test(core): fix flaky calibration-flash e2e assertion

[jbrecht]

Root cause

The e2e gate intermittently failed at assert.ok(result.videoClockOffsetMs > 0, 'calibration flash was detected') (~1 run in 3). It is not a detection-robustness problem — #46's range tolerance is fine.

The flash (#ff00ff) is painted right after clock.zero(). recordVideo starts at context creation, but VP8 coalesces the identical pre-flash about:blank frames, so the first encoded frame is sometimes the flash itself at pts_time:0. Detection then correctly returns offset 0 — and the assertion's > 0 treated that valid detection as a failure. The offset 0 is also the correct value (the flash genuinely is at clock zero).

I confirmed this against the raw recording from a failed run: chroma at frame 0 = UAVG 201 / VAVG 219 (clean magenta), pts_time 0, so detectFlashOffsetMs returns 0.

detected offset: 0
old assertion (> 0):   FAIL   <- the flake
new assertion (!=null): PASS

Fix (test-only)

• e2e: re-detect the flash on the raw recording and assert it was found (!== null) and that videoClockOffsetMs equals that detected offset — instead of asserting the offset is positive. A genuine miss (null) still fails the gate, so the original intent (catch real misses) is preserved.
• Unit: add a parseFlashFromMetadata regression test pinning that a first-frame magenta flash parses to 0, not null.

Production code is unchanged — the pipeline already distinguished null (warn, fall back) from a found offset, and 0 is the right answer when the flash is on frame 0.

Verification

• 159 core unit tests pass; pnpm build clean.
• Full e2e integration gate run: exit 0.
• Proven against the exact recording that previously failed.

Note

This adds an ## Unreleased CHANGELOG entry; #53 does too, so whichever merges second needs a trivial one-line merge to combine the bullets under a single ## Unreleased. Both should land before the next version bump.

🤖 Generated with Claude Code