feat(amf): add native AMD AMF encoder

[RamazanKara]

Summary

Native AMD AMF encoder for Windows (D3D11, zero-copy), replacing the FFmpeg-based amdvce path. The PR contains AMD encoder code, AMD configuration/UI/docs, build integration, and AMD-focused tests, based on 1.18.1-stable.2 (6fe86ba4).

The native path supports H.264, HEVC, and AV1 with SDR/HDR where the selected driver mode supports it, dynamic bitrate, IDR and reference-frame invalidation, bounded queue/backpressure handling, and ordered startup/teardown. amdvce selects the native encoder; amdvce_legacy is the explicit rollback (exposed in the encoder dropdown and the docs Choices table). Automatic probing never silently falls back to the legacy path.

Design contracts

• Verified property application. Every requested AMF property is set and read back; a rejected or post-Init-drifted value fails encoder creation instead of silently streaming with a different configuration. The client reference-frame limit and the H.264 intra-refresh two-reference floor are computed by one shared helper (effective_reference_frame_limit) used identically at configure time and post-Init verification, so the encoder never rejects its own raise.
• FFmpeg-parity submit semantics. Only AMF_INPUT_FULL/AMF_DECODER_NO_FREE_SURFACES retry a submission. AMF_NEED_MORE_INPUT counts as consumed input with deferred output, which is what PreAnalysis lookahead priming returns while the pipeline fills.
• Bounded everything. Submit retries are capped (20 attempts, 2 s wall clock), the input ring is capped at 32 plus transit, the surface pool grows lazily one texture at a time, and the teardown fence and the driver-destruction watchdog each get an independent 5 s budget. A genuinely wedged driver quarantines native AMF for the host's lifetime rather than hanging the process.
• Wire continuity under load. A submission dropped by driver backpressure never consumes a frame index; the next capture reuses it, so the RTP frameIndex sequence stays contiguous and the client is never tricked into reference-invalidation round-trips for frames that never shipped.
• Generation-tagged surfaces. Every input-surface reservation is stamped with a generation and each AMF wrapper carries the value it was created under; a driver-delayed release callback from an old wrapper cannot recycle a surface the VCN may still be reading.
• Recoverable runtime fatals. A TDR, sustained submit backpressure, or an output stall rebuilds the same native session, bounded to three attempts without a stable run — never a silent legacy fallback, and fail-closed while quarantined.
• Curated configuration surface. The UI and docs expose only choices a user can reason about: usage, rate control, QVBR level, quality preset, VBAQ, PreAnalysis, HRD, H.264 coder, and the AV1 content/latency modes. Deep driver knobs (input queue depth, LTR slots, Smart Access Video, the LOWLATENCY_MODE override, high-motion boost) are handled by the encoder's own policy and remain available only as undocumented expert config keys.
• Cold-path work stays off the hot path. Bitstream extraction and input-pool texture allocation run outside the encoder state lock, the output pump waits interruptibly instead of sleeping in fixed quanta, and the encode loop performs one RTTI lookup per session rather than per frame, per-frame timestamp bookkeeping lives in a fixed ring instead of a heap-allocating map, and healthy session ends or runtime bitrate rebuilds detach their bounded teardown worker instead of stalling the encode thread.

Xbox intra-refresh

enableIntraRefresh == 1 follows the NVENC behavior with a 300-frame refresh cycle:

• H.264 refreshes 16×16 macroblocks across 299 refresh frames.
• HEVC refreshes 64×64 CTBs across 299 refresh frames.
• AV1 uses continuous intra-refresh mode with a 300-frame cycle.
• H.264 avoids conflicting LTR configuration; clients that send a single reference (or no limit at all) receive the required two-reference floor.
• Requested properties are read back; driver rejection fails the native session rather than silently ignoring the Xbox request.

The mapping is covered by unit tests and exact-head CI. Desktop Moonlight does not send the Xbox request, so confirmation on actual Xbox hardware is still welcome.

Quality presets

Selection	H.264	HEVC	AV1
Speed	1	10	100
Balanced	0	5	70
Quality	2	0	30

Auto leaves the AMF quality property unset. Explicit presets are applied after usage and read back; a rejected or changed value fails the native session instead of silently selecting a different preset.

Also included

• fix(nvhttp): PIN-stdin pairing (-0) built its certificate response and then fell through to a session lookup on a moved-from uniqueid, replacing the response with 400 Invalid uniqueid — stdin-driven pairing could never complete. Included here because the automated AMD latency validation pairs through exactly this path.

Validation

• Exact-head Windows CI, native AMF lifecycle tests, packaging, and installer creation: workflow 30048792848.
• Standalone AMF lifecycle suite: 25/25 checks (the same file runs under GTest on Arch/macOS CI). Frontend config-schema tests: 14/14.
• Closed-loop hardware run at this head: RX 7900 XT, AMF runtime 1.5.2.0, Moonlight 6.1.0, AV1, 1920×1080 @ 120 FPS, 50 Mbps, deterministic full-screen workload, 50 s per profile, installed sunshine.exe SHA-256 a7d54ff43b30de76a51c9c86a9a71779e05c0d3ba75d6dd6ee2ec91c83fb7456. Every profile established a real client connection, delivered video, and completed with zero AMF warnings and zero fatals.

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

The suite's recommendation rule (lowest encode p95, differences ≤0.05 ms treated as noise, q4 baseline preferred on ties) selects Queue 1 / Speed. Zero dropped submissions were recorded across every profile, and encode medians hold a 2.3–2.5 ms band run to run.

A tester build for this head is available: native-amf-polish-20260724 (unsigned; Windows may show a publisher warning). More AMD and Xbox reports are welcome — please include GPU, driver, codec, HDR state, selected AMD settings, reconnect/quit behavior, and whether the stream remained stable.

[RunAway189]

I've quickly tested the latest build and the streaming got broken on AV1/HEVC/H264. The interesting thing is that Artemis told me my GPU wasn't capable of HDR when I started the stream. I only got a purple screen when the stream started. Sorry for not providing the log but I couldn't since I was in a hurry.