Offline upstream validation

The combined changes cover input/certificate hardening, explicit pairing approval, encoder/capture lifecycle, conditional probe reuse, rational timing and PipeWire negotiation, nanors runtime dispatch, optional VA-API controls, stable KMS names, and isolated multiseat input/runtime images. Production multiseat activation and worker adapters remain unwired. Runtime dependency/provider success is not game streaming acceptance.

Private capture regression

Build the full native video suite with CUDA support, then run on an NVIDIA host with labwc, wlr-randr, and vkcube installed:

python3 tools/tests/video_lifecycle_harness.py \
  --test-binary build/tests/test_polaris_video \
  --render-device /dev/dri/renderD128 \
  --evidence build/private-video-evidence

The evidence directory must be new. The harness creates its own home and Wayland runtime, starts an animated private compositor, and exercises nine capture cycles: abort before the first frame, repeated reconnects, explicit 59.940 FPS, a 60 FPS stream on a 120 Hz output, SHM fallback, and successful DMA-BUF retry. It asserts the observed transport, checks render-descriptor stability after warmup, and retains an encoded packet through encoder/compositor teardown to check ownership. It starts no streaming server, input authority, launcher, or game.

A process-local child subreaper and retained pidfds provide bounded cleanup, including reparented descendants and termination signals. The receipt records forced cleanup and remaining resources; either prevents an acceptance pass. Private logs/configuration remain for inspection. The test skips in ordinary native runs unless explicitly enabled through this harness.

Measured comparison

Three interleaved runs compared the same test against lifecycle baseline 7cdc9c90e712f1b5113e50dc69212897ca0289a4 and the combined implementation on an RTX 4090, NVIDIA 610.57.04, Core i9-14900K, using GCC 16, CUDA 13.2, Debug builds, 1280x720 H.264 NVENC at 10,000 kbps, and the same private animated compositor. These are capture/encoder measurements; they exclude network and client latency.

Warm reconnect median	Baseline	Combined implementation
Probe call	51.74 ms	5.25 ms
Capture-to-encode completion	1.460 ms	1.473 ms
Capture stop	41.57 ms	41.47 ms
Delivered cadence	60.02 FPS	60.14 FPS
Resident memory	305.3 MiB	307.9 MiB

Each nine-cycle run performed nine full probes on baseline and five full probes plus four reuse hits on the implementation. Output and capture-transport changes required full probing. Warm probe time fell about 90%; the encode/stop/cadence results do not establish a performance improvement. Short frame samples do not resolve the 0.060 FPS difference between 59.940 and 60. Separate private PipeWire producer measurements cover rational cadence directly.

The baseline leaked three render descriptors on each DMA-BUF reconnect. Its render-descriptor count grew from 5 to 26 over the nine cycles; the implementation held at 2. The baseline deliberately fails that regression while completing all frame/transport checks. The implementation retains additional bounded provider proofs: warm total descriptors were 237 versus 80, and final median RSS was 322.0 MiB versus 315.3 MiB. Probe reuse trades retained provider references and memory for fewer live probes; it is not a memory optimization.

The independent nanors benchmark measured a 22.8–33.0% reduction in parity encode CPU time, with identical parity, and a small codec-creation cost increase. See fec-runtime-dispatch.md for its separate method and limitations. No end-to-end game-streaming gain is claimed.

Remaining hardware boundaries

VA-API controls preserve automatic defaults and have driver/codec option tests; AMD/Intel hardware evidence is required before promoting new defaults. KMS qualified and numeric selection have physical capture evidence; disconnect and ambiguous binding behavior also have native DRM boundary regressions. The four runtime profiles have isolated provider/input evidence, including independent workers and cleanup. Worker-local encoding, launcher management, production media routing, seat-aware status, and real concurrent game streams remain the next milestone.