refactor(producer): extract audioStage from executeRenderJob

[jrusso1020]

Stacked on #725 (PR 1.4). Rebase to `main` once that merges.

What

Phase 1 PR 1.5 of the distributed-render refactor. Moves the audio mixing sub-stage of executeRenderJob into packages/producer/src/services/render/stages/audioStage.ts. Trivial extraction — the body is a thin wrapper around processCompositionAudio.

Source range moved: ~25 lines between the "── Stage 3: Audio processing ──" comment and the start of "── Stage 4: Frame capture ──".

Why

Continues the Phase 1 mechanical extraction. Phase 1 ships zero new functionality.

How

• New file audioStage.ts exports runAudioStage(input) → AudioStageResult returning { audioOutputPath, hasAudio, audioProcessMs }.
• Sequencer calls the stage at the same code point with the same inputs.
• Removes the now-unused processCompositionAudio import from renderOrchestrator.ts (oxlint flagged it).

Preserved invariants

• audioOutputPath is still join(workDir, "audio.aac") regardless of whether any audio was produced.
• hasAudio reflects audioResult.success from processCompositionAudio; it is false when there are no audio elements (call is skipped) and also when the mixer returns success: false.
• perfStages.audioProcessMs is set whether or not the mixer ran, at the same code point.
• The 20% "Processing audio tracks" updateJobStatus fires at the same point.

Test plan

• bunx oxlint + bunx oxfmt --check — clean.
• bun run --filter @hyperframes/producer typecheck + build — clean.
• bun test packages/producer/src/services/ — 176 pass, same single pre-existing failure unrelated to this PR.
• docker run hyperframes-producer:test --sequential font-variant-numeric many-cuts variables-prod — 3/3 PASS, audio correlations 1.000 / 0.994 / 0.975 (identical to prior PRs in the stack; font-variant-numeric and many-cuts exercise the audio path).
• Full regression matrix on CI via the regression workflow.

🤖 Generated with Claude Code

[jrusso1020]

• refactor(producer): move updateJobStatus to render/shared.ts #737
• refactor(producer): document executeRenderJob as a thin sequencer #735
• refactor(producer): extract encodeStage and assembleStage #734
• refactor(producer): extract captureHdrStage (HDR / shader-transition path) #733
• refactor(producer): extract captureStreamingStage (single-machine fusion) #731
• refactor(producer): extract captureStage (SDR disk path) #730
• refactor(producer): extract audioStage from executeRenderJob #726 👈 (View in Graphite)
• refactor(producer): extract extractVideosStage + add materializeSymlinks param #725
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[vanceingalls]

Verdict: approve.

Trivial extraction. audioStage is a thin wrapper around processCompositionAudio. Body lifted byte-clean: same audioOutputPath = join(workDir, "audio.aac") even when audio is empty, same hasAudio = audioResult.success, same perfStages.audioProcessMs timer semantics (set whether or not the call ran).

No-op concerns are also handled cleanly — the stage's timer covers both branches via stage3Start → Date.now() - stage3Start, identical to the pre-extraction code.

Nits:

• audioStage.ts:18 — job: RenderJob is declared on AudioStageInput but never read in the body. #735 actually removes this, so it's self-cleaning, but worth flagging that future stage inputs should be added only when consumed (foundation review feedback).
• Same runtime cycle pattern as the other stages — audioStage.ts:3 imports RenderJob from ../../renderOrchestrator.js (type-only, but the import line is still there). Type-only imports are erased so this is harmless at runtime; #735 will tighten it.

Praise: smallest stage in the stack, smallest diff, easiest to verify. The pattern of explicitly preserving "the timer is still set so the perf summary stays consistent across renders" in the JSDoc is exactly the discipline this stack needs.

— Vai

The base branch was changed.

[miguel-heygen]

Clean mechanical extraction — no behavior changes, no introduced bugs. Verified imports, error handling, and cleanup invariants are preserved. LGTM. — Magi