#802: Prepare CPU sources before Program and isolate optional exports

[iamfatness]

#802: cold 1080p background creation/upload could exceed the Program deadline, and an expired Ready queue head could amplify the original loss. Windows module composition now enables existing completed-source preparation and isolated optional exports, with explicit rollback overrides. The Program buffer retires an expired Ready head before waiting for the same following slot, preserving the original loss count and configured buffer depth.

The maintained real-meeting harness requires every selected raw feed, exercises cold media identities, and fails acceptance on Program losses or independently judged OS-camera pixel discontinuity. Receiver exit zero is insufficient. Before/after counter baselines now align with the fixed receiver warm-up.

Validation:

• Release native build, eight hardware Program-buffer tests and ten I420/source-preparation tests passed. Reverting the queue repair fails the real NV12 regression.
• Full local suite: 1,566 passed; two webcam-framer failures also reproduced in the unchanged baseline and remain tracked in Built-in OH framer overlay for webcam output only #839.
• Seven receiver-judge tests passed, including intentional loss, invalid pixels, wrong cadence and incomplete capture.
• All 17 CI/analysis checks completed without failure on integrated revision 456d762, including the merged docs: reconcile shipped UI/UX plans and installed acceptance (#848) #858/Fix idle recording Stop reporting phantom finalization (#859) #860 changes.
• Installed-layout three-minute live smoke passed Program and OS-camera continuity through three cold 1080p loads.
• Thirty-minute qualification FAILED despite no SDK crash/disconnect: eight real feeds (all 1080p at the end), all 20 backgrounds admitted, 201 Program underruns and 132 overflows. OS-camera pixels over its slightly longer window contained 242 repeated and 173 missing identities, maximum interval 49.632 ms.

Before any Program loss, independent webcam read contention already caused 37 repeats/missing identities. Later Program failures begin with an empty queue and also include late Preparing/Ready heads. Program source uploads/creations remain zero, but producer timing is not qualified. Stage windows locate long ingest/monitor handoff intervals without establishing the owner or cause of every miss. The next run needs existing Zoom handoff timing and exact bounded delivery trace attribution; reader contention needs a matched off/on fault comparison with the loaded DLL pinned.

The owner requested integrating all open PRs on October 11 after the failed-soak report. This merges incremental improvements and retained failure evidence; it does not waive zero-miss acceptance. Acceptance for #802/#517 remains outstanding, including physical display, A/V, recording/streaming, device recovery or fleet acceptance. The candidate native executable was substituted into the registered installed owner directory with an original backup; shell and camera DLL were unchanged. No beta was published. Detailed evidence and limits: docs/qa/cpu-preparation-production-2026-10-11.md.