Live validation on RTX 4090, 1920x1080/60 Program plus YouTube GPU stream, switching between the saved Tiles and Active Speaker scenes. Candidate 85ecaca (main plus Program-buffer startup preallocation and diagnostic logging).
The recording encoder queue continued dropping video after startup: 927 dropped video items at 43.133 seconds, 1,940 at 95.883 seconds, and 1,946 after finalization. Dropped audio remained at 54. The explicit startup video-drop counter was separately 158. Final per-source ISO dropped counts ranged from 51 to 704, with eight ISO outputs present; Program plus all ISOs reported completed and files were finalized. Program buffer remained at zero underruns and zero deadline misses throughout this recording.
A preceding main-plus-logging candidate (607810c) also dropped queue items when recording began (298 video / 49 audio), but those counters stopped growing in that shorter seven-ISO observation. This is not yet attributed to the Program-buffer startup change; compare identical source counts, rates and scene workloads before asserting regression.
Investigate the serial recording writer/ISO encode throughput and per-source queue fairness without lowering capture resolution or hiding dropped frames. Also verify whether recording health should expose the sustained loss instead of reporting healthy/completed.
Local evidence: CoreVideoPro-merge-prs/artifacts/live-2026-09-14/{warm-recording,warm-finalized,baseline-recording,baseline-finalized}.json. Release/manual QA caveat; not a passing recording performance result.
Follow-up: a second recording on candidate 85ecaca held the eight-person Tiles wall steady with no Takes. Cumulative recording drop counters grew from 3,222 video / 108 audio at 17.066 seconds to 8,482 video / 124 audio at 99.083 seconds. Thus loss continues without scene switching. The concurrent Program buffer remained at zero underruns and deadline misses, and all six source/audio live checks passed. Compare deltas within each recording: these counters are cumulative across recording sessions in the native process. Evidence: steady-recording-start.json and steady-recording-end.json in the same local evidence directory.
Live validation on RTX 4090, 1920x1080/60 Program plus YouTube GPU stream, switching between the saved Tiles and Active Speaker scenes. Candidate 85ecaca (main plus Program-buffer startup preallocation and diagnostic logging).
The recording encoder queue continued dropping video after startup: 927 dropped video items at 43.133 seconds, 1,940 at 95.883 seconds, and 1,946 after finalization. Dropped audio remained at 54. The explicit startup video-drop counter was separately 158. Final per-source ISO dropped counts ranged from 51 to 704, with eight ISO outputs present; Program plus all ISOs reported completed and files were finalized. Program buffer remained at zero underruns and zero deadline misses throughout this recording.
A preceding main-plus-logging candidate (607810c) also dropped queue items when recording began (298 video / 49 audio), but those counters stopped growing in that shorter seven-ISO observation. This is not yet attributed to the Program-buffer startup change; compare identical source counts, rates and scene workloads before asserting regression.
Investigate the serial recording writer/ISO encode throughput and per-source queue fairness without lowering capture resolution or hiding dropped frames. Also verify whether recording health should expose the sustained loss instead of reporting healthy/completed.
Local evidence: CoreVideoPro-merge-prs/artifacts/live-2026-09-14/{warm-recording,warm-finalized,baseline-recording,baseline-finalized}.json. Release/manual QA caveat; not a passing recording performance result.
Follow-up: a second recording on candidate 85ecaca held the eight-person Tiles wall steady with no Takes. Cumulative recording drop counters grew from 3,222 video / 108 audio at 17.066 seconds to 8,482 video / 124 audio at 99.083 seconds. Thus loss continues without scene switching. The concurrent Program buffer remained at zero underruns and deadline misses, and all six source/audio live checks passed. Compare deltas within each recording: these counters are cumulative across recording sessions in the native process. Evidence: steady-recording-start.json and steady-recording-end.json in the same local evidence directory.