OnMic — what's new

- "Re-read tracks from sequence" is now a button, matching the "Auto
  Detect cameras" button for a consistent look.
- The Log panel is collapsed by default — click "Log" to expand it.

- "Detect cameras" is roughly 5x faster — it now samples motion at a
  lower resolution and uses fewer sample points, with no loss in which-
  camera-shows-whom accuracy.
- Corrected the sign-in screen: new accounts get a 14-day free trial
  (it previously said 30 by mistake).
- Internal: the panel's bundle version now advances with each release so
  Premiere always loads the new version instead of a cached copy.

- The "Advanced" settings card is now collapsed by default and moved to
  the bottom of the settings (just above the Create / Preview buttons).
  Click its header to expand it. All the advanced options (delay cuts,
  ignore-cut threshold, output mode, backup, snap-to-pauses, trim dead
  air) are unchanged and still remembered with your saved setups.
- The "Profiles" card is now collapsed by default too — click its header
  to expand and load/save setups.

- Group shots now resolve to the specific people in them, not "everyone":
  each face is found and its mouth matched to a mic, so a two-shot of two
  people tags exactly those two.
- Wide shots are recognised by face size — a camera whose faces are much
  smaller than your close-up shots is tagged as the whole room (everyone),
  since a detector can't reliably find every face in a wide (profile /
  behind mics). Fixes a 4-person wide being tagged as just the 2 visible.
- A face clearly on camera is always tagged as a real person (no longer
  dropped just because its lip-to-mic match was weak).
- Face finding keeps the frame's real aspect ratio (letterboxed) instead
  of squashing 16:9 into a square.
- The "who's talking when" timeline uses AI voice detection (silero) when
  available instead of raw mic loudness.

- FIX: the panel could load its UI shell but never actually start up —
  stuck on "Connecting to Premiere…", no version number, no log. Cause:
  the panel's scripts load at the end of the page, and the DOM had often
  already finished parsing by the time startup ran, so the "page ready"
  event it waited for had already fired and never came again. Startup now
  runs immediately when the page is already parsed. (Verified on a real
  project: connects, reads tracks, builds the speaker/camera rows.)

- FIX: startup is now fault-isolated. Each boot step (connect to
  Premiere, build the speaker/camera UI, load the AI model, wire the
  buttons, Detect cameras) runs independently, so if any one of them
  hits a snag it's logged and skipped instead of silently leaving the
  panel stuck on "Connecting to Premiere…" with no UI. If a real problem
  occurs it's now shown on the status line instead of hidden.
- Manual speaker/camera tagging is and remains the main workflow;
  "Detect cameras" is an optional helper that fills the tags in for you.

- Detect cameras now locates each person's MOUTH and measures lip motion
  there, instead of whole-frame motion — far more reliable for deciding
  who a camera is on. It runs a small on-device face detector (YuNet, via
  the same onnxruntime already bundled; no Python, no cloud), finds the
  mouth region once per camera, then correlates lip movement with each
  mic's talk timeline.
- Multiple faces in one shot are recognized as a wide/group camera.
- If a face can't be found (or the footage can't be decoded), it falls
  back to whole-frame motion, and anything it can't read is left for you
  to tag manually. Still a proposal you review before cutting.

- NEW (beta): "Detect cameras" under Tag Speakers. Instead of tagging
  who each camera shows by hand, OnMic can guess it for you. Because the
  mics already tell it who's talking each moment, it samples frames from
  each camera, measures on-screen motion (a talking person moves), and
  matches that to each speaker's talk timeline — the camera that moves
  when Abby talks is Abby's; a camera that moves for everyone is read as
  a wide shot. It fills in the tags as a proposal you review before
  cutting; nothing is locked in.
- Works on footage the panel can decode (e.g. H.264); any camera it
  can't read is simply left for you to tag manually. Manual tagging is
  unchanged and always available.