E2: Retained measurement summary

Original summary SHA-256: c9f378be140406b160e2509e5723ba2ccd7403285831951f39952ff94fcb5803

Subsequent desktop validation rejected the fixed delay as a shared product change: 57.43 to 48.39 unique final FPS. The browser result below is retained as measured evidence.

Final-pixel timing comparison. 9 September 2026, Mountain Time. Apple M2 / 24 GiB, AC power.

Unique final-pixel delivery increased from 45.595 to 51.642 FPS in one sequential 20-second comparison. The difference was 6.047 FPS. The candidate cleared the 50 FPS average checkpoint in this run. The minimum complete interval remained near 42 FPS. This does not resolve delivery stalls.

| Measurement | Baseline | Delayed render |
|---|---:|---:|
| Unique final-pixel FPS | 45.595 | 51.642 |
| Minimum complete interval FPS | 41.920 | 41.929 |
| Observer callback mean, ms | 2.171 | 2.057 |
| Final valid reads | 1204 | 1205 |
| Final unique markers | 915 | 1037 |
| Final duplicate markers | 289 | 168 |
| Invalid markers | 0 | 0 |
| Reordered markers | 0 | 0 |
| Observer busy skips | 0 | 0 |
| Capture-window lower bound, FPS | 45.546 | 51.593 |
| Capture-window upper bound, FPS | 45.645 | 51.692 |

The baseline used the shipped media and render schedules. The candidate combined an upload immediately before render with a requested 7.625 ms delay. Product modules, native binary, captured engine, 1920×1080 GLSL and testPattern source were unchanged between windows. Each window used fresh browser sessions and a five-second warmup.

The existing WebGL2 observer copied marker strips from the final default framebuffer into pooled asynchronous readback buffers. It also sampled the acquisition texture. Final and acquisition totals agreed in both windows; these aggregate totals do not establish pairwise equality for every capture. All final markers were valid and ordered. Both full-camera-texture byte checks and deliberate one-pixel failure controls passed before timing.

Each endpoint had one pending readback per channel. The reader has no drain operation. Each stop abandoned one pending capture. Reported FPS counts completed marker reads. The capture-window bounds subtract the pending start capture and add the pending end capture. These bounds cover endpoint uncertainty, not experimental variance or observer overhead.

Native camera enqueue stayed near 60 FPS, with no native drops, rejections or failures. The baseline sender had six pressure drops; the candidate had none. The observer took about 2 ms per callback and can affect scheduling and GPU work. Earlier comparisons without pixel readback increased distinct browser-frame selection from about 45/s to 58.8–59.0/s in both orders. That larger number is not measured final-pixel FPS.

Marker age starts at sender marker insertion and ends at the observer capture timestamp. It excludes source rendering before marker insertion and is not physical display latency.

| Marker age, ms | Baseline | Delayed render |
|---|---:|---:|
| Median | 32.0 | 42.0 |
| 95th percentile | 49.0 | 58.0 |
| Maximum | 81.0 | 63.0 |

This is one browser pair. Desktop behavior, other content, long-run timing stability and sustained 60 FPS remain unverified. No product changes, commits or pushes occurred. All owned test sessions exited, and normal Sync was restored.