Fix replay audio timing reliability

[rgarcia]

Summary

• Synchronize the PulseAudio input to x11grab with -isync 0. Both use wall-clock timestamps, but ffmpeg otherwise subtracts each input start independently, pulling later-opened audio forward and ending it early in the replay timeline.
• Bind the Playwright daemon socket before importing the selected browser engine; import esbuild only when transforming code. Previously all three packages initialized before listening. Keep the existing 5-second socket deadline.
• Explicitly apply and verify the fixture's intended 1280×720 display through the API. WIDTH/HEIGHT alone do not configure headful Xorg; CI was recording its 3840×2160 fallback instead.
• Retain the real-browser audio MP4, packet/stream timestamps and isolated service logs in e2e CI. Log measured durations and lifecycle timings; run the normal parallel e2e suite uncached.

The existing 2-second audio-duration assertion, audible-tone check, recording-before-first-Playwright ordering and e2e concurrency are unchanged. No audio padding, test retries or production deployment.

Regression evidence

• New ffmpeg regression models devices opening at timestamps 10 and 13 and both finishing at 16, retaining production encode/mux arguments. Pre-fix: video ends at 6.021387s, audio at 3.029333s. Fixed: 6.000000s / 6.008000s, before and after remux. Decoding still yields only about 3 seconds of audio, not padded silence.
• New daemon regression holds dependency evaluation until explicitly released. Pre-fix the socket remains absent; fixed it accepts connections while the engine import is blocked, then connects after release. This proves the startup dependency, not CPU contention in historical CI failures.

Validation

• Local make test-unit (vet and Go race tests) and make test-runtime pass. An earlier baseline unit run hit an unrelated devtoolsproxy Chromium TempDir cleanup race; the complete rerun passed.
• Full local parallel real-container suite passes with both updated headful/headless images and verified 1280×720 capture: video 10.100000s, audio 9.971750s, gap 0.128250s, no non-monotonic DTS warnings.
• The first candidate passed three independent CI jobs, but one had a 1.795333s gap, 28 non-monotonic audio DTS warnings and ffmpeg final speed 0.858x while encoding unintended 4K video. Validation was stopped proactively rather than accepting a near-threshold result. Correcting fixture setup reduces raw frame bandwidth from 2,654,208 to 294,912 kbit/s (9×), without changing suite parallelism or production recording settings. The local corrected full-suite run sustained 1.04x with no DTS warnings.
• Final commit 8897c2a881ce2a08a5faac4554e305399d5f972e passed ten consecutive independent full parallel e2e CI jobs. All normal CI checks and latest-commit BugBot are green; BugBot reported no issues. No code changed during validation.

Ten-run CI evidence

Workflow server-test.yaml, run 34002849643, attempts 1–10. Each is a distinct job execution of the full normal parallel suite with -race -count=1; the target ran exactly once without skips or retries. Source SHA 8897c2a881ce2a08a5faac4554e305399d5f972e; tested merge SHA cd4a15bdaa7dcbbb649e6ca4c8fb06e3673b26b2 throughout.

Attempt	Independent e2e job	Video / format (s)	Audio (s)	Outcome
1	101404737767	9.400000	9.269750	PASS
2	101405310678	8.900000	8.830000	PASS
3	101405973470	9.800000	9.664396	PASS
4	101406546784	9.600000	9.526396	PASS
5	101407211513	9.400000	9.185375	PASS
6	101407785465	9.600000	9.559938	PASS
7	101408374366	10.500000	10.433687	PASS
8	101408938930	9.900000	9.759292	PASS
9	101409567192	9.100000	8.994167	PASS
10	101410190324	9.300000	9.162354	PASS

Largest gap: 0.214625s against the unchanged 2s assertion. Each run has a replay-audio-34002849643-<attempt> artifact containing the MP4, stream/packet probe and service logs. Every video is verified as 1280×720. Minor Pulse timestamp warnings remain visible in those logs; no universal overloaded-4K recording guarantee is claimed.