OnMic — what's new

- FIX: with "Back up timeline first" on, duplicating the sequence stole
  focus, so the razor cuts landed on the new backup copy while the
  enable/disable landed on the original — the edit was split across two
  sequences and looked like nothing got enabled/disabled. OnMic now
  switches focus back to your original sequence before editing, so all
  cuts and switches apply to the same (original) sequence; the duplicate
  stays as the clean backup. (Also fixes the same risk for dead-air trim.)

- NEW (beta): Trim dead air. In Advanced, set a threshold and click Trim;
  OnMic ripple-deletes pauses longer than that (keeping a small natural
  breath). Safe by design: it backs up the sequence first, and skips any
  pause where the tracks aren't continuous across the cut (so it can never
  desync audio and video). Undoable with Cmd+Z.
- Because it permanently restructures the timeline, try it on a copy first
  and tell us how it goes.

- NEW: Snap cuts to pauses (Advanced, on by default). Camera switches
  now land in natural silences instead of mid-word, so cuts feel human.
- NEW: Reaction shots (Layout → Off/Low/Medium/High). During longer
  pauses, OnMic cuts to a listener for a beat, then back — something a
  pure speaker-follower can't do. Off by default.
- NEW: Speaking-balance report after each analysis — a visual breakdown
  of each person's talk % and on-screen %, plus silence and reaction count.

- The 30-day free trial is now tied to the machine, not just the account.
  The panel sends a privacy-preserving hash (SHA-256 of the Mac's hardware
  UUID — never the raw id) so each Mac gets one trial. Making a new free
  WhereToPodcast account no longer resets the trial. Paid plans and Premium
  access are unaffected.

- NEW: "OnMic on WhereToPodcast ↗" link in the panel footer opens the
  product page (wheretopodcast.com/onmic).
- NEW: update notifications. The panel learns the latest shipped version
  from the verify response and shows a non-blocking "Update available"
  banner (with a download link) when it's running an older build.
- Latest version is stored in the site DB (onmic_meta.latest_version), so
  it can be bumped on each release without redeploying the site.
- Click the version number at the top of the panel to open the public
  changelog (wheretopodcast.com/onmic/changelog).

- NEW: OnMic now requires a WhereToPodcast (Clerk) login. The panel opens
  the browser to wheretopodcast.com/activate; you just sign in and a signed
  token bounces back to the panel over a 127.0.0.1 loopback (nothing to type).
- 30-day free trial from first activation; after that, the site's existing
  paid tier (an active premium / premium_plus account) keeps OnMic unlocked.
- Account chip shows trial days left / Subscribed; Sign out included.
- Offline grace: if WhereToPodcast can't be reached, recently-verified
  access keeps working for up to 72h so a connection blip never blocks an edit.
- Server side (wheretopodcast repo, branch onmic-auth): /activate page,
  /api/onmic/verify, onmic_users trial table, HMAC token lib.

- NEW: "Back up timeline first" — before applying cuts, OnMic duplicates
  the active sequence into the Project panel (named "<sequence> — OnMic
  backup <date time>"), so your original timeline is always preserved.
  On by default; toggle in the Advanced card; saved with profiles.
- If the backup can't be created, OnMic stops and changes nothing rather
  than risk editing without a safety copy.

** Major: OnMic no longer needs Python at all. **
  - Analyzer fully ported to JavaScript (js/engine.js): WAV decode
    (16/24/32-bit int + 32-bit float, multichannel, windowed), clip/
    offset timeline reconstruction, RMS, level normalization, dominance/
    overlap, smoothing — verified to match the old Python output.
  - Mix/isolated channel detection ported to JS too (no detect_channels.py).
  - AI voice detection (silero) now runs IN THE PANEL via ONNX Runtime
    Web — no Python, no PyTorch. Fixed the silero v5 call to feed the
    required 64-sample context (576-sample input) + scalar sr; max
    speech prob went 0.002 → ~1.0. Auto-boosts quiet mics to a normal
    speech level (with soft-limiting) before inference. Falls back to an
    energy+zero-crossing voice filter if the model is ever unavailable.
  - Panel runs everything internally; the Python spawn is gone.
  - AI model + runtime are fetched once (fetch_vendor) and bundled, so
    end users still get a single one-click install.
  - Distribution is now Python-free: "Install OnMic.command" just allows
    the panel and copies it in (no venv/pip). The panel writes its working
    files to ~/Library/Application Support/OnMic when no dev project folder
    is set. build_dmg.command bundles the self-contained panel + AI.