perf(producer): --gpu renders encode while frames are captured - #6631

What

Multi-worker mp4 renders with --gpu now encode while frames are still being captured, as software encodes already do. The encode left after capture drops to about zero.

Why

#5952 added encode-while-capturing for multi-worker disk captures but limited it to software libx264. With --gpu, the hardware encoder waited until every frame was captured and then encoded them all serially. That made --gpu slower than the default on a fast Mac: VideoToolbox spent 6.1 s after a 42 s film's capture finished, while overlapped libx264 had 0.1 s left.

The exclusion's stated reason was that overlap reuses the streaming encoder, which had to build the same arguments as the disk encoder. For the GPU encoders it already does: both builders emit identical arguments from -c:v on, for videotoolbox, nvenc, qsv, vaapi and amf. A new test pins that.

Why under 100 lines

One lever: the GPU encoder joins the existing encode-while-capturing path, which already builds the same encoder args. The change is dropping one exclusion, with a test that pins the args match.

Related work

Refs #5952.

How

canEncodeWhileCapturing drops its useGpu !== true condition. Everything else is unchanged: h264 from JPEG frames, not HDR, not chunked, not png/gif, and capture must leave a core free. If the overlapped encode fails, the frames are still encoded after capture as before.

Test plan

• Unit tests added: encodeWhileCapturing.test.ts (GPU renders are eligible) and deliveryEncoderOptions.test.ts (disk and streaming builders emit the same GPU arguments for all five GPU encoders).
• Fails without the fix: encodeWhileCapturing.test.ts > runs only when the capture workers leave a core free for the encoder (1 failed / 4 passed with only the tests applied). The argument-parity test passes on main too; it guards the invariant this change relies on.
• tsc --noEmit (engine, producer) exit 0; engine deliveryEncoderOptions.test.ts 9 passed; producer encodeWhileCapturing.test.ts 5 passed; oxfmt and oxlint clean.
• Manual testing performed: before/after on the real CLI (below).
• Documentation updated (not applicable)
• Comments follow CONTRIBUTING.md "Comments"

Before / after

hyperframes render --gpu, published [email protected] vs the same bundle with only this change, Apple M4 Pro (14 cores, 24 GB, VideoToolbox), repo fixtures, interleaved, medians. wall - compile drops compile, which is network-bound for these fixtures.

Output: VideoToolbox output is not byte-stable on its own (two runs of main on the first fixture gave two different files), so this compares timing only. Encoder, arguments and settings are unchanged.