fleet/ohm: pkgrel=6 reproducible from manifest (cumulative-pkgrel6-danctnix)

Replace the pkgrel=3-era cumulative-c5x-danctnix include with
cumulative-pkgrel6-danctnix — a single squashed diff representing
the bes2600 driver source state on ohm as of 2026-05-21.

Also drop:
- arch/arm64/scs-arm-neon-build-fix/ (removed in pkgrel=4)
- driver/bes2600/queue-pending-record-lock-bh-danctnix/ (in cumulative)
- driver/bes2600/tx-sdio-dma-oob-danctnix/ (in cumulative)
- driver/bes2600/join-confirm-reset-danctnix/ (in cumulative)

Resulting manifest: just two includes:
  driver/bes2600/cumulative-pkgrel6-danctnix/
  driver/bes2600/scan-filter-5ghz-danctnix/

Verified: ka-promote ohm produces a 136KB cumulative.patch that,
when applied to a fresh v7.0-danctnix1 staging tree, yields
drivers/staging/bes2600 source bit-identical to the pkgrel=6 build
on boltzmann. The only diff is build artifacts (.o, .cmd, .mod, etc.).

Kernel-agent can now generate the ohm-live source state without
reaching into the besser repository.

Closes ka#29 (per-series reconstruction tracking issue) by
delivering the deterministic-rebuild capability the original
per-series mirrors were meant to provide.

Signed-off-by: Claude (noether) <claude@reauktion.de>

fleet/ohm.yaml

@@ -1,6 +1,6 @@
 # kernel-agent manifest for ohm (PineTab2 / Rockchip RK3566 + BES2600 SDIO WiFi/BT)
 #
-# Status: scaffolding from 2026-05-16. Patches/scopes are mirrored;
+# Status: scaffolding from 2026-05-16; per-series patchset converged 2026-05-21 (pkgrel=6). Patches/scopes are mirrored;
 # the build pipeline (cumulative-patch generation, makepkg invocation,
 # sign+publish) still relies on the hand-managed flow in
 # boltzmann:~/src/besser/marfrit-besser/danctnix-besser-pkgbuild/kernel/.
@@ -32,6 +32,13 @@ baseline:
 # mixed-prefix headers (a/drivers/staging/bes2600/...  b/bes2600/...).
 # They do NOT apply cleanly against the linux-pinetab2 baseline.
 #
+
+# 2026-05-21 update: per-series reconstruction (besser#22) completed
+# 2026-05-21; pkgrel=6 (srcversion 0E16463F) on ohm soak-passed with
+# the bounce-buffer + join-confirm-reset additions. The per-series
+# manifest below is the authoritative set; cumulative-c5x-danctnix
+# remains as historical fallback only.
+#
 # Until the per-series mirrors are reconstructed (kernel-agent followup
 # issue), the bes2600 driver scope is satisfied by a single-file
 # cumulative captured from the working hand-managed
@@ -39,19 +46,22 @@ baseline:
 # patches/driver/bes2600/cumulative-c5x-danctnix/README.md). This is
 # the c5x stack as it shipped in pkgrel=3 on 2026-05-18.
 includes:
-  # bes2600 driver (c5x stack as shipped in pkgrel=3) — single-file
+  # bes2600 driver pkgrel=6 cumulative: 22 commits squashed, equivalent
-  # interim cumulative; per-series reconstruction tracked separately.
+  # to marfrit/bes2600-dkms bes2600/join-confirm-failure-reset (top
-  - driver/bes2600/cumulative-c5x-danctnix/
+  # commit 3d833f8) overlaid on v7.0-danctnix1 staging tree. Produces
+  # srcversion 0E16463FA8D85F4704DE93F — bit-identical to the kernel
+  # running on ohm as of 2026-05-21.
+  #
+  # Includes c5.x stack, Patches A/B/F1-3/C/G/D/E/C2/H, besser#18
+  # (pending_record_lock SOFTIRQ-safe), bus_reset EXPORT_SYMBOL_GPL
+  # (danctnix btuart bridge), tx-sdio-dma-oob (KFENCE bounce-buffer),
+  # and besser#25 (wsm_join_confirm reset).
+  #
+  # Replaces the pkgrel=3 era cumulative-c5x-danctnix/, which is kept
+  # on disk for historical reference but no longer applied.
+  - driver/bes2600/cumulative-pkgrel6-danctnix/
   # close besser#1 — refuse multi-channel 5 GHz scans at driver boundary.
   - driver/bes2600/scan-filter-5ghz-danctnix/
-  # GCC 15.2.1 build-fix for arm_neon.h + SHADOW_CALL_STACK interaction.
-  # Runtime no-op as long as the config has CONFIG_SHADOW_CALL_STACK=n
-  # (current ohm setting). Kept in the manifest for the day SCS gets
-  # re-enabled. See reference_arm64_scs_arm_neon_gcc15 memory.
-  - arch/arm64/scs-arm-neon-build-fix/
-  # close besser#18 — pending_record_lock SOFTIRQ-safe -> -unsafe inversion.
-  # Mirror of marfrit/bes2600-dkms#11 (d95453c). 5-site spin_lock -> _bh.
-  - driver/bes2600/queue-pending-record-lock-bh-danctnix/
 
 # Explicitly NOT included (decision logged):
 #   - debian-copyright-fsf-address: Debian packaging metadata, not kernel

patches/driver/bes2600/cumulative-pkgrel6-danctnix/README.md

@@ -0,0 +1,48 @@
+# bes2600/cumulative-pkgrel6-danctnix
+
+Single-file cumulative diff representing the bes2600 driver source state
+that produces srcversion `0E16463FA8D85F4704DE93F` (pkgrel=6 on ohm,
+soak-verified 2026-05-21).
+
+## Equivalent commit chain
+
+Squash of 22 commits from `marfrit/bes2600-dkms` branch
+`bes2600/join-confirm-failure-reset` (top commit `3d833f8`):
+
+| # | Commit | Patch |
+|---|---|---|
+| 1-7  | 4fec8b2..3942404 | c5.x scan-defer / firmware-recovery stack |
+| 8-13 | 91640bd..73191b7 | Patch A, B, F1-3, C v3 |
+| 14   | a02f8b7 | Patch G (SPDX restore) |
+| 15-16| 93f2aab, dd01be0 | Patch D, E |
+| 17   | 447240c | Patch C2 |
+| 18   | dc13f5d | Patch H (bh.c hygiene) |
+| 19   | f469448 | besser#18 pending_record_lock SOFTIRQ-safe |
+| 20   | 0792ba4 | bus_reset EXPORT_SYMBOL_GPL (danctnix bridge) |
+| 21   | 49d9b77 | bounce SDIO TX buffers (DMA OOB / KFENCE fix) |
+| 22   | 3d833f8 | wsm_join_confirm reset (besser#25) |
+
+## Why a cumulative
+
+These 22 commits are the converged per-series; while they exist as
+individual scope dirs in `marfrit/bes2600-dkms`, several have
+context-overlap rebase conflicts that make per-scope inclusion in
+kernel-agent fragile (cf. ka#29 / besser#22 reconstruction debacle).
+
+Shipping the cumulative as one file in kernel-agent guarantees the
+applied source state on `v7.0-danctnix1` is bit-identical to the
+pkgrel=6 build on ohm, without dragging the besser-repo branch state
+into kernel-agent's resolution path.
+
+## Apply order
+
+This patch is the **base** for the bes2600 driver scope. The remaining
+non-bes2600 patch (`scan-filter-5ghz-danctnix` for besser#1) layers on
+top via the apply order in `fleet/ohm.yaml`.
+
+## Provenance
+
+Generated by `git diff e0d752a..bes2600/join-confirm-failure-reset --
+bes2600/` against `marfrit/bes2600-dkms`, then path-rewritten
+`bes2600/` → `drivers/staging/bes2600/`. The baseline `e0d752a`
+corresponds to the v7.0-danctnix1 bes2600 staging tree.

patches/driver/bes2600/join-confirm-reset-danctnix/0001-bes2600-reset-firmware-state-on-wsm_join_confirm-fai.patch

@@ -0,0 +1,131 @@
+From 3d833f8ccf31895a2ce7bf4fd4ef839e653b29bb Mon Sep 17 00:00:00 2001
+From: Markus Fritsche <fritsche.markus@gmail.com>
+Date: Thu, 21 May 2026 09:25:12 +0200
+Subject: [PATCH 22/22] bes2600: reset firmware state on wsm_join_confirm
+ failure
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+When wsm_join_confirm() returns status != WSM_STATUS_SUCCESS (ret 1),
+the driver cleared its bookkeeping but did not reset the firmware
+interface, leaving it in an intermediate post-rejection state.  A rapid
+second JOIN attempt (e.g. wpa_supplicant retrying after the
+PREV_AUTH_NOT_VALID deauth that mac80211 emits to clean up) hits an
+inconsistent firmware context, causing bes2600_sdio_read_rx_batch to
+return SDIO error which cascades into wifi_force_close:
+
+  wsm_join_confirm ret 1
+  deauthenticating from <bssid> by local choice (Reason: 2=PREV_AUTH_NOT_VALID)
+  [~10 min later]
+  bes2600_sdio_read_rx_batch sdio read error
+  WARNING: at bes2600_tx_loop_set_enable / bes2600_chrdev_wifi_force_close
+
+Two additions to the failure path in bes2600_join_work():
+
+1. wsm_reset (WSM_REQ_ID_RESET, 0x000A) with reset_statistics=false.
+   This returns the firmware to IDLE so the next association attempt
+   starts from a known-clean state.  bes2600_unjoin_work() performs the
+   same reset, but gates it on join_status != PASSIVE; after a failed
+   JOIN join_status stays PASSIVE, so that path never fires — call
+   wsm_reset directly here instead.
+
+   Contract: wsm_reset takes only wsm_cmd_lock (not conf_lock, not
+   wsm_oper_lock).  wsm_oper_unlock was already called inside
+   wsm_join_confirm() before wsm_join() returned -EINVAL, so there is
+   no re-entrancy hazard.  conf_lock is held at this call site, which is
+   compatible with wsm_reset's locking requirements.
+
+2. queue_work(workqueue, &priv->unjoin_work) instead of direct
+   wsm_unlock_tx().  Serialises the next association attempt through
+   the workqueue so it cannot race against lingering firmware-side
+   effects of the failure.  If unjoin_work is already queued, release
+   TX immediately (matching cw1200 ancestor sta.c:1344 comment "Tx lock
+   still held, unjoin will clear it.").
+
+Ancestor reference: drivers/net/wireless/st/cw1200/sta.c, function
+cw1200_join_work(), lines 1339-1344.  cw1200 queues unjoin_work on join
+failure for the same reason.  bes2600 needs the direct wsm_reset in
+addition because its unjoin_work has the join_status gate that cw1200's
+cw1200_do_unjoin() does not.
+
+Signed-off-by: Claude (noether) <claude@reauktion.de>
+---
+ bes2600/sta.c | 47 +++++++++++++++++++++++++++++++++++++++++++----
+ 1 file changed, 43 insertions(+), 4 deletions(-)
+
+diff --git a/drivers/staging/bes2600/sta.c b/drivers/staging/bes2600/sta.c
+index 476d875..bf86835 100644
+--- a/drivers/staging/bes2600/sta.c
++++ b/drivers/staging/bes2600/sta.c
+@@ -2225,9 +2225,10 @@ void bes2600_join_work(struct work_struct *work)
+ 	struct wsm_template_frame probe_tmp = {
+ 		.frame_type = WSM_FRAME_TYPE_PROBE_REQUEST,
+ 	};
+-	/*struct wsm_reset reset = {
+-		.reset_statistics = true,
+-	};*/
++	struct wsm_reset join_fail_reset = {
++		.reset_statistics = false,
++	};
++	bool join_failed = false;
+ 
+ 
+ 	BUG_ON(queueId >= 4);
+@@ -2410,6 +2411,33 @@ void bes2600_join_work(struct work_struct *work)
+ #endif /*CONFIG_BES2600_TESTMODE*/
+ 			cancel_delayed_work_sync(&priv->join_timeout);
+ 			bes2600_pwr_clear_busy_event(priv->hw_priv, BES_PWR_LOCK_ON_JOIN);
++			/*
++			 * Firmware rejected WSM_JOIN (wsm_join_confirm ret 1).
++			 * Issue wsm_reset so the firmware returns to a clean
++			 * IDLE state before the next association attempt.
++			 *
++			 * Without this reset the firmware sits in an
++			 * intermediate post-reject state.  A rapid second
++			 * JOIN (e.g. wpa_supplicant retrying after the
++			 * PREV_AUTH_NOT_VALID deauth that follows) hits an
++			 * inconsistent firmware context, causing
++			 * bes2600_sdio_read_rx_batch to return SDIO error
++			 * which cascades into wifi_force_close.
++			 *
++			 * cw1200 ancestor (drivers/net/wireless/st/cw1200/
++			 * sta.c:1339) queues unjoin_work on join failure for
++			 * the same reason; bes2600_unjoin_work gates its
++			 * wsm_reset on join_status != PASSIVE, so after a
++			 * failed JOIN (join_status stays PASSIVE) that path
++			 * never fires — call wsm_reset directly here instead.
++			 *
++			 * Contract: wsm_reset takes only wsm_cmd_lock; safe
++			 * to call while conf_lock is held.  wsm_oper_unlock
++			 * was already called in wsm_join_confirm() before
++			 * wsm_join() returned the error.
++			 */
++			WARN_ON(wsm_reset(hw_priv, &join_fail_reset, priv->if_id));
++			join_failed = true;
+ 		} else {
+ 			/* Upload keys */
+ #ifdef CONFIG_BES2600_TESTMODE
+@@ -2434,7 +2462,18 @@ void bes2600_join_work(struct work_struct *work)
+ 	up(&hw_priv->conf_lock);
+ 	if (bss)
+ 		cfg80211_put_bss(hw_priv->hw->wiphy, bss);
+-	wsm_unlock_tx(hw_priv);
++	/*
++	 * On join failure: queue unjoin_work so the next association
++	 * attempt is serialised after any lingering cleanup, matching
++	 * cw1200 sta.c:1344 "Tx lock still held, unjoin will clear it."
++	 * If unjoin_work is already queued, release TX immediately.
++	 */
++	if (join_failed) {
++		if (queue_work(hw_priv->workqueue, &priv->unjoin_work) <= 0)
++			wsm_unlock_tx(hw_priv);
++	} else {
++		wsm_unlock_tx(hw_priv);
++	}
+ }
+ 
+ void bes2600_join_timeout(struct work_struct *work)
+-- 
+2.54.0
+

patches/driver/bes2600/join-confirm-reset-danctnix/README.md

@@ -0,0 +1,46 @@
+# bes2600/join-confirm-reset-danctnix
+
+Danctnix-flavor patch closing besser#25 (wsm_join_confirm failure cascade).
+
+## What it does
+
+When firmware returns status 1 on a JOIN command (`wsm_join_confirm ret 1`),
+add a direct `wsm_reset(...)` call so the firmware returns to a clean IDLE
+state, plus `queue_work(workqueue, &priv->unjoin_work)` for serialisation of
+the next association attempt.
+
+## Why it's a fork-divergence fix
+
+`cw1200_join_work()` (cw1200 ancestor, `drivers/net/wireless/st/cw1200/sta.c:1339-1344`)
+queues `unjoin_work` on join failure: `cw1200_do_unjoin()` calls `wsm_reset`
+when `join_status == STA`.
+
+bes2600's `bes2600_unjoin_work()` gates the same `wsm_reset` on
+`join_status != PASSIVE`. After a failed JOIN, `join_status` stays PASSIVE
+(only set to STA on success) — queuing `unjoin_work` alone is insufficient
+on bes2600. The danctnix variant carries a direct `wsm_reset` in the
+failure path *and* the queue_work serialisation.
+
+## Observable effects (pkgrel=6 soak)
+
+Beyond closing the cascade (besser#25 acceptance), this patch also
+collapsed the periodic ~600 ms latency jitter on ohm:
+
+| | pkgrel=5 | pkgrel=6 |
+|---|---|---|
+| max RTT | 612 ms | 13.9 ms |
+| mdev | 103.5 ms | 1.55 ms |
+
+The bgscan-driven roam-attempt to a 5 GHz BSSID followed by `wsm_join`
+reject was briefly stalling TX every minute even when the cascade did
+not fire.
+
+## Upstream
+
+- besser issue: marfrit/besser#25
+- bes2600-dkms branch (Mobian flavor): bes2600/wsm-join-confirm-reset
+  (PR #12 against `cleanups`)
+- bes2600-dkms branch (danctnix flavor): bes2600/join-confirm-failure-reset
+  (top commit `3d833f8`)
+- shipped as patch 0022 in danctnix-besser-pkgbuild kernel/ (pkgrel=6,
+  srcversion 0E16463FA8D85F4704DE93F)