feat(ingest): store extracted video frames at a chosen scale

[JArmandoAnaya]

What changed

Ingest-time downscale for video sources: registration takes an optional scale_percent
(1–100, default 100), and every extracted frame is stored at that percent of the clip's
native size per side.

• Identity. The scale is a decomposition parameter, so it joins the source's identity
  beside the rate and the clip ranges: the same clip at another scale is a second source.
  Migration 18 reshapes the origin index to
  uq_source_project_kind_path_fps_ranges_scale; an unscaled clip omits the JSON key, so
  rows written in any generation share one spelling and nothing forks retroactively.
  Migration 16 became drop-only in the same change — only the head reshape may execute the
  shared index declaration, because that declaration is always the current spelling.
• Extraction. _filtergraph gains a final scale=W:H stage with dimensions computed
  in Python from the probe (max(1, (native·pct+50)//100), integer half-up — mirrored
  exactly by the screen so preview and stored size cannot drift). The whole-clip 100%
  command stays byte-identical, pinned by test.
• Surfaces. REST scale_percent form field, CLI --scale, MCP scale — all
  video-only, refused on a directory of stills the way --fps is. VideoProvenanceOut
  publishes the value; openapi.json, the generated client, and the MCP tool reference
  are regenerated.
• UI. The video step gains a "Stored size" block: an always-visible readout
  (1280×720 · native, or 1280×720 → 640×360 · 50%), a native range slider with
  endpoint labels, and a purpose line stating the pixel cost. The step-2 fact grid shows
  the same arrow. The ingest preview player is now always muted, with the volume control
  hidden — a vision dataset has no use for the audio track.

Per-image scaling for image batches was implemented and then withdrawn during review; the
history carries both halves, and the final tree scales video only. Image directories store
stills at their decoded size — a uniform size remains the export recipes' job.

Found, not fixed

Nothing — the one defect found along the way (an object URL revoked under StrictMode's
double-mount, breaking dev-server thumbnails) was in this branch's own code and is gone
with it.

Test plan

• uv run pytest tests/kernel tests/server tests/cli tests/mcp — green.
• uv run lint-imports, uv run mypy src/visionset/kernel, uv run ruff check . — green.
• openapi.json / generated client / docs/content/mcp-tools.md drift — clean.
• pnpm -r build && pnpm -r test && pnpm -r lint — green; real-server cycle suite green.
• Verified live in Chromium against a real server: scaled registration, stored frame
  dimensions on disk (640×360 from a 1280×720 clip at 50%), identity forking, the muted
  player, and the images panel unchanged.