Unify frame sampling, add fast mode, and support bundled-app transcription

[Justrada]

What changed

Replaces the codec-specific I-frame extraction path with a single
binary-subdivision sampler that works for any codec. Frames are grabbed
in coverage-priority order (endpoints → midpoint → quarters → eighths…) and
bounded by both a max_frames cap and a time_budget, so faster machines
and shorter clips naturally get denser coverage — no per-codec branching.

Highlights

• core/frames.py — drop the ALL_INTRA_CODECS special case. New
  sample_frames() (capped subdivision + similarity filtering) and
  sample_fast_frames() (N equidistant frames, no filtering).
• core/schemas.py — new settings: max_frames, time_budget,
  fast_mode, fast_frame_count.
• core/transcribe.py — mlx-whisper now falls back to a subprocess
  using whichever Python interpreter has it installed, so transcription works
  in the packaged app; the in-process path is kept for running from source.
• core/llm.py — thread a logger into LLM clients; clearer
  transcript-summary logging plus a retry on the first failure.
• AutoBin.spec — bundle mlx_whisper + huggingface_hub into the
  frozen build.
• gui/ — surface fast mode and the new sampling knobs through the
  settings panel, and wire them through the orchestrator and workers.
• .gitignore — ignore *.LRV test footage and node_modules / the
  remotion out/ dir.

Why

Per-codec branching was fragile and gave inconsistent coverage. A unified,
budget-bounded sampler is simpler and adapts to the machine/clip. Fast mode
gives a quick low-cost pass (few frames, tiny whisper, single LLM batch).
Subprocess transcription unblocks the packaged/bundled app.

Reviewer notes

• Unrelated untracked items in the working tree (a remotion/ experiment
  with node_modules, a docs flowchart, and a test .LRV clip) are
  intentionally excluded from this PR; the heavy ones are now gitignored.
• No automated tests in this repo — changes exercised manually via the GUI.

🤖 Generated with Claude Code