Design: parakeet-asr

Process split

GitHub Pages is static. The ASR API is a Bun process on fractal1, sibling to aicam-bay-feed (:8745), bound to the Tailscale IPv4 only on :8750. UPnP is on; never 0.0.0.0.

caller --HTTP--> Bun Hono :8750  --stdin/IPC-->  Python onnx-asr sidecar
                 OpenAPI here                    INT8 TDT v3, CPU EP

Hono owns HTTP, Zod, and the spec. The sidecar owns the model. Bun does not dlopen CUDA. The sidecar sets CUDA_VISIBLE_DEVICES= and providers=["CPUExecutionProvider"]. Four intra-op threads.

Papyrus morphist-asr is the policy source (CPU-when-AVX2, 16 kHz mono f32, INT8 ONNX, keep the dGPU free). It is utterance-batch, CoreML on Apple, and a Tauri crate. We do not vendor it.

Discovery

Copy daBOM, not the SvelteKit mount:

URL	Body
GET /.well-known/openapi.json	OpenAPI 3.1, canonical; servers: [{ url: '/api/v1' }]
GET /api/v1/openapi.json	same bytes
GET /.well-known/api-catalog	RFC 9727 linkset

app.doc31('/openapi.json', …) generates the document from createRoute. Agents read well-known. Do not hard-code path lists in clients in this repo.

Audio

Input is 16 kHz mono PCM (or WAV that already is that). Wrong sample shape is 422. Non-audio content-type is 415. The sidecar does not resample or mix channels. Clip POST is one utterance. Stream is windowed TDT (VAD or N-second chunks) emitting NDJSON {text, t0, t1, final}. v1 stream clients are non-browser. No Unified RNNT in this change.

The sidecar is a child of the Bun process on stdin/IPC. It does not open a network listener. Bun is the sole tailnet socket.

GPU

A new CUDA process during transcribe is a failed acceptance, not a warning. Bay feed stays the 3090 tenant.