Add voice-follow to mc-prompter: streaming local ASR tracks the speaker (Phase B)

[bmadcode]

What this is

Phase B of mc-prompter, PR 2 of the 3-PR stack (PR 1: #4, the classic teleprompter). This adds voice-follow: the scroll tracks the speaker through the script using streaming local ASR and a deterministic alignment engine. No LLM anywhere in this PR; everything is local and free.

What you get

• The prompter-lab workspace: ensure_workspace.py builds a persistent venv and downloads the ASR models (~465 MB, consent required, idempotent, resumable against truncated downloads). Without the workspace, tier 1 keeps working untouched.
• Streaming ASR: sherpa-onnx 1.13.4 with the NVIDIA nemotron-speech-streaming-en-0.6b int8 export (the streaming sibling of the parakeet family already used for cutting) plus silero VAD, on a dedicated worker thread with a bounded drop-oldest queue and an honest BEHIND indicator. An important lane finding is recorded in the skill docs: sherpa-onnx model exports are coupled to the runtime generation (a 2023 zipformer export silently produces garbage under 1.13.4), so runtime and model pins move together and zipformer-small stays a planned lane that exits with a pointer.
• The alignment engine (align.py, pure stdlib): BPE-aware, tail-revision-tolerant bounded-window matching with a monotonic committed anchor, hold-on-adlib and re-anchor recovery, number/abbreviation normalization, and free skipping of TAKE paragraphs (the lines a creator does not re-record). Its fixture suite runs on seven REAL captured partial-event streams from the actual model (verbatim, ad-lib, skipped sentence, digits-vs-words, tail revisions, take-skip, stopword-heavy ad-lib).
• Browser capture: getUserMedia with echo cancellation, noise suppression, and auto gain explicitly OFF (they degrade ASR), 16 kHz worklet frames over the existing WebSocket, single capture owner with a token grant, backpressure at both ends. Mic capture stays on the server machine; tablets and phones remain display/control surfaces.
• The preflight panel: device picker with persistence, live level meter, applied-constraints readback, live partial line, and a tracking check that passes only when the anchor is genuinely following. Voice-follow is off by default.
• Follow-mode UX: eyeline-led eased scroll, match tinting, HOLD and BEHIND chips, click-any-word to re-anchor, and clean fallbacks for reconnects, mic unplugs, and token rejection.

Verification

• 210+ unit tests green under the CI quality gate command, genericity lint clean, all Phase A behavior regression-tested
• A real-model replay harness measured capture-to-anchor latency at 72 ms median over a 32 s read with a fully monotonic anchor
• Live browser end-to-end on this machine: simulated mic audio through the exact capture path, preflight tracking check passing, the scroll following the voice to precisely where the audio stops, HOLD on silence, zero console errors
• Adversarial 3-lens review after all of the above found 20 more issues (orphaned process trees, BOM parsing, WS leader races, hide-takes pacing math, a take-paragraph stall in the aligner); all fixed with regression tests and zero recall loss on the captured fixtures

🤖 Generated with Claude Code

https://claude.ai/code/session_01BbneC1PckyV8vkDzhGYkpN

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: b6491273-31ff-4566-acdf-df4f51f03c55

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat-prompter-b

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}