E4: Retained measurement summary

Original summary SHA-256: 851a17ad3a8c8e3ef2f6c004a4dfc558b535380af036f22cf7223ebff0e25c5c

# Camera enqueue and browser output

9 September 2026 · Apple M2 / 24 GiB · AC power

The camera queue accepted 1,799 frames during a 30.0135-second receiver measurement. Native counters recorded zero drops, rejected frames, or failures. The receiver observed 1,471 unique final markers. This places the measured gap after successful camera enqueue. It does not locate the gap within CMIO, Chrome capture, rendering, or observation.

| Measurement | Count | Count per second |
| --- | ---: | ---: |
| Camera queue accepted | 1,799 | 59.93969 |
| Receiver video callbacks | 1,478 | 49.24451 |
| Receiver unique final markers | 1,471 | 49.01128 |
| Receiver render callbacks | 1,801 | 60.00633 |


The receiver decoded 1,801 valid markers, including 330 repeats. It recorded no invalid or reordered markers. The minimum complete interval was 45.09 unique FPS. The sender recorded 1,799 sends and 2 admission backpressure drops at nearby sender endpoints.

The test used the existing production WebSocket → Sync Camera → Noisedeck media-effect path, with GLSL and testPattern() at 1920 × 1080. The served media module was the shipped b7c0cb0a baseline. The native binary was the previously validated d167ee4f build. The local allocation candidate and the rejected buffer-read experiment were excluded. A and B ran in separate Chrome processes. A five-second warmup preceded measurement.

Four existing getStats requests bracketed the receiver start and end snapshots. At each endpoint, the two native snapshots had identical counters. Both inner and outer bounds therefore contain exactly 1,799 accepted frames. The request/snapshot brackets lasted 9.713 ms at the start and 11.612 ms at the end. The diagnostic used the sender's existing serialized owner control session. It added no per-frame hooks or counters.

Only the camera publisher was enabled. In this configuration, native accepted means that CMSimpleQueueEnqueue succeeded. With multiple publishers, the aggregate accepted counter can instead mean that any publisher accepted the frame. Queue acceptance does not establish CMIO relay completion or browser presentation.

The final VideoFrame observer averaged 9.728 ms of elapsed submission time per capture. It adds work. This is one instrumented 30-second observation, not proof of uninstrumented throughput, a desktop result, full-frame pixel parity, or long-session stability. Video callback counts also do not establish unique frame identity. Neither GC nor GPU contention is established as the cause.

All source and cleanup checks passed. The installed Sync app was restored. No product change, commit, or push was made. Existing source, raw results, summaries, and the diagram are retained outside SCM.