A capture hook, so the UI can be recorded deterministically

[spacedevin]

Recording Deckard playing a set meant screencasting it in real time — which drops frames under load
and never renders the same way twice.

The UI already subscribes to transport.position on the signal bus to drive the playhead, grid and
timeline. So a recorder can publish that position itself, one frame at a time, and screenshot each
one: every frame is then drawn at exactly the timecode it will occupy in the video.

The seam was missing. signals is created inside the App component and nothing outside can reach it.
With ?capture=1, this exposes the runtime handle and the bus:

if (typeof window !== "undefined" && window.location &&
    String(window.location.search).indexOf("capture=1") >= 0) {
  window.__deckardCapture = { rt: rt, signals: signals }
}

It extends the same rt escape-hatch object the app already keeps for setActiveDeck and the
undo/redo entry points, rather than inventing a new mechanism. Gated on the query string, so it is
not an API the app ships to ordinary users, and a normal load is unchanged.

What it is for

A capture harness in the chuggie repo (scripts/deckard-mix) that loads preset sets onto decks,
crossfades between them, renders the mix offline, and records the UI against it. It drives Deckard's
real model, real preset sets and real sequencer — the same move renderDeckLoopToBuffer() already
makes internally, widened from one mono deck to a stereo multi-deck mix.

Building it surfaced two bugs worth knowing about here, both in the render path rather than in this
diff:

• Voices retire themselves with setTimeout against wall clock. An offline render finishes a
  minute of audio in about a second, so those cleanup timers fire mid-render and tear down notes
  that are still sounding. Deferring them for one render recovered 7 dB of RMS and a truncated
  tail. Live playback is unaffected — wall clock and audio time agree there — but anything rendering
  offline through these voices is losing note tails today.
• Noise voices seed from Math.random(), so no two renders match. The harness installs a seeded
  PRNG; after that, two runs differ by less than one 16-bit LSB.

Verification

• npm test — 516 pass, 0 fail.
• Hook verified in a real browser: __deckardCapture.rt.project reachable, signals.publish
  present, rt.setProject re-renders after loading sets onto decks.
• A 4-bar capture produced 230 frames at 30fps muxed against the rendered audio, with the playhead
  confirmed advancing between frames.