Feat/library stems midi first class

[danieljtrujillo]

feat(library): make stems & MIDI first-class]

Stems and MIDI in the library can now be played, favorited, deleted, and
routed anywhere audio goes, instead of only offering a right-click send menu.

• DB schema v5: favorite column on stems + midis; get/set-favorite/delete
  methods for both.
• Endpoints: PATCH/DELETE /api/library/stems/{id} and
  PATCH/DELETE /api/midi/file/{id} (delete removes the file + row, leaves
  the parent track and siblings intact).
• Library STEMS/MIDI rows get inline play/pause, favorite star, and delete;
  MIDI right-click now routes to editor/init/inpaint/chimera.
• Shared lib/midiSynth.ts extracts the offline synth voice + WAV encoder
  (renderNotesToBlob / renderStepNotesToBlob / renderMidiBufferToBlob);
  PianoRoll delegates to it so previews, the editor bounce, and library MIDI
  all sound identical. midiIdToSendable() makes any library MIDI a lazy
  SendableAudio.

fix: module-list reliability, Windows MIDI/stems incidents, Suno a11y

Settings 'Backend Modules' could show a false 'No modules found' when opened
during a backend (re)start, because the modal fetched on open and silently
blanked on any error. Now the catalog loads once on backend-ready into a
shared moduleStore (retry-until-success, cached) and Settings just reads it,
so there is nothing to fail on open; a real error shows an error + Retry.

Also:

• MIDI (Windows): basic-pitch's emoji status output no longer crashes MIDI
  conversion on cp1252 consoles — its stdout/stderr is captured into a text
  buffer (engine.py), so transcription failures report the real cause.
• Stems sidecar: probe now checks all critical packages (demucs/torch/
  torchaudio/torchcrepe), and ensure_running installs deps BEFORE spawning
  when any are missing instead of letting run_backend.py self-install and burn
  the 300s readiness window; the timeout error now names the dep state.
• Suno API key field is wrapped in a with an associated label and
  autoComplete=off (silences the 'password field not in a form' warning).
• docs: record Phase I (stems/MIDI), G15 (device errors + Quest video-in),
  and H1/H2 status in the plan.

fix(generate): reliably surface freshly-generated tracks in the library (G1)

After a generation completes the backend writes artifacts synchronously, but
the library list index can lag that write by a beat — when it did, the entry
wasn't in the just-refreshed list and the track silently failed to appear
('manual reload'). Now the post-generation reconciliation retries the refresh
a few times until the expected <job_id>_ id shows up, and on genuine
failure surfaces a visible status message instead of only logging.

Verified the rest of the add-to-library chain end-to-end: the /api/library/
import endpoint (200 + full entry), the frontend contract, importEntry's
in-place store update, and the backend's <job_id>_ id scheme all match.

docs: record G15 research findings (Quest video-in without MQDH — scrcpy+OBS
virtual cam Tier 1 needs only a VJ device picker; raw passthrough is Tier 2
via the Unity Passthrough Camera API).