pi3-64 hardware video playback (30fps, zero drops) + audio - #3164

Fixes #3084 (pi3-64 video renders black). On the Raspberry Pi 3 arm64
board (VideoCore IV, eglfs) QtMultimedia's VideoOutput cannot present a
video node on the vc4 GL driver — the frame is black — and compositing a
full-screen video widget through eglfs GL caps at ~9 fps regardless. This
adds a pi3-64-only path that keeps decode + scanout on the hardware and
composites the video on a vc4 DRM overlay plane, reaching a stable 30 fps
with zero ongoing dropped frames, with audio.

Video (VideoView::gstPlayOverlay):
- filesrc ! qtdemux ! h264parse ! v4l2h264dec (bcm2835 HW decoder) !
  queue ! kmssink on a vc4 overlay plane, using eglfs's own DRM master fd
  + connector via QPlatformNativeInterface; the plane is HW-composited
  with the eglfs UI plane.
- The decoder is driven capture-io-mode=mmap so kmssink COPIES each frame
  into its own scanout buffer instead of pinning the decoder's DPB pool —
  that pin deadlocked every plain decoder-direct attempt. The decoder
  emits I420 at full rate; the alternative ISP convert to NV12 is a second
  M2M pass that caps at ~18 fps, so it is only a fallback
  (ANTHIAS_GST_ISP_CONVERT=1).
- Falls back to a CPU-raster appsink path, then to QMediaPlayer, if the
  DRM resources or an overlay plane can't be resolved.

Audio (VideoView::gstStartAudio) runs in a SEPARATE pipeline:
- filesrc ! qtdemux audio ! decodebin ! audioconvert ! audioresample !
  alsasink on the HDMI card. Independent clock/preroll/bus so no audio
  condition (missing track, slow device, decode gap) can stall the video
  — a shared pipeline froze the video at the first frame.
- Adds gstreamer1.0-libav for the audio decoders (avdec_ac3/aac/mp3/…);
  decodebin reported a missing plug-in on AC3 with only faad (AAC).

Also: only pi3-64 links GStreamer (AnthiasViewer.pro CONFIG, Dockerfile
build + runtime deps); _build_webview_env enables the path via
ANTHIAS_VIDEO_RASTER/OVERLAY; start_viewer chowns the GStreamer registry
cache to avoid a per-launch rescan; main.cpp installs a fatal-signal
backtrace handler.

Validated on a real Pi 3 (1080p30 H.264/AC3, HDMI): 30 fps, dropped flat
at the startup ramp only, clean across the loop boundary, HDMI PCM
RUNNING, respawns=1.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

The RPi archive ships RPi-tuned +rpt3 builds of gstreamer plugins-base,
plugins-bad and alsa (the archive + keyring are already configured by
_rpt1-ffmpeg-pin.j2 — pi3-64 is in its board list). plugins-bad carries
the vc4 kmssink the overlay-plane path scans out through, so the RPi build
is the right one for production parity with Pi OS. plugins-good (the V4L2
HW decode/convert elements) and libav (audio decoders) aren't in the RPi
archive, so they stay on Debian. The +rpt3 plugins share upstream 1.26.2
with Debian's libgstreamer1.0-0 core (not in the RPi archive), so the ABI
matches.

Validated on a real Pi 3 (clean reboot): base/bad/alsa install as
1.26.2*+rpt3*, good/libav as Debian; 30fps zero ongoing drops, audio
playing, respawns=1 — unchanged from the Debian-plugins build.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

- main.cpp: scope the fatal-signal backtrace handler to the pi3-64 build
  (#ifdef ANTHIAS_GSTREAMER) instead of installing it fleet-wide. It was
  added to debug the pi3-64 overlay crashes; other boards keep the default
  crash behaviour. (It already re-raised via SIG_DFL, so core dumps were
  never suppressed — but this keeps the PR to its board scope.)
- gstPlay: normalise a file:// URI to a filesystem path before handing it
  to filesrc (Anthias passes bare local paths today; this is defensive).
- Start audio once in gstPlay so BOTH the overlay AND the appsink-raster
  fallback get sound — previously the raster fallback played silent
  (audio was only started in gstPlayOverlay). gstStartAudio now escapes
  its own path; gstPlayOverlay drops the unused options param.
- play(): log when gstPlay exhausts every GStreamer path (there is no
  QMediaPlayer fallback on this board) instead of returning silently.

Validated on a real Pi 3: video 30fps zero ongoing drops + audio (3.5mm)
still play; pi3-64 (with GStreamer) and pi4-64 (without) both compile.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

- videoview.cpp: clamp ANTHIAS_GST_QUEUE — a 0/negative/non-numeric value
  made queueBuffers 0, i.e. an unbounded queue (max-size-buffers/time/bytes
  all 0) → OOM on the 1 GB board; fall back to the default 4.
- gstStartAudio: check gst_element_set_state()'s result; on
  GST_STATE_CHANGE_FAILURE log + tear the audio pipeline down instead of
  leaving a dead pipeline reporting "started".
- containerFps: was written from the kmssink pad-probe (streaming thread)
  and read by sampleStats (GUI thread) as a plain double — a data race.
  Make the handoff atomic (QAtomicInteger fps×1000 via containerFps() /
  setContainerFps()).
- main.cpp: the fatal-signal handler now uses only async-signal-safe calls
  (write() + backtrace_symbols_fd) instead of fprintf/fflush.
- test_viewer.py: also assert ANTHIAS_VIDEO_OVERLAY is set on pi3-64 and
  cleared on other boards.

Validated on a real Pi 3: 30fps zero ongoing drops, container-fps=30.00,
audio (3.5mm) playing; pi3-64 (with GStreamer) and pi4-64 (without) both
compile.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

- sampleStats: on the gstMode appsink-raster fallback, query the GStreamer
  pipeline position instead of QMediaPlayer::position() (the player isn't
  the one playing there, so position-ms was misleading).
- gstPlayOverlay: guard a null QGuiApplication::primaryScreen() before
  using it for the DRM native-resource lookups; document why driFd<=0 is
  the correct "not available" test here (eglfs returns the fd through a
  void* API — a real DRM master fd is always >2, "no fd" comes back as 0).
- main.cpp: soften the crash-handler comment — backtrace()/
  backtrace_symbols_fd() are glibc extensions, not formally guaranteed
  async-signal-safe (best-effort last-gasp diagnostic).

Compiles on pi3-64 (with GStreamer) and pi4-64 (without); Pi 3 still
plays 30fps with audio.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

- gstPlay: treat a missing appsink (gst_bin_get_by_name null) as a hard
  failure — otherwise the pipeline goes PLAYING to a silent black screen.
- play(): only start the stats sampler when gstPlay() succeeds, so a
  failed start doesn't emit SAMPLE lines with position=-1/elapsed=-1.
- gstStartAudio: escape the ALSA device string before interpolating it
  into the gst_parse_launch description (same as the filesrc path).
- gstStop: drop the stashed frame (under the frame mutex) and reset the
  drain flag, so a late-queued onGstFrame() can't repaint a stale frame
  after teardown.

pi3-64 compiles; changes are pi3-64-scoped (#ifdef ANTHIAS_GSTREAMER) so
other boards are unaffected. Pi 3 still plays 30fps with audio.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

The gstMode STOP stats line hard-coded position-ms=0, which read like
playback never advanced. Query the pipeline position (it's still up at
that point; gstStop() runs after) and log it, or -1 if the query fails.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…und 6)

- anthias_gst_bus_loop: on GST_MESSAGE_ERROR, tear the pipeline down
  (queued VideoView::stop()) and return FALSE to drop the watch, instead
  of logging and leaving a dead pipeline on a stale frame — the next asset
  re-shows and rebuilds.
- gstPlay (appsink-raster path): install the same bus watch so pipeline
  errors are surfaced/recovered and EOS looping goes through one path
  (gstRestartLoop). Drops the now-redundant appsink EOS callback
  (anthias_gst_on_eos); appsink posts EOS to the bus like any GstBaseSink.

Validated on a real Pi 3: overlay path still plays 30fps, loops cleanly
across the 60s boundary (no respawn), audio playing. pi3-64-scoped
(#ifdef ANTHIAS_GSTREAMER); other boards unaffected.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Comment-only. The code evolved past several comments:
- appsink note said sync=true; the pipeline uses sync=false (GUI-side
  coalescing paces the paint) — corrected.
- gstStop said the appsink path has no bus watch; it now installs one
  (round 6) — corrected.
- two comments claimed decoder-direct I420 "renders blue" on the vc4
  plane, but decoder-direct I420 is the working default; reworded to
  describe the formats/mmap-copy accurately without the (thrashed-board
  artifact) "blue" assertion.

No code change.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…nd 8)

- STOP: on the overlay path report gstRawSamples as frames-rendered (the
  Qt frame counters stay 0 there, kmssink's tally is the real count).
- appsink-raster path: latch the source framerate from the negotiated
  caps (latchSourceFps) so its SAMPLE computes expected/dropped instead of
  reporting -1.
- Remove dead gstUri / gstRequeueUri (leftover playbin about-to-finish
  requeue path; the current pipelines don't use playbin).
- Fix the gstRestartLoop header comment (loops via the pipeline bus watch,
  not the removed appsink EOS callback).

Compiles on pi3-64 + pi4-64; overlay hot path unchanged from the
already-validated build.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

gstRawSamples was reset only in gstPlayOverlay, so the appsink-raster path
carried the previous asset's cumulative count into its stats. Reset it
once in gstPlay() right after teardown (covers overlay + appsink); drop
the now-redundant reset in gstPlayOverlay.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…ound 10)

- onGstFrame(): release gstDrainPending at the END of the drain, not the
  start. pushGstFrame() uses testAndSetOrdered(0,1) so exactly one drain is
  ever in flight; clearing at the end keeps it that way (frames arriving
  mid-drain just update gstLatestFrame) — bounded GUI event queue, no
  unbounded QImage backlog.
- gstPlay(): clear the raster surface right after teardown so the previous
  asset's last frame can't linger (stale-frame flash) while the new
  pipeline prerolls or if it fails to start.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

- main.cpp: install the fatal-signal handler via sigaction() with
  SA_RESETHAND instead of signal(). The handler no longer calls signal()
  (not async-signal-safe) — SA_RESETHAND restores the default disposition
  so the re-raise takes the default path (core dump / exit).
- onGstFrame: release gstDrainPending with storeRelease (was storeRelaxed)
  so it pairs with pushGstFrame's testAndSetOrdered acquire — the drain's
  writes are visible before the next drain can be queued.

(Copilot also flagged toImage() in onRasterPainted; that is the dormant
QMediaPlayer-raster path — pi3-64 uses gstMode, which feeds frames via
onGstFrame/setImage and never runs that toImage-in-paint path.)

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Comment-only. The two "CPU-raster presenter" comments predate the overlay
default; reworded to: pi3-64 uses the non-VideoOutput presenter — the
GStreamer vc4 overlay-plane path (preferred, ANTHIAS_VIDEO_OVERLAY) with
the CPU-raster blit as fallback.

(Copilot also suggested a QtTest that constructs VideoView with
ANTHIAS_VIDEO_RASTER=1; that headless C++ harness is a reasonable
follow-up — the env wiring is already covered by tests/test_viewer.py.)

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>