Native AMD AMF tester build — polish batch (2026-07-24)

AMD-only tester build for Vibepollo PR Nonary#342, on the 1.18.1-stable.2 base.

Build

• Commit: b290c63 (branch feat/native-amd-amf-encoder)
• Installer SHA-256: f794d9bbea79e861b97cab45efd546669b61a6bdcec342adafaeb4ab310d319f
• Installed sunshine.exe SHA-256: a7d54ff43b30de76a51c9c86a9a71779e05c0d3ba75d6dd6ee2ec91c83fb7456
• CI: https://github.com/RamazanKara/Vibepollo/actions/runs/30048792848
• Unsigned tester build: Windows may show a publisher warning.

What changed since the 2026-07-23 tester

• Wire continuity: a submission dropped by driver backpressure no longer consumes a frame index, so Moonlight never sees a phantom gap and stops issuing spurious reference-invalidation under load. Drops are visible in the loop stats.
• Runtime bitrate rebuilds and healthy session ends detach the bounded teardown worker instead of stalling the encode thread 100–300 ms.
• Per-frame timestamp bookkeeping moved to a fixed ring — no heap traffic on the submit path.
• Input-surface reservations are generation-tagged; a driver-delayed release callback can never recycle a surface the VCN may still be reading.
• Encoder probing bounds now apply to AMF only; NVENC/QSV/software negotiation is untouched by this PR.
• Landing on software encoding after hardware probe failures is now loudly logged.

Encode latency (RX 7900 XT, AV1, 1920x1080 @ 120 FPS, 50 Mbps, 50 s per profile, zero AMF warnings/fatals, zero dropped submissions):

Profile	Encode median / p95 / max	Processing median
Queue 16 / Balanced	2.47 / 2.51 / 23.79 ms	2.68 ms
Queue 4 / Balanced	2.44 / 2.50 / 10.95 ms	2.64 ms
Queue 1 / Balanced	2.47 / 4.10 / 11.48 ms	2.70 ms
Queue 1 / Speed	2.28 / 2.34 / 11.95 ms	2.55 ms

The worst-case q16 spike from the previous run (103 ms) is gone; best-profile p95 improved from 2.56 ms to 2.34 ms.

Please report GPU, driver, codec, HDR state, selected AMD settings, reconnect/quit behavior, and stream stability.