fix(whisperkit): streaming translate slows after ~30-40s — task-aware confirmation lag

[initcore0]

Symptom

WhisperKit backend, Russian speech, translate-to-English on: works well at first, gets noticeably sluggish after ~30–40 s of continuous dictation.

Root cause (researched, fork + app both mapped)

The streaming loop (AudioStreamTranscriber) re-decodes everything after lastConfirmedSegmentEndSeconds on every pass, and confirmation permanently holds back the newest requiredSegmentsForConfirmation (2) segments — so the re-decoded tail settles at ~2–3 segments (10–25 s of audio). On task: .translate every pass must also re-generate English for that whole tail (more tokens, lower confidence, more decoder fallbacks than same-language transcribe), so once the tail fills out, per-pass cost outgrows the loop's hard-coded ≥1 s new-audio cadence and partials lag — exactly a "fine at first, slow after 30–40 s" curve. The knob was never passed at our call site; both tasks silently ran the upstream default.

Fix

WhisperKitStreamingDecodePolicy.requiredSegmentsForConfirmation(translate:) — translate holds back 1 segment (clip point advances a segment sooner → steady-state re-decode window roughly halves), transcribe keeps the upstream default of 2. Passed explicitly in makeStreamHandle. Tests pin the per-task values, the ≥1 floor (0 would freeze half-heard words with no revision room), and that the translate threshold still satisfies the #222 confirmed-end advancement invariant (no 30 s-truncation regression).

Also fixes docs/WHISPERKIT_PILOT.md, which still documented withoutTimestamps: true — the exact configuration #222 removed — plus its stale "file transcription only" limitation.

Testing

• swift test: 1937 tests green (3 new in WhisperKitStreamingDecodePolicyTests)
• ./build.sh (WhisperKit default): clean, no new warnings
• ⚠️ Needs a live pass: Russian dictation with translate on, 60–90 s continuous, watching whether partial cadence stays steady past the 40 s mark. Not merging until that's done.

Follow-ups (researched, out of scope here)

• Bigger win available: transcribe natively + translate the text via the Apple Translation path (PR feat(translate): on-device text translation via Apple Translation framework (macOS 15+) #223's engine-agnostic mechanism, currently policy-gated to ASR-only engines). Constant low latency regardless of session length; tradeoff is the live preview showing the source language until paste.
• Fork-side levers if needed later: the loop's hard-coded ≥1 s cadence, and the full-session buffer copy per pass (negligible below ~5 min).
• MAK-72 still tracks the missing streaming-latency instrumentation.

🤖 Generated with Claude Code