linux-ampere-fourier: bump to kafr2 (v7.0-rc3 + fleet/ampere.yaml manifest)

Drops the f8f3ad9 baseline ("18 commits ahead, BLACK-SCREENS ampere"
per fleet/ampere.yaml bisect note) in favor of the kernel-agent-
managed 10-patch set produced by ka-promote from fleet/ampere.yaml.

Baseline: mainline v7.0-rc3 (3daa4f5dc6cc), plus the 10 scope-tagged
patches under marfrit/kernel-agent/patches/{soc,module,board,driver}/:
  - 1 soc/rk3588 pwm15 pinctrl
  - 6 board/coolpi-cm5-genbook DTS patches
  - 3 driver/media Sarma VP9-on-VDPU381 patches (PR #24 closure)

New _commit 48a8c78 reflects this tree state in ~/src/linux-rockchip
(branch vp9-build on ampere, exactly v7.0-rc3 + 10 manifest patches).

End-to-end verified before this iteration was cut (hand-build of the
same tip on 2026-05-18 booted ampere via arch_vp9_test extlinux
label):
  - VP9 decode bit-exact HW==SW==libva (sha c8624d7c42db66525f53a02a515bc38d0a17ef39f692660cc7bebb1e2d2e1b48)
  - AV1 decode bit-exact HW==SW via kdirect (sha 30d2091158d92f3c5e0a807217c3e7307f873267673d92632e7fb147383e7dd1, av1-vpu-dec is mainline 7.0)

prebuild.sh canonical sha256 cleared — gzip-version-dependent, the
script warns-not-fails on mismatch. First successful kafr2 build can
pin a canonical sha here if a reproducibility audit ever needs it.

Cross-references:
  - marfrit/kernel-agent#12 (VP9 enablement closure)
  - marfrit/kernel-agent PR #24 (Sarma patch import + ampere.yaml bump)

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/PKGBUILD

@@ -24,7 +24,7 @@ _srcname=FFmpeg
 _version='8.1'
 _commit='b57fbbe50c9b2656fad86a1a7eeabfd2b2a50935'  # v4l2-request-n8.1 tip 2026-04-24
 pkgver=8.1.r123329.b57fbbe
-pkgrel=4
+pkgrel=5
 epoch=2
 pkgdesc='FFmpeg with V4L2 Request API hwaccel (Rockchip / Allwinner stateless decode)'
 arch=('aarch64')
@@ -78,8 +78,9 @@ 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')
+        '0001-libudev-bypass-fallback.patch'
+        '0002-nv15-to-p010-unpack.patch')
+sha256sums=('SKIP' 'SKIP' 'SKIP')
 
 pkgver() {
   cd "${_srcname}"
@@ -91,6 +92,7 @@ pkgver() {
 prepare() {
   cd "${_srcname}"
   patch -Np1 -i "${srcdir}/0001-libudev-bypass-fallback.patch"
+  patch -Np1 -i "${srcdir}/0002-nv15-to-p010-unpack.patch"
 }
 
 build() {

arch/linux-ampere-fourier/PKGBUILD

@@ -1,14 +1,21 @@
 # Maintainer: Markus Fritsche <mfritsche@reauktion.de>
 #
-# linux-ampere-fourier — CoolPi GenBook (RK3588) kernel from the
-# marfrit-iterated linux-rk3588-marfrit branch (pinned commit).
+# linux-ampere-fourier — CoolPi GenBook (RK3588) kernel built from the
+# kernel-agent fleet/ampere.yaml manifest applied to mainline v7.0-rc3.
 #
-# Baseline carries the marfrit RK3588 delta on top of mainline v7.0-rc3:
-# 10 Markus commits (board DTS + Kconfig + suspend/wakeup + bluetooth)
-# plus 8 upstream cherry-picks (Shawn Lin's pcie3 phy series, Collabora
-# clk/dts/dw-dp fixes, Sebastian Reichel's Rock 5 ITX hdmirx). The 6
-# board-relevant patches are scope-tagged in marfrit/kernel-agent under
-# patches/{soc,module,board}/ — see fleet/ampere.yaml manifest.
+# 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
@@ -21,9 +28,9 @@
 
 pkgbase=linux-ampere-fourier
 pkgname=("$pkgbase" "$pkgbase-headers")
-pkgver=7.0rc3.kafr1
+pkgver=7.0rc3.kafr2
 pkgrel=1
-pkgdesc='CoolPi GenBook kernel (linux-rk3588-marfrit @ f8f3ad9 + RK3588 cherry-picks + GenBook DTS)'
+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)
@@ -34,13 +41,24 @@ makedepends=(
 )
 options=('!strip')
 
-# Pinned tip of marfrit/linux-rk3588-marfrit. 18 commits ahead of v7.0-rc3:
-#   10 by Markus Fritsche  (board/SoC/module/Bluetooth/Kconfig)
-#    4 by Shawn Lin        (phy: rockchip-snps-pcie3 stability series)
-#    2 by Cristian Ciocaltea (clk + dw-dp bridge — Collabora track)
-#    1 by Sebastian Reichel (Rock 5 ITX hdmirx — not used by ampere)
-#    1 by Pedro Alves      (CLK_SET_RATE_PARENT VOP2 fix)
-_commit=f8f3ad934433bd7e1207d9b0b37e817a692b7ee9
+# 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

arch/linux-ampere-fourier/prebuild.sh

@@ -19,8 +19,11 @@
 set -euo pipefail
 
 TREE="${LINUX_RK3588_MARFRIT_TREE:-${HOME}/src/linux-rockchip}"
-COMMIT=f8f3ad934433bd7e1207d9b0b37e817a692b7ee9
-SHA256_EXPECTED=b4eca11e883fe6f7f306d8751c3efa3afed9cc3465c74a3320de1b7204f5f330
+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"

arch/lmcp/PKGBUILD

@@ -3,7 +3,7 @@
 # Source of truth: git.reauktion.de/marfrit/lmcp
 
 pkgname=lmcp
-pkgver=1.1.1
+pkgver=1.2.1
 pkgrel=1
 pkgdesc="Lightweight MCP (Model Context Protocol) server in pure Lua"
 arch=('any')
@@ -14,7 +14,7 @@ depends=('lua' 'lua-socket')
 # 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=('80c2e815aa61a2d3baab051c51cd247bdefa9dd03d72c4867b99c49b6eae9cb9')
+sha256sums=('bf9cce1a84c66b1b74c5aec923c5960d60ae33c221afc8d47ce0d74b8f7ee609')
 
 package() {
     cd "${pkgname}"

debian/lmcp/build-deb.sh

@@ -7,10 +7,10 @@
 # package (Architecture: all, depends on lua + lua-socket).
 set -euo pipefail
 
-PKGVER=1.1.1
-UPSTREAM_TAG=v1.1.1
+PKGVER=1.2.1
+UPSTREAM_TAG=v1.2.1
 PKGREL=1
-LMCP_TARBALL_SHA256=80c2e815aa61a2d3baab051c51cd247bdefa9dd03d72c4867b99c49b6eae9cb9
+LMCP_TARBALL_SHA256=bf9cce1a84c66b1b74c5aec923c5960d60ae33c221afc8d47ce0d74b8f7ee609
 HERE=$(dirname "$(readlink -f "$0")")
 
 # Reproducible build: pin all file mtimes + ar member timestamps to a fixed

debian/lmcp/debian/changelog

@@ -1,3 +1,24 @@
+lmcp (1.2.1-1) bookworm trixie; urgency=medium
+
+  * tools.d/ plugin scan (closes lmcp#22): server.lua now scans
+    LMCP_TOOLS_DIR (default /opt/lmcp/tools.d on POSIX) for *.lua
+    files and invokes each as a function receiving (server, run).
+    Lets hosts ship local tool extensions alongside the packaged
+    generics without forking server.lua. Existing single-file
+    deployments without a tools.d/ directory: no behaviour change.
+  * LMCP_HOST + LMCP_CONF env vars: packaged server.lua now threads
+    these into lmcp.new(opts.host, opts.conf). Hosts that need
+    single-interface binding (e.g. hertz on 192.168.88.18) or a
+    conf-file-based bearer token (e.g. /opt/herding/etc/hertz-tools.conf)
+    can drive the packaged entrypoint directly via systemd env
+    instead of carrying a forked server.lua.
+  * Together with the above, hertz, ampere, and any future host with
+    custom tools can migrate from /opt/lmcp/server.lua forks to a
+    plain plugin file + standard systemd env. apt upgrade then
+    delivers all packaged improvements automatically.
+
+ -- Markus Fritsche <mfritsche@reauktion.de>  Mon, 18 May 2026 01:30:00 +0000
+
 lmcp (1.1.1-1) bookworm trixie; urgency=medium
 
   * Bug fix: omit empty inputSchema.properties at tool registration.