Current status: rejected and removed before this plan. The historical hold statement below is superseded.

E6: Retained measurement summary

Original summary SHA-256: 1dc0fae37b5990e948553effcf8063ad15c72787b7c818aa0069adad74d399ad

# Media allocation reuse measurements

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

Hold locally. Mean unique delivery was lower in all four comparisons; the reversed web followup repeated the decline. Allocation savings do not establish a delivered-performance improvement.

The shared media manager reuses immutable upload options and the uploaded-size callback for each media object and step index. Map traversal uses a stable visitor. Upload cadence and the existing synchronous and asynchronous completion behavior remain unchanged.

In an isolated Node/V8 profile, sampled allocations attributed to the manager decreased from 35.48 MB to 0.20 MB across 200,000 updates. Every upload continued. This is sampled component evidence, not whole-app allocation or FPS evidence.

| Run | Unique FPS | Minimum interval FPS | Uploads | Elapsed ms/update |
| --- | ---: | ---: | ---: | ---: |
| Web B1 | 48.36 | 40.00 | 1800 | 0.279 |
| Web C1 | 48.21 | 40.95 | 1801 | 0.254 |
| Web C2 | 49.42 | 44.00 | 1800 | 0.270 |
| Web B2 | 51.44 | 44.51 | 1801 | 0.284 |
| Desk B | 56.61 | 53.03 | 1801 | 0.283 |
| Desk C | 56.19 | 52.94 | 1800 | 0.276 |
| Web C3 | 49.22 | 44.09 | 1800 | 0.250 |
| Web B3 | 49.52 | 45.00 | 1800 | 0.249 |


Mean delivery changes were -0.15, -2.01, -0.42, and -0.30 FPS. The final reversed web pair also showed no elapsed-loop time improvement (0.250 ms candidate versus 0.249 ms baseline). These short runs do not prove causation, but they do not meet the confidence requirement for release. Recurring stalls remain unresolved.

All eight delivery measurements used testPattern() through the actual WebSocket sender, Sync Camera, and Noisedeck media effect at 1920×1080. Web order was baseline, candidate, candidate, baseline. Desktop order was baseline, candidate. A final web pair used candidate, baseline to check the delivery differences. Each run used fresh owned processes, the same native binary, a five-second warmup, and a 30-second measurement. Frozen media modules isolate the change against b7c0cb0a. Raw paths, hashes, timing distributions, and interval results are recorded in summary.json.

The allocation profile used Node 24.7.0 and a 256-byte sampling interval, including objects collected by major and minor garbage collection. Each fresh process warmed up for 20,000 updates, collected garbage, and then profiled 200,000 updates. The fixture supplied preallocated dimensions and renderer results to isolate the manager. Module imports were rewritten to point to the unchanged app dependencies.

Eleven focused tests cover steady uploads, metadata changes, replaced state and pipelines, async completion, source reindexing with pending uploads, options reuse, media filtering, nested updates, Map mutation, and thrown errors. Targeted lint passed. Independent source review found no release blockers. Broader validation and any publication are recorded separately in validation.json and shipping.json.

Allocation sampling isolates the manager in Node/V8 with a preallocated renderer result; it is not browser or whole-app allocation accounting.

The older async-upload helper still allocates Promise handlers. The engine and other frame-path allocations remain.

Elapsed media-loop time includes preemption and is not thread CPU time.

The VideoFrame observer adds work. Thirty-second runs cannot establish sustained 60 FPS or long-term stability.

Web and desktop use different frozen app/browser fixtures. Desktop runs use the real source entry, not packaged builds.

Markers establish observed identity order, not full-frame pixel equality or physical display cadence.

The full local Node suite recorded 861 passes, one failure, and one skip. The failure imports the absent sibling Noisemaker extrude fixture. No hosted validation was triggered because no commit was pushed.

Resource cleanup completed for every run, source checks passed, and the installed Sync app was restored. This report, diagram, fixtures, and raw evidence remain outside source control.