E3: Retained measurement summary

Original summary SHA-256: 7c85b03367f6a255d40c2ffd030688dc1693e485d9a28b6232bfebd85c3a5fcf

Desktop render timing comparison. 9 September 2026, Mountain Time. Apple M2 / 24 GiB, AC power.

Reject the fixed delay as a shared product change. Unique final-pixel delivery fell from 57.426 to 48.394 FPS, a change of -9.032 FPS. The browser gain did not carry over to desktop.

| Measurement | Baseline | Delayed render |
|---|---:|---:|
| Unique final-pixel FPS | 57.426 | 48.394 |
| Minimum complete interval FPS | 53.995 | 43.996 |
| Observer callback mean, ms | 0.851 | 0.569 |
| Unique final markers | 1723 | 1452 |
| Duplicate final markers | 77 | 347 |
| Invalid markers | 0 | 0 |
| Reordered markers | 0 | 0 |
| Observer busy skips | 0 | 0 |
| Median marker age, ms | 34 | 27 |
| 95th percentile marker age, ms | 36 | 45 |
| Maximum marker age, ms | 37 | 77 |

Both windows used fresh desktop sender and receiver processes, a five-second warmup and thirty seconds of measurement. Both used current Noisedeck app and desktop entry sources, the captured engine, GLSL, 1920×1080, testPattern and the validated d167 native binary. Electron 42.3.0 / Chromium 148.0.7778.180 ran the actual source entry with frozen shell dependencies. This is not packaged-release certification.

The shared timing and upload helpers were byte-identical to the browser experiment. The candidate combined a pre-render media upload with a requested 7.625 ms render delay. Actual mean timer delay was 9.660 ms. The candidate added no native-desktop-only implementation.

The existing WebGL2 reader decoded markers from actual final rendered pixels. All reads were valid and ordered, with zero busy skips. Full-camera-texture byte comparisons and one-pixel failure controls passed before both windows. One capture per channel remained pending at each endpoint and was abandoned on stop. Candidate capture-window bounds were 48.360–48.427 FPS; these cover endpoint uncertainty, not experimental variance or reader overhead.

The baseline sender had zero pressure drops. The candidate had four. Neither sender reported failures. Median marker age improved in the candidate, but its 95th percentile and maximum increased. Marker age runs from sender marker insertion to observer capture; it is not physical display latency.

The first launch failed before measurement because the older harness referenced a removed startup-preview property. That failed attempt and its cleanup evidence are retained in desktop-runs. The retry used the current close/readiness contract, matching the browser harness. No product file changed. Twenty lifecycle, ownership and measurement checks passed before launch. All owned sessions exited, both source receipts stayed stable, and normal Sync was restored.

A constant delay is not a reliable shared optimization. A further timing experiment must use observed frame freshness and allow zero delay when the existing cadence is already effective. No fixed-delay integration, commit or push is justified by these results.