┌─ vision thread ────────────────────────────┐   ┌─ audio thread (host-owned) ──┐
│ camera ─► GPU inference ─► manifold filter │──▶│ ring-buffer read (wait-free) │
│                                             │   │ interpolate to block time   │
│                                             │   │ write automation events     │
└─────────────────────────────────────────────┘   └──────────────────────────────┘
                   │
                   └──► UI thread (viewport, config) — read-only snapshot

The audio thread

This is where 90% of all real-time audio bugs live. Non-negotiable rules:

1. No allocation. Not Vec::push, not Box::new, not format!, not a HashMap::insert. Preallocate everything.
2. No locks. Not Mutex, not RwLock, not parking_lot. Use rtrb (wait-free SPSC) and atomic primitives.
3. No blocking I/O. No file I/O, no network, no logging that writes to a file. tracing events on the audio thread must route through a lock-free channel.
4. No panics. Panics in the audio thread are UB for the host.
5. Bounded work. The callback must return in less than the block time, every call, even in the worst case. Think in 99.9th-percentile terms.

Cross-thread communication

• Vision → Audio: rtrb ring buffer carrying Sample<G, C> values. Writer is the single vision-thread producer; reader is the audio thread. Wait-free on the reader side.
• Audio → UI: atomic snapshots of parameter state.
• UI → Vision: a separate rtrb ring carrying commands (reconfigure, reset, change model).

Timestamp discipline

Every sample carries a host-monotonic Timestamp. The audio thread knows now_audio - sample_timestamp and interpolates accordingly, so latency is measured, not guessed. Interpolation happens in the tangent space of the relevant manifold — you cannot slerp a homography by blending matrix entries.

Why not async

tokio and similar runtimes spawn worker threads that can starve an audio callback on a shared core and allocate unpredictably. They have no place in the real-time path. The vision thread can use a minimal runtime for GPU futures, but the audio thread sees none of it.