perf(producer): phase-level render instrumentation

[jrusso1020]

What

Adds phase-level observability to the render pipeline — no behavior change, all additive fields.

• ExtractionResult.phaseBreakdown: splits the existing durationMs into resolveMs, probeMs, hdrPreflightMs, vfrPreflightMs, extractMs, plus HDR/VFR preflight counts.
• InjectorCacheStats: optional out-param on createVideoFrameInjector that counts LRU hits, misses, in-flight-coalesced reads, and peak entries.
• RenderPerfSummary gets videoExtractBreakdown, injectorStats, and tmpPeakBytes.

Why

This is PR 1 of a 5-PR stack distilled from `hyperframes-notes/producer-render-architecture-review-2026-04-21.md`. The doc flags specific bottlenecks (HDR/VFR preflight cost, injector disk reads, /tmp pressure) but explicitly says every claim is a hypothesis until profiled on real compositions. Downstream PRs — segment-scoped HDR preflight, extraction cache, WebP unified format, gated hwaccel — need these numbers to validate they actually helped.

Concretely, after this lands the validation plan from the doc (§ Validation plan) becomes runnable: the three reference compositions (short SDR, medium alpha-mix, long 4K) each emit a structured `perfSummary` with per-phase timings that can be diffed across PRs.

How

• Extractor (`videoFrameExtractor.ts`): adds a `phaseBreakdown` object populated in-place as each phase runs. HDR and VFR re-encodes are wrapped in try/finally so failed attempts still account for their wall-time. Phase 3 is timed as a single wall-clock span since it's already parallel across inputs.
• Injector (`videoFrameInjector.ts`): the frame-dataURI LRU accepts an optional `InjectorCacheStats` that it mutates in-place. When absent, behavior is byte-for-byte identical to before (single sentinel check at each counter site). Exported `createEmptyInjectorCacheStats` and a test-only `__createFrameDataUriCacheForTests` so the LRU semantics can be verified without a running Chrome.
• Orchestrator (`renderOrchestrator.ts`): creates one shared `injectorCacheStats` and threads it into the three sequential-capture injector constructions. Parallel-worker injectors (built via factory) run in separate worker processes and can't mutate the host stats object, so those paths leave `injectorStats` undefined. Samples `workDir` size just before cleanup as a lower bound on peak disk usage (by then the extracted-frame dir, captured-frame dir, and intermediate audio/video artifacts all coexist).

What's deliberately not in this PR

• Browser launch timing (`browserLaunchMs`) and first-frame latency — would require threading timers through `executeParallelCapture`/worker boundaries. The existing `browserProbeMs` + `captureMs` stages cover this at lower resolution and are enough to sequence the downstream PRs.
• Mid-render `/tmp` watermark polling — end-of-render sample is close enough to peak for the validation plan's purposes.

Test plan

• Unit tests added/updated — 5 new `InjectorCacheStats` tests (hit/miss, in-flight coalescing, peakEntries, eviction, absent-stats regression) and a `phaseBreakdown` assertion on the existing VFR fixture that verifies the VFR preflight is counted and timed.
• Full engine suite: 318 tests pass (was 312).
• Typecheck clean in engine + producer.
• Lint + format clean on all changed files.
• Documentation updated — not applicable (internal instrumentation).

Follow-ups (PRs 2–5 in the stack)

Each of these will reference the numbers this PR emits to validate the improvement:

1. Segment-scope `convertSdrToHdr` — mirror of fix(engine): auto-normalize VFR video inputs to CFR before frame extraction #360's VFR fix; watch `videoExtractBreakdown.hdrPreflightMs` drop.
2. Extraction cache keyed by `(srcHash, windowStart, windowEnd, fps, hasAlpha)` — `extractMs` goes to ~0 on cache hit.
3. WebP as the single extraction format — unifies the jpg/png split; `tmpPeakBytes` decreases on alpha compositions.
4. Hardware-accelerated SDR decode, gated on `!hasAlpha` and segment-length floor — `extractMs` decreases on eligible inputs.

[jrusso1020]

• perf(producer): revert webp extraction default (20x slower than jpg on this host) #441
• perf(engine): consolidate cache key + reuse Phase 2 metadata + cleanups #437
• perf(engine): gated hwaccel auto for SDR opaque extraction #435
• perf(engine): webp extraction format + injector MIME routing #434
• perf(engine): content-addressed extraction cache for video frames #433
• perf(engine): segment-scope convertSdrToHdr re-encode #432
• perf(producer): phase-level render instrumentation #430 👈 (View in Graphite)
• main

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

[jrusso1020]

Closed: rebuilt as #444 off current main (post Vance's HDR work #371/#373/#375). Scope narrowed to extractor-side instrumentation + tmpPeakBytes; the injector-stats plumbing that conflicted with upstream #371's lifecycle refactor was dropped. See hyperframes-notes/perf-stack-rebuild-plan-2026-04-23.md