVHWFramesContext *hwfc,
++                                                 uint32_t pixelformat,
++                                                 uint32_t buffersize)
++{
++    AVV4L2RequestFramesContext *fctx = hwfc->hwctx;
++    AVV4L2RequestFramesContextInternal *fctxi = fctx->internal;
++    struct media_device_info device_info;
++    struct media_v2_topology topology = {0};
++    struct media_v2_interface *interfaces;
++    char *path;
++    dev_t devnum;
++    int ret;
++
++    if (ioctl(fctxi->media_fd, MEDIA_IOC_DEVICE_INFO, &device_info) < 0)
++        return AVERROR(errno);
++
++    if (ioctl(fctxi->media_fd, MEDIA_IOC_G_TOPOLOGY, &topology) < 0)
++        return AVERROR(errno);
++
++    if (!topology.num_interfaces)
++        return AVERROR(ENOENT);
++
++    interfaces = av_calloc(topology.num_interfaces,
++                           sizeof(struct media_v2_interface));
++    if (!interfaces)
++        return AVERROR(ENOMEM);
++
++    topology.ptr_interfaces = (__u64)(uintptr_t)interfaces;
++    if (ioctl(fctxi->media_fd, MEDIA_IOC_G_TOPOLOGY, &topology) < 0) {
++        ret = AVERROR(errno);
++        goto fail;
++    }
++
++    ret = AVERROR(ENOENT);
++    for (unsigned i = 0; i < topology.num_interfaces; i++) {
++        if (interfaces[i].intf_type != MEDIA_INTF_T_V4L_VIDEO)
++            continue;
++
++        devnum = makedev(interfaces[i].devnode.major,
++                         interfaces[i].devnode.minor);
++        path = v4l2request_devnum_to_video_path_brute(devnum);
++        if (!path)
++            continue;
++
++        ret = v4l2request_probe_video_device(hwfc, path, pixelformat, buffersize);
++        if (!ret) {
++            av_log(hwfc, AV_LOG_INFO,
++                   "Using V4L2 media driver %s (brute-force) for %s\n",
++                   device_info.driver, av_fourcc2str(pixelformat));
++            av_free(path);
++            break;
++        }
++        av_free(path);
++    }
++
++fail:
++    av_free(interfaces);
++    return ret;
++}
++
++/* Brute-force fallback for v4l2request_probe_media_devices(). Iterates
++ * /dev/media[0..15], opens each, probes via topology+stat. */
++static int v4l2request_probe_media_devices_brute(AVHWFramesContext *hwfc,
++                                                 uint32_t pixelformat,
++                                                 uint32_t buffersize)
++{
++    AVV4L2RequestFramesContext *fctx = hwfc->hwctx;
++    AVV4L2RequestFramesContextInternal *fctxi = fctx->internal;
++    char path[32];
++    int ret = AVERROR(ENOENT);
++
++    for (int i = 0; i < 16; i++) {
++        snprintf(path, sizeof(path), "/dev/media%d", i);
++
++        fctxi->media_fd = open(path, O_RDWR);
++        if (fctxi->media_fd < 0)
++            continue;
++
++        ret = v4l2request_probe_video_devices_brute(hwfc, pixelformat,
++                                                    buffersize);
++        if (!ret)
++            return 0;
++
++        close(fctxi->media_fd);
++        fctxi->media_fd = -1;
++    }
++
++    return ret;
++}
++
+ static int v4l2request_open_decoder(AVHWFramesContext *hwfc)
+ {
+     AVV4L2RequestFramesContext *fctx = hwfc->hwctx;
+     uint32_t buffersize;
+     struct udev *udev;
+     int ret;
+@@ -712,12 +826,23 @@
+ 
+     buffersize = FFMAX(hwfc->width * hwfc->height * 3 / 2, 256 * 1024);
+ 
+     // Probe all media devices (auto-detection)
+     ret = v4l2request_probe_media_devices(hwfc, udev, fctx->pixelformat, buffersize);
+ 
++    // Brute-force fallback when libudev fails. Firefox-fourier hits this
++    // because Mozilla's RDD sandbox blocks fd-relative openat used by
++    // systemd's chase() symlink resolver inside udev_enumerate_scan_devices.
++    if (ret < 0) {
++        av_log(hwfc, AV_LOG_INFO,
++               "libudev probe failed (%d), falling back to brute-force /dev/media*\n",
++               ret);
++        ret = v4l2request_probe_media_devices_brute(hwfc, fctx->pixelformat,
++                                                    buffersize);
++    }
++
+     udev_unref(udev);
+     return ret;
+ }
+ 
+ static AVBufferRef *v4l2request_v4l2_buffer_alloc(AVHWFramesContext *hwfc,
+                                                   struct v4l2_format *format)

debian/ffmpeg-v4l2-request-fourier/0002-nv15-to-p010-unpack.patch

@@ -0,0 +1,178 @@
+From 0cd6e669735e453ec8772f111065bbb2f70a5bc6 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Mon, 18 May 2026 07:27:10 +0000
+Subject: [PATCH] avutil/hwcontext_v4l2request: unpack NV15 to P010 in
+ transfer_data_from
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+V4L2_PIX_FMT_NV15 (RK3399/RK3588 rkvdec 10-bit 4:2:0 capture) is mapped to
+sw_format = AV_PIX_FMT_YUV420P10 in v4l2request_capture_pixelformats[]. The
+existing transfer_get_formats explicitly blanked the format list for that
+sw_format, so 'ffmpeg -hwaccel v4l2request -vf hwdownload,format=p010le' on
+a Hi10P / Main10 input failed at filter init with EINVAL before reaching
+the actual decode (which itself succeeds — 2 frames decoded cleanly).
+
+Expose AV_PIX_FMT_P010 as the transfer target for NV15-backed surfaces and
+unpack the packed 10-bit samples into the standard high-bits-of-16 layout
+inside transfer_data_from. Luma and chroma share the same packing format
+(5 bytes per 4 samples, little endian); chroma plane is W × H/2 samples
+for 4:2:0.
+
+The other 'needs custom unpack' sw_formats (YUV420P / Allwinner NV12_32L32
+tiled and YUV422P10 / rkvdec NV20) keep the original ENOSYS path because
+they need different unpack code that isn't covered by this patch.
+
+Closes marfrit/marfrit-packages#21.
+---
+ libavutil/hwcontext_v4l2request.c | 111 +++++++++++++++++++++++++++++-
+ 1 file changed, 110 insertions(+), 1 deletion(-)
+
+diff --git a/libavutil/hwcontext_v4l2request.c b/libavutil/hwcontext_v4l2request.c
+index b6633d9081..3842160dfb 100644
+--- a/libavutil/hwcontext_v4l2request.c
++++ b/libavutil/hwcontext_v4l2request.c
+@@ -1073,6 +1073,56 @@ fail:
+     return ret;
+ }
+ 
++/*
++ * Unpack one NV15-packed 10-bit plane (5 bytes per 4 samples, little endian)
++ * into a P010-style plane (10 bits in the high bits of a 16-bit container).
++ * `dst_stride` is in bytes; `src_stride` is bytes per row of NV15 data.
++ */
++static void v4l2request_nv15_unpack_plane_to_p010(const uint8_t *src,
++                                                  uint16_t *dst,
++                                                  unsigned width,
++                                                  unsigned height,
++                                                  unsigned src_stride,
++                                                  unsigned dst_stride)
++{
++    for (unsigned y = 0; y < height; y++) {
++        const uint8_t *s = src + y * src_stride;
++        uint16_t *d = (uint16_t *)((uint8_t *)dst + y * dst_stride);
++        unsigned x;
++
++        for (x = 0; x + 4 <= width; x += 4) {
++            uint16_t a = (uint16_t)s[0] | ((uint16_t)(s[1] & 0x03) << 8);
++            uint16_t b = ((uint16_t)s[1] >> 2) | ((uint16_t)(s[2] & 0x0F) << 6);
++            uint16_t c = ((uint16_t)s[2] >> 4) | ((uint16_t)(s[3] & 0x3F) << 4);
++            uint16_t e = ((uint16_t)s[3] >> 6) | ((uint16_t)s[4] << 2);
++
++            d[0] = (uint16_t)(a << 6);
++            d[1] = (uint16_t)(b << 6);
++            d[2] = (uint16_t)(c << 6);
++            d[3] = (uint16_t)(e << 6);
++
++            d += 4;
++            s += 5;
++        }
++
++        if (x < width) {
++            unsigned rem = width - x;
++            uint16_t pix[4] = { 0, 0, 0, 0 };
++
++            pix[0] = (uint16_t)s[0] | ((uint16_t)(s[1] & 0x03) << 8);
++            if (rem >= 2)
++                pix[1] = ((uint16_t)s[1] >> 2) | ((uint16_t)(s[2] & 0x0F) << 6);
++            if (rem >= 3)
++                pix[2] = ((uint16_t)s[2] >> 4) | ((uint16_t)(s[3] & 0x3F) << 4);
++            if (rem >= 4)
++                pix[3] = ((uint16_t)s[3] >> 6) | ((uint16_t)s[4] << 2);
++
++            for (unsigned j = 0; j < rem; j++)
++                d[j] = (uint16_t)(pix[j] << 6);
++        }
++    }
++}
++
+ static int v4l2request_transfer_get_formats(AVHWFramesContext *hwfc,
+                                             enum AVHWFrameTransferDirection dir,
+                                             enum AVPixelFormat **formats)
+@@ -1082,6 +1132,22 @@ static int v4l2request_transfer_get_formats(AVHWFramesContext *hwfc,
+     if (dir == AV_HWFRAME_TRANSFER_DIRECTION_TO)
+         return AVERROR(ENOSYS);
+ 
++    /*
++     * NV15-backed surfaces (sw_format = YUV420P10) are exposed as P010 to
++     * downstream filters: the unpack below converts the packed 10-bit
++     * samples into the standard high-bits-of-16 layout. Hi10P / Main10
++     * VAAPI/v4l2-request decode reaches userspace through this path.
++     */
++    if (hwfc->sw_format == AV_PIX_FMT_YUV420P10) {
++        fmts = av_malloc_array(2, sizeof(*fmts));
++        if (!fmts)
++            return AVERROR(ENOMEM);
++        fmts[0] = AV_PIX_FMT_P010;
++        fmts[1] = AV_PIX_FMT_NONE;
++        *formats = fmts;
++        return 0;
++    }
++
+     fmts = av_malloc_array(2, sizeof(*fmts));
+     if (!fmts)
+         return AVERROR(ENOMEM);
+@@ -1089,8 +1155,13 @@ static int v4l2request_transfer_get_formats(AVHWFramesContext *hwfc,
+     fmts[0] = hwfc->sw_format;
+     fmts[1] = AV_PIX_FMT_NONE;
+ 
++    /*
++     * Tiled-NV12-32L32 (Allwinner) and NV20 (rkvdec 4:2:2 10-bit) still need
++     * dedicated unpacks before hwdownload can consume them; leave them as
++     * "no transfer formats" so the filter graph reports the limitation
++     * rather than silently producing garbage.
++     */
+     if (hwfc->sw_format == AV_PIX_FMT_YUV420P ||
+-        hwfc->sw_format == AV_PIX_FMT_YUV420P10 ||
+         hwfc->sw_format == AV_PIX_FMT_YUV422P10)
+         fmts[0] = AV_PIX_FMT_NONE;
+ 
+@@ -1110,6 +1181,44 @@ static int v4l2request_transfer_data_from(AVHWFramesContext *hwfc,
+     map = av_frame_alloc();
+     if (!map)
+         return AVERROR(ENOMEM);
++
++    /*
++     * For NV15→P010, map the raw NV15 bytes (sw_format) and unpack into
++     * dst's P010 storage. Otherwise fall through to the original byte-copy
++     * path used for 1:1 sw_format matches (NV12, NV16, AFBC handled by DRM).
++     */
++    if (hwfc->sw_format == AV_PIX_FMT_YUV420P10) {
++        /*
++         * Only P010 is advertised by transfer_get_formats for this sw_format;
++         * a caller that bypasses get_formats and asks for anything else would
++         * silently corrupt output via av_frame_copy on NV15-packed bytes.
++         * Reject explicitly.
++         */
++        if (dst->format != AV_PIX_FMT_P010) {
++            ret = AVERROR(ENOSYS);
++            goto fail;
++        }
++
++        map->format = hwfc->sw_format;
++        ret = v4l2request_map_frame(hwfc, map, src);
++        if (ret)
++            goto fail;
++
++        v4l2request_nv15_unpack_plane_to_p010(map->data[0],
++                                              (uint16_t *)dst->data[0],
++                                              dst->width, dst->height,
++                                              map->linesize[0],
++                                              dst->linesize[0]);
++        /* NV15 chroma plane is W × H/2 samples (4:2:0, UV interleaved). */
++        v4l2request_nv15_unpack_plane_to_p010(map->data[1],
++                                              (uint16_t *)dst->data[1],
++                                              dst->width, dst->height / 2,
++                                              map->linesize[1],
++                                              dst->linesize[1]);
++        ret = 0;
++        goto fail;
++    }
++
+     map->format = dst->format;
+ 
+     ret = v4l2request_map_frame(hwfc, map, src);
+-- 
+2.47.3
+

debian/ffmpeg-v4l2-request-fourier/0003-h264-idct4-daedalus-fourier.patch

@@ -0,0 +1,137 @@
+From f760c0541586f43334c02611fcb4c212c08ad576 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Thu, 21 May 2026 21:40:22 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 4x4 IDCT through
+ daedalus-fourier
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+H264DSPContext.idct_add (called per 4x4 block from the intra-4x4
+decode path in h264_mb.c) now dispatches through
+daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
+
+The recipe layer picks the substrate; for cycle 6 (H.264 IDCT 4x4)
+the recipe is CPU NEON, so this is effectively a NEON-to-NEON
+substitution with one extra dispatch call and recipe-table lookup.
+Provides the first end-to-end exercise of the daedalus-fourier
+kernel pack inside the libavcodec.so decode hot path; follow-up
+patches wire IDCT 8x8, luma-v deblock, and qpel mc20.
+
+The library context is process-global, lazily initialised under
+pthread_once on first call.  We pick the no-QPU constructor because
+libavcodec.so is loaded into arbitrary host processes
+(firefox-fourier, mpv-fourier, daedalus_v4l2_daemon, ...) and we
+cannot assume the host has a usable Vulkan instance.  Higher cycles
+(deblock luma-v, MC) that benefit from the QPU will provision their
+own recipe-selected context once that path is wired.
+
+Bulk paths (idct_add16, idct_add16intra, idct_add8 — used for
+non-intra4x4 macroblocks) remain on the stock NEON .S implementations
+and will be batched through daedalus_recipe_dispatch_h264_idct4 with
+n_blocks>1 in a follow-up.
+
+Bit-exact against ff_h264_idct_add_neon (daedalus-fourier cycle 6
+green; see marfrit/daedalus-fourier/CYCLE_LOGS.md).
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2.
+---
+ libavcodec/aarch64/Makefile               |  3 +-
+ libavcodec/aarch64/h264_idct_daedalus.c   | 49 +++++++++++++++++++++++
+ libavcodec/aarch64/h264dsp_init_aarch64.c |  3 +-
+ 3 files changed, 53 insertions(+), 2 deletions(-)
+ create mode 100644 libavcodec/aarch64/h264_idct_daedalus.c
+
+diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
+index 41ab025..7b95fb1 100644
+--- a/libavcodec/aarch64/Makefile
++++ b/libavcodec/aarch64/Makefile
+@@ -3,7 +3,8 @@ OBJS-$(CONFIG_AC3DSP)                   += aarch64/ac3dsp_init_aarch64.o
+ OBJS-$(CONFIG_FDCTDSP)                  += aarch64/fdctdsp_init_aarch64.o
+ OBJS-$(CONFIG_FMTCONVERT)               += aarch64/fmtconvert_init.o
+ OBJS-$(CONFIG_H264CHROMA)               += aarch64/h264chroma_init_aarch64.o
+-OBJS-$(CONFIG_H264DSP)                  += aarch64/h264dsp_init_aarch64.o
++OBJS-$(CONFIG_H264DSP)                  += aarch64/h264dsp_init_aarch64.o \
++                                           aarch64/h264_idct_daedalus.o
+ OBJS-$(CONFIG_HUFFYUVDSP)               += aarch64/huffyuvdsp_init_aarch64.o
+ OBJS-$(CONFIG_H264PRED)                 += aarch64/h264pred_init.o
+ OBJS-$(CONFIG_H264QPEL)                 += aarch64/h264qpel_init_aarch64.o
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+new file mode 100644
+index 0000000..538d223
+--- /dev/null
++++ b/libavcodec/aarch64/h264_idct_daedalus.c
+@@ -0,0 +1,49 @@
++/*
++ * H.264 4x4 IDCT + add — daedalus-fourier substitution shim.
++ *
++ * Routes H264DSPContext.idct_add through
++ * daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
++ * The recipe layer picks the substrate (CPU NEON by default for
++ * cycle 6; future cycles may dispatch to V3D opportunistically).
++ *
++ * FFmpeg's 4x4 block memory layout matches daedalus's column-major
++ * convention: block[r + 4*c] = coefficient at (row r, col c).  Both
++ * sides destructively zero the block after the transform.
++ *
++ * The library context is process-global and lazily initialised under
++ * pthread_once.  We pick the no-QPU constructor here because
++ * libavcodec.so is loaded into arbitrary host processes
++ * (firefox-fourier, mpv-fourier, daedalus_v4l2_daemon, ...) and we
++ * cannot assume the host has a usable Vulkan instance.  Higher cycles
++ * (deblock, MC) that benefit from the QPU initialise their own
++ * recipe-selected context once that path is wired.
++ */
++
++#include <pthread.h>
++#include <stddef.h>
++#include <stdint.h>
++
++#include <daedalus.h>
++
++#include "libavutil/attributes.h"
++#include "libavcodec/h264dsp.h"
++
++static daedalus_ctx     *g_dctx;
++static pthread_once_t    g_dctx_once = PTHREAD_ONCE_INIT;
++
++static void daedalus_ctx_init_once(void)
++{
++    g_dctx = daedalus_ctx_create_no_qpu();
++}
++
++void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
++
++void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
++{
++    static const daedalus_h264_block_meta meta = { .dst_off = 0 };
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_idct4(g_dctx, dst, (size_t)stride,
++                                        block, 1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+index c684574..b993df2 100644
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
++++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
+@@ -66,6 +66,7 @@ void ff_biweight_h264_pixels_4_neon(uint8_t *dst, uint8_t *src, ptrdiff_t stride
+                                     int weights, int offset);
+ 
+ void ff_h264_idct_add_neon(uint8_t *dst, int16_t *block, int stride);
++void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct_dc_add_neon(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct_add16_neon(uint8_t *dst, const int *block_offset,
+                              int16_t *block, int stride,
+@@ -139,7 +140,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
+         c->biweight_pixels_tab[1] = ff_biweight_h264_pixels_8_neon;
+         c->biweight_pixels_tab[2] = ff_biweight_h264_pixels_4_neon;
+ 
+-        c->idct_add        = ff_h264_idct_add_neon;
++        c->idct_add        = ff_h264_idct_add_daedalus;
+         c->idct_dc_add     = ff_h264_idct_dc_add_neon;
+         c->idct_add16      = ff_h264_idct_add16_neon;
+         c->idct_add16intra = ff_h264_idct_add16intra_neon;
+-- 
+2.47.3
+

debian/ffmpeg-v4l2-request-fourier/0004-h264-idct8-daedalus-fourier.patch

@@ -0,0 +1,107 @@
+From 1b286ddb4efaca26ec9b9e290e989fec77dc1c77 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Fri, 22 May 2026 10:18:21 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 8x8 IDCT through
+ daedalus-fourier
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+H264DSPContext.idct8_add (called per 8x8 block from the High-profile
+intra-8x8-DCT decode path in h264_mb.c) now dispatches through
+daedalus_recipe_dispatch_h264_idct8 instead of ff_h264_idct8_add_neon.
+
+The recipe layer picks the substrate; for cycle 7 (H.264 IDCT 8x8)
+the recipe is CPU NEON, so this is effectively a NEON-to-NEON
+substitution layered on top of the cycle-6 IDCT 4x4 wiring.  Same
+pthread_once global context, same destructive-zero semantics; FFmpeg
+column-major 8x8 storage block[r + 8*c] matches daedalus's convention.
+
+Bulk path c->idct8_add4 (used for inter 8x8-DCT macroblocks) remains
+on the in-tree NEON .S code and will be batched through
+daedalus_recipe_dispatch_h264_idct8 with n_blocks>1 in a follow-up.
+
+Bit-exact against ff_h264_idct8_add_neon (daedalus-fourier cycle 7
+green).
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 7.
+---
+ libavcodec/aarch64/h264_idct_daedalus.c   | 29 ++++++++++++++++-------
+ libavcodec/aarch64/h264dsp_init_aarch64.c |  3 ++-
+ 2 files changed, 23 insertions(+), 9 deletions(-)
+
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+index 538d223..cbb98af 100644
+--- a/libavcodec/aarch64/h264_idct_daedalus.c
++++ b/libavcodec/aarch64/h264_idct_daedalus.c
+@@ -1,14 +1,16 @@
+ /*
+- * H.264 4x4 IDCT + add — daedalus-fourier substitution shim.
++ * H.264 4x4 / 8x8 IDCT + add — daedalus-fourier substitution shims.
+  *
+- * Routes H264DSPContext.idct_add through
+- * daedalus_recipe_dispatch_h264_idct4 instead of ff_h264_idct_add_neon.
+- * The recipe layer picks the substrate (CPU NEON by default for
+- * cycle 6; future cycles may dispatch to V3D opportunistically).
++ * Routes H264DSPContext.idct_add  → daedalus_recipe_dispatch_h264_idct4
++ *        H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
++ * instead of the in-tree ff_h264_idct{,8}_add_neon assembly.  The
++ * recipe layer picks the substrate (CPU NEON by default for cycles
++ * 6 + 7; future cycles may dispatch to V3D opportunistically).
+  *
+- * FFmpeg's 4x4 block memory layout matches daedalus's column-major
+- * convention: block[r + 4*c] = coefficient at (row r, col c).  Both
+- * sides destructively zero the block after the transform.
++ * FFmpeg's 4x4 and 8x8 block memory layouts match daedalus's
++ * column-major convention: block[r + N*c] = coefficient at
++ * (row r, col c) for N ∈ {4, 8}.  Both sides destructively zero the
++ * block after the transform.
+  *
+  * The library context is process-global and lazily initialised under
+  * pthread_once.  We pick the no-QPU constructor here because
+@@ -37,6 +39,7 @@ static void daedalus_ctx_init_once(void)
+ }
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
++void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+ {
+@@ -47,3 +50,13 @@ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+     daedalus_recipe_dispatch_h264_idct4(g_dctx, dst, (size_t)stride,
+                                         block, 1, &meta);
+ }
++
++void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride)
++{
++    static const daedalus_h264_block_meta meta = { .dst_off = 0 };
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_idct8(g_dctx, dst, (size_t)stride,
++                                        block, 1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+index b993df2..741e551 100644
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
++++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
+@@ -79,6 +79,7 @@ void ff_h264_idct_add8_neon(uint8_t **dest, const int *block_offset,
+                             const uint8_t nnzc[15 * 8]);
+ 
+ void ff_h264_idct8_add_neon(uint8_t *dst, int16_t *block, int stride);
++void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct8_dc_add_neon(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct8_add4_neon(uint8_t *dst, const int *block_offset,
+                              int16_t *block, int stride,
+@@ -146,7 +147,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
+         c->idct_add16intra = ff_h264_idct_add16intra_neon;
+         if (chroma_format_idc <= 1)
+             c->idct_add8   = ff_h264_idct_add8_neon;
+-        c->idct8_add       = ff_h264_idct8_add_neon;
++        c->idct8_add       = ff_h264_idct8_add_daedalus;
+         c->idct8_dc_add    = ff_h264_idct8_dc_add_neon;
+         c->idct8_add4      = ff_h264_idct8_add4_neon;
+     } else if (have_neon(cpu_flags) && bit_depth == 10) {
+-- 
+2.47.3
+

debian/ffmpeg-v4l2-request-fourier/0005-h264-deblock-luma-v-daedalus-fourier.patch

@@ -0,0 +1,121 @@
+From 68731c41d7ea68be0e912b128cb4e71fb56e8263 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Fri, 22 May 2026 12:15:16 +0200
+Subject: [PATCH] avcodec/aarch64/h264dsp: route H.264 luma-v deblock through
+ daedalus-fourier
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+H264DSPContext.v_loop_filter_luma (non-intra bS<4 vertical luma
+deblock, called per macroblock-row edge from the slice deblock
+loop) now dispatches through
+daedalus_recipe_dispatch_h264_deblock_luma_v instead of
+ff_h264_v_loop_filter_luma_neon.
+
+The recipe layer picks the substrate; for cycle 8 the daedalus
+docstring marks the kernel "CPU primary; QPU opportunistic", but
+the libavcodec.so context here is built with
+daedalus_ctx_create_no_qpu — process-global pthread_once init,
+shared with cycles 6/7.  QPU opportunism stays gated off until a
+follow-up adds an explicit feature flag (no implicit Vulkan init
+in arbitrary host processes).  In the meantime cycle 8 is a
+plumbing-only substitution, NEON-to-NEON via the daedalus recipe.
+
+Intra (bS=4) loop filter — c->v_loop_filter_luma_intra — stays on
+the in-tree NEON .S code; daedalus's daedalus_h264_deblock_meta
+only covers the non-intra path per its docstring.
+
+FFmpeg `int alpha/beta/int8_t tc0[4]` → daedalus_h264_deblock_meta
+(int32_t alpha/beta + inline int8_t tc0[4]).  pix already points
+to row 0 of the bottom block per FFmpeg's deblock convention,
+satisfying daedalus's `dst_off >= 4 * dst_stride` constraint.
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 cycle 8.
+---
+ libavcodec/aarch64/h264_idct_daedalus.c   | 36 +++++++++++++++++++----
+ libavcodec/aarch64/h264dsp_init_aarch64.c |  4 ++-
+ 2 files changed, 33 insertions(+), 7 deletions(-)
+
+diff --git a/libavcodec/aarch64/h264_idct_daedalus.c b/libavcodec/aarch64/h264_idct_daedalus.c
+index cbb98af..92365fa 100644
+--- a/libavcodec/aarch64/h264_idct_daedalus.c
++++ b/libavcodec/aarch64/h264_idct_daedalus.c
+@@ -1,11 +1,14 @@
+ /*
+- * H.264 4x4 / 8x8 IDCT + add — daedalus-fourier substitution shims.
++ * H.264 4x4 / 8x8 IDCT + luma-v deblock — daedalus-fourier substitution shims.
+  *
+- * Routes H264DSPContext.idct_add  → daedalus_recipe_dispatch_h264_idct4
+- *        H264DSPContext.idct8_add → daedalus_recipe_dispatch_h264_idct8
+- * instead of the in-tree ff_h264_idct{,8}_add_neon assembly.  The
+- * recipe layer picks the substrate (CPU NEON by default for cycles
+- * 6 + 7; future cycles may dispatch to V3D opportunistically).
++ * Routes H264DSPContext.idct_add           → daedalus_recipe_dispatch_h264_idct4
++ *        H264DSPContext.idct8_add          → daedalus_recipe_dispatch_h264_idct8
++ *        H264DSPContext.v_loop_filter_luma → daedalus_recipe_dispatch_h264_deblock_luma_v
++ * instead of the in-tree ff_h264_*_neon assembly.  The recipe layer
++ * picks the substrate (CPU NEON for cycles 6 + 7 by default; cycle 8
++ * is CPU primary with QPU opportunistic — the ctx below is no-QPU,
++ * so cycle 8 stays on the CPU NEON path until a separate change
++ * gates QPU init on a daedalus-fourier feature flag).
+  *
+  * FFmpeg's 4x4 and 8x8 block memory layouts match daedalus's
+  * column-major convention: block[r + N*c] = coefficient at
+@@ -40,6 +43,8 @@ static void daedalus_ctx_init_once(void)
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride);
+ void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride);
++void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0);
+ 
+ void ff_h264_idct_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+ {
+@@ -60,3 +65,22 @@ void ff_h264_idct8_add_daedalus(uint8_t *dst, int16_t *block, int stride)
+     daedalus_recipe_dispatch_h264_idct8(g_dctx, dst, (size_t)stride,
+                                         block, 1, &meta);
+ }
++
++void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0)
++{
++    daedalus_h264_deblock_meta meta = {
++        .dst_off = 0,
++        .alpha   = alpha,
++        .beta    = beta,
++    };
++    meta.tc0[0] = tc0[0];
++    meta.tc0[1] = tc0[1];
++    meta.tc0[2] = tc0[2];
++    meta.tc0[3] = tc0[3];
++
++    pthread_once(&g_dctx_once, daedalus_ctx_init_once);
++
++    daedalus_recipe_dispatch_h264_deblock_luma_v(g_dctx, pix, (size_t)stride,
++                                                 1, &meta);
++}
+diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c b/libavcodec/aarch64/h264dsp_init_aarch64.c
+index 741e551..85ac381 100644
+--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
++++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
+@@ -27,6 +27,8 @@
+ 
+ void ff_h264_v_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                      int beta, int8_t *tc0);
++void ff_h264_v_loop_filter_luma_daedalus(uint8_t *pix, ptrdiff_t stride,
++                                         int alpha, int beta, int8_t *tc0);
+ void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+                                      int beta, int8_t *tc0);
+ void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, ptrdiff_t stride, int alpha,
+@@ -114,7 +116,7 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c, const int bit_depth,
+     int cpu_flags = av_get_cpu_flags();
+ 
+     if (have_neon(cpu_flags) && bit_depth == 8) {
+-        c->v_loop_filter_luma   = ff_h264_v_loop_filter_luma_neon;
++        c->v_loop_filter_luma   = ff_h264_v_loop_filter_luma_daedalus;
+         c->h_loop_filter_luma   = ff_h264_h_loop_filter_luma_neon;
+         c->v_loop_filter_luma_intra= ff_h264_v_loop_filter_luma_intra_neon;
+         c->h_loop_filter_luma_intra= ff_h264_h_loop_filter_luma_intra_neon;
+-- 
+2.47.3
+

debian/ffmpeg-v4l2-request-fourier/0006-h264-restore-low-delay.patch

@@ -0,0 +1,82 @@
+From 0d1292ea99bc4e5fa2da438259fa01a2374e3e04 Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <mfritsche@reauktion.de>
+Date: Fri, 22 May 2026 14:18:25 +0200
+Subject: [PATCH] avcodec/h264: restore AV_CODEC_FLAG_LOW_DELAY semantics
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+FFmpeg 8.x dropped the H.264 decoder's low_delay path —
+AV_CODEC_FLAG_LOW_DELAY no longer prevents
+h264_select_output_frame from running the display-order DPB
+output queue.  V4L2-stateless-style consumers (daedalus-v4l2
+daemon, libva-v4l2-request-fourier) that set the flag end up
+seeing the 2-1-4-3 pair-swap pattern on B-frame streams again.
+
+Restore the documented semantics:
+
+  - Early-exit at the top of h264_select_output_frame when the
+    flag is set: emit the just-decoded picture immediately as
+    next_output_pic, mirror the corruption / recovery-point
+    tracking the main path performs, and skip the entire
+    delayed_pic[] / POC reorder machinery.
+
+  - Suppress the SPS-driven has_b_frames clobber in
+    h264_field_start when the flag is set, so the per-slice
+    bitstream_restriction_flag re-pickup cannot reintroduce a
+    nonzero reorder buffer mid-stream.
+
+This is a fork-only change required by the daedalus-v4l2 daemon's
+one-frame-per-send_packet contract; upstream FFmpeg consumers that
+expect display-order output remain untouched (flag default = off).
+
+Refs reauktion/daedalus-v4l2#11 — substitution arc step 2 deblock
++ flag-restoration follow-up.
+---
+ libavcodec/h264_slice.c | 23 +++++++++++++++++++++++
+ 1 file changed, 23 insertions(+)
+
+diff --git a/libavcodec/h264_slice.c b/libavcodec/h264_slice.c
+index 97fab70..a7bfbd6 100644
+--- a/libavcodec/h264_slice.c
++++ b/libavcodec/h264_slice.c
+@@ -1308,6 +1308,28 @@ static int h264_select_output_frame(H264Context *h)
+     cur->mmco_reset = h->mmco_reset;
+     h->mmco_reset = 0;
+ 
++    /* AV_CODEC_FLAG_LOW_DELAY restore (FFmpeg 8.x dropped the H.264
++     * decoder's low_delay path).  Bypass the display-order DPB
++     * output queue: emit the just-decoded picture immediately, in
++     * decode order, one per send_packet.  V4L2-stateless-style
++     * consumers (daedalus-v4l2 daemon, libva-v4l2-request-fourier)
++     * do their own POC-based reorder downstream and require this
++     * behaviour. */
++    if (h->avctx->flags & AV_CODEC_FLAG_LOW_DELAY) {
++        h->next_output_pic    = cur;
++        h->next_outputed_poc  = cur->poc;
++        h->frame_recovered   |= cur->recovered;
++        cur->recovered       |= h->frame_recovered & FRAME_RECOVERED_SEI;
++        if (!cur->recovered) {
++            if (!(h->avctx->flags  & AV_CODEC_FLAG_OUTPUT_CORRUPT) &&
++                !(h->avctx->flags2 & AV_CODEC_FLAG2_SHOW_ALL))
++                h->next_output_pic = NULL;
++            else
++                cur->f->flags |= AV_FRAME_FLAG_CORRUPT;
++        }
++        return 0;
++    }
++
+     if (sps->bitstream_restriction_flag ||
+         h->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT) {
+         h->avctx->has_b_frames = FFMAX(h->avctx->has_b_frames, sps->num_reorder_frames);
+@@ -1415,6 +1437,7 @@ static int h264_field_start(H264Context *h, const H264SliceContext *sl,
+     sps = h->ps.sps;
+ 
+     if (sps->bitstream_restriction_flag &&
++        !(h->avctx->flags & AV_CODEC_FLAG_LOW_DELAY) &&
+         h->avctx->has_b_frames < sps->num_reorder_frames) {
+         h->avctx->has_b_frames = sps->num_reorder_frames;
+     }
+-- 
+2.47.3
+