 marfrit
|
08a04c3791
|
Iteration 4 Phase 4: load-bearing fix landed — fresh request_fd per frame
Phase 4's diagnostic journey: TRY_EXT_CTRLS retry didn't pinpoint
(kernel obfuscation extends to TRY for compound controls). Per-control
TRY isolation showed all 4 H.264 controls fail individually on the same
fd → pivot from "bad control content" to "bad request_fd state."
Replaced REINIT with close+realloc; iter1+2+3 carryover frame-11
EINVAL empirically eliminated.
Fork commits: 74d8dd1 (DPB FFmpeg-semantics fixes), 385dee1 (fresh
request_fd per frame, load-bearing), f21bdf0 (debug TRY iso),
b81ce69 (B-slice L1 copy-paste fix from Phase 5 review).
Empirical: 49.7s of bbb_1080p30 stream-time decoded clean on
firefox-fourier without MOZ_DISABLE_RDD_SANDBOX=1.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-05-05 14:16:00 +00:00 |
|