fix(video): treat NOT_READY as transient with one IDR on resume

[mariotaku]

Summary

Companion to ss4s #16 (FeedGuard exclusive primitive) — closes the original race that #593 worked around, but does it correctly so HDR toggle and resolution changes actually work mid-stream instead of being silently suppressed.

Background

SS4S_PlayerVideoSetHDRInfo(NULL) on ndl-webos5 (and SS4S_PlayerVideoSizeChanged on most LG/NDL backends) internally calls NDL_DirectMediaUnload + NDL_DirectMediaLoad. Before ss4s #16, an in-flight driver->Feed could be running on the half-torn-down decoder and crash the stream. PR #593 worked around this by skipping the SetHDRInfo(NULL) call entirely for MAIN10 streams.

ss4s #16 added a proper drain primitive (BeginExclusive / EndExclusive on the FeedGuard): during the destructive op, Feed returns SS4S_VIDEO_FEED_NOT_READY so it never reaches the driver. But the app side still tears down the stream on NOT_READY because it falls into the default error path:

} else {
    commons_log_error("Session", "Video feed error %d", result);
    session_interrupt(session, false, STREAMING_INTERRUPT_DECODER);
    return DR_OK;
}

So the race is gone, but every HDR toggle still kills the stream.

Fix

Handle NOT_READY as a transient signal in vdec_delegate_submit:

Feed return	New behavior
OK (first one after NOT_READY)	Return DR_NEED_IDR — server resends a keyframe so the decoder resyncs from known-good data instead of a P-frame referencing the gap
OK (steady state)	DR_OK, normal stats bookkeeping
NOT_READY	Drop the frame silently (DR_OK), set need_idr_on_resume = true. No session_interrupt, no DR_NEED_IDR spam — exactly one IDR request per outage no matter how long it lasts.
REQUEST_KEYFRAME	DR_NEED_IDR (existing)
Anything else	session_interrupt(STREAMING_INTERRUPT_DECODER) (existing — real driver errors still tear the stream down)

Resets need_idr_on_resume = false in vdec_delegate_setup so stream restarts don't inherit stale state.

Bumps

Submodule pin 60d980a2 → ab442770. Picks up:

• fix(video): drain Feed during SetHDRInfo and SizeChanged (#16) — the exclusive primitive this PR pairs with
• ci: run tests on every push and pull request (#17)
• ci: add AddressSanitizer pipeline (with dynamic libasan) (#18)

Comparison with #593

Aspect	#593 (skip on MAIN10)	this PR + ss4s #16
Fixes the crash on MAIN10 HDR-off	✅	✅
Same fix for SizeChanged race	❌	✅ (free — same primitive)
HDR can actually be disabled mid-stream on 8-bit streams	unchanged	✅
HDR can be disabled mid-stream on MAIN10 streams	❌ (suppressed)	✅
Brief frame drop + IDR resync during toggle	n/a	~ms of dropped frames, single IDR request, no tear-down

Recommend closing #593 in favor of this combined fix.

Test plan

• cmake --build build/webos-debug --target moonlight clean
• Stream H.265 MAIN10 with HDR, toggle HDR off mid-stream — previously crashed (without fix: skip NDL HDR disable for MAIN10 to prevent decoder crash #593) or no-op'd (with fix: skip NDL HDR disable for MAIN10 to prevent decoder crash #593); should now show ~brief gap then keyframe, stream continues with TV in SDR mode
• Stream H.264 (8-bit), toggle HDR off — should show brief gap then continue (with fix: skip NDL HDR disable for MAIN10 to prevent decoder crash #593 alone this already worked)
• Change stream resolution mid-stream — SizeChanged uses the same primitive; should also survive without tear-down (previously this race was latent — same shape as the HDR one but no one had hit it)
• Regular streaming with no HDR/size events — should be byte-for-byte identical to before (only NOT_READY path is new)

🤖 Generated with Claude Code