Session-2 (continuation of d614462 INCONCLUSIVE) executed disambiguation
plan and converged on a definitive verdict. Three independent observations
ruled out ALL architectural-failure hypotheses:
Step A: race-tolerant offscreen target attach (committed separately;
enabled visibility into the offscreen recorder + remux pipeline).
Step B: pre-kill and post-kill segment-count probes via the existing
`__mokoshOffscreenQuery 'get-segment-count'` bridge op (no new
test-only symbols introduced; FORBIDDEN_HOOK_STRINGS inventory
unchanged at 12 entries). Observed segments.length transition:
POST-PRIME=0 → PRE-KILL=3 → POST-KILL=3
Segments structurally survive the SW kill (offscreen still responds
to bridge query post-kill). Hypothesis A (architectural RAM loss
across SW termination) REFUTED.
Step C: SPIKE_SKIP_SW_KILL=1 env-var mode skips worker.close(). The
resulting videoSize is IDENTICAL to the canonical run (8505 bytes).
Hypothesis C (CDP-induced offscreen collateral teardown) REFUTED.
Since SW was not killed, its console listener stayed connected,
exposing the full Remux pipeline output:
[SW:Remux] Segment ts=1: 0 frames, duration=0ms, trackInfo=320x180
[SW:Remux] Segment ts=2: 0 frames, duration=0ms, trackInfo=320x180
[SW:Remux] Segment ts=3: 0 frames, duration=0ms, trackInfo=320x180
[SW:Remux] Remux complete: 0 frames, total timeline=0ms, output=8505 bytes
Each segment Blob has a valid track header (PixelWidth/Height parsed
successfully) but ZERO VP9 frames. Hypothesis B (canvas-captureStream
throttling in headless idle) CONFIRMED.
VERDICT: REFUTED-architecture (canvas-captureStream issue).
The architecture (offscreen-RAM `segments: Blob[] = []`) works
correctly; the spike's test methodology is invalid. The
`installFakeDisplayMedia` synthetic stream (canvas.captureStream(30)
on a hidden -9999px-offset 320x180 canvas) cannot sustain frame
production during a 5-min headless idle window despite the
`setInterval(drawFrame, 33ms)` belt-and-suspenders mitigation. This
matches the documented Chromium throttling of MediaRecorder on
invisible-canvas sources (Chrome bug 653548; auto-throttled-screen-capture
design doc; sendrec.eu blog "Why Canvas Breaks Your Screen Recorder").
ROUTING RECOMMENDATION (out of scope for this debug session):
- Do NOT proceed with the IndexedDB persistence plan-fix proposed by
Plan 04-04 SUMMARY. The plan-fix would NOT close SC #1 because the
spike would STILL produce 8505 bytes after IDB lands — the failure
is in the test's fake stream, not in segment persistence.
- Open a new plan slot (likely Plan 04-08 or a Phase 5 plan) that
reframes SC #1 verification methodology. Options:
(a) real getDisplayMedia in non-headless Puppeteer with
--auto-select-desktop-capture-source;
(b) video-file-backed MediaStream source (HTMLVideoElement
playing a bundled WebM) — bypasses canvas-captureStream
throttling entirely;
(c) reduce SC #1 wall-clock idle threshold to a value short
enough that canvas-captureStream survives (e.g., 30s) AND
add a separate manual operator-empirical test for 5-min.
ROADMAP SC #1 status: REMAINS OPEN. The architecture is sound; the
empirical verification gate is broken. Plan 04-04 SUMMARY's
characterization ("spike FAILED → architectural plan-fix needed") is
TECHNICALLY CORRECT on the first clause but INCORRECT on the second —
the spike's failure mode is in test infrastructure, not in production
code.
Files in this commit:
- tests/uat/spike-a33-sw-persistence.ts: added probeSegmentCount
helper using existing __mokoshOffscreenQuery bridge op; 3
checkpoints (POST-PRIME / PRE-KILL / POST-KILL); SPIKE_SKIP_SW_KILL=1
env-var skips worker.close() for Step C disambiguation.
- .planning/debug/sw-offscreen-persistence-investigation-session-2.md:
NEW session-2 debug note documenting full evidence trail + verdict
derivation + routing recommendation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mark
4ea1bbb7a8