Synthesise a soundtrack on the same clock as the cuts

[LeadcodeDev]

Part of #326. Batch 2 — depends on #336 for TimePoint and the beat grid.

AudioTrack requires a src. An LLM can write a score and cannot hand over a WAV, so every generated scenario is silent. Both reference reels synthesise their entire soundtrack from oscillators, noise and filters, with no audio file anywhere, and the SVG one renders it offline into a buffer before muxing. This is the single largest gap between a generated scenario and a generated reel, measured by how much of the perceived result it carries.

Owned files

• crates/rustmotion-core/src/audio/ (new: voices.rs, score.rs, synth.rs, dsp.rs, mod.rs)
• crates/rustmotion/src/encode/audio/ (the offline render and the mux path)

Do NOT write: schema/scenario.rs (#336 registers the audio.synth field), anything under engine/.

Deliverables

1. A declarative score sharing the visuals' clock:

   "audio": { "bpm": 115,
     "voices": { "kick": { "type": "sine", "freq": [150, 42], "sweep": 0.14, "decay": 0.38 },
                 "hat":  { "type": "noise", "filter": {"type":"highpass","freq":7500}, "decay": 0.05 } },
     "score":  [ { "voice": "kick", "every": "1b", "from": "@2.2s", "to": "@12.6s" },
                 { "voice": "hat",  "every": "1b", "offset": "0.5b", "from": "@2.2s" } ],
     "master": { "compressor": { "threshold": -14, "ratio": 4 }, "limiter": true } }

   every, from, at and offset go through Place scenes on an absolute timeline with a beat grid #336's TimePoint parser. Sharing bpm and beat_offset with the cuts is the whole point: it is what makes an edit land on a downbeat rather than near one.

2. A small, boring synth. Sine, square, saw and triangle oscillators; a noise buffer; biquad filters in lowpass, highpass, bandpass and notch; ADSR envelopes; a bus with gain, a compressor and a limiter. Rendered offline to an f32 buffer at 48 kHz.

3. Coexistence with file-based tracks. audio accepts both the existing AudioTrack entries and a synth block in the same scenario, mixed into one bus.

4. The limiter is a default, not an option. The SVG reel's author found his first mix clipping at 0 dB only by reading ffmpeg output after the fact.

5. --frames a-b keeps its audio slice. A segment renders its own slice of the synthesised buffer, not a restart from zero, matching what the existing file-based path already guarantees.

Out of scope for this workstream: convolution reverb, sample loading, any effects rack. A synth that grows is a synth that stops being reviewable.

Acceptance

• A scenario with no audio file and a 60-event score produces a muxed soundtrack
• Two renders of that scenario produce byte-identical audio
• Peak never exceeds -0.1 dBTP with the limiter on

[LeadcodeDev]

Delivered on main at 187cf7c (squash of #340). Verified present on main, not taken on the PR's word.