E5: Retained measurement summary

Original summary SHA-256: 338d0141372ad35bc7905135bf5711e5db8f93ccbbfa50211f4a114568f36197

# Production camera CPU profile

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

A 30-second profile of the shipped WebSocket → Sync Camera → Noisedeck media-effect path identifies sender readback as the largest named active operation. The final-pixel observer was disabled. This run does not measure unique delivered FPS.

| Sampled operation | Sender A, ms | Receiver B, ms |
| --- | ---: | ---: |
| getBufferSubData | 6698.590 | 0.000 |
| Garbage collector | 15.782 | 45.847 |
| Idle | 18705.182 | 28049.162 |

Sender samples also include 1336.638 ms in WebSocket send and 665.566 ms in frame encoding. CDP sampled every 2 ms. Time attributed to an API can include waiting and preemption; it is not a GPU timer or an exact thread CPU measurement. This short profile does not exclude GC as a cause of individual stalls.

The served media module is the shipped b7c0cb0a baseline. The local allocation candidate was excluded. Source checks and resource cleanup passed, and the installed Sync app was restored. Raw profiles and source hashes are retained beside this report.

The subsequent buffer-read placement experiment reduced mean elapsed read time by 8.34%, but unique delivery fell from 49.79 to 47.74 FPS. It was stopped after one pair. See buffer-read placement (retained local evidence). A separate camera-counter measurement (retained local evidence) then recorded 59.94 accepted camera enqueues/s and 49.01 unique final markers/s. Neither diagnostic introduced a product change.