Use the moq libraries instead of hand-rolling the wire codec

[kixelated]

A pass on top of the POC swapping the hand-rolled wire codec on both sides for the upstream moq-rs (Python) and @moq/publish (browser) libraries. Net ~−1,500 lines vs. the POC.

Benefits

• Opus-encoded audio on the wire. No more raw 16-bit PCM streaming. Bandwidth drops from ~384 kbps (24 kHz mono PCM) to ~32 kbps for the bot's audio, ~−12×. libopus also handles packet loss concealment, so dropouts sound like brief artifacts instead of clicks.

• Battle-tested Rust QUIC stack. moq-rs wraps moq-native, which gets built-in caching, fan-out, prioritization, congestion control, and WebTransport ALPN negotiation. We were re-implementing slivers of all of these in Python on top of aioquic. Gone.

• max_latency_ms SLA on the subscriber. Default bumped to 500 ms (was 100 in the previous draft). Under healthy networks frames flow with no buffering; under congestion the consumer waits up to 500 ms for late groups to fill, then skips ahead. This is the structural win over WebRTC — there's no "choose audio loss vs. stall," you set a deadline and the library obeys it. Tunable via MOQParams.audio_in_max_latency_ms.

• No moq-relay process for local dev. New --moq-serve flag binds the bot to a local UDP socket via moq.Server. The bot is the relay. scripts/moq-dev-setup.sh is deleted. One terminal, not two.

• Self-signed certs minted in-process. tls_generate=[\"localhost\"] produces a cert on startup; server.cert_fingerprints() exposes the SHA-256, the runner threads it back through /start, and /api/config hands it to the browser for WebTransport pinning. No PEM files, no symlinks across repos.

• @moq/publish does browser audio capture. Replaces ~250 lines of AudioWorklet + Int16 PCM + framing with Publish.Broadcast({ audio: { source: mic.source } }). We get Opus encoding via WebCodecs, catalog publishing, and the subscribe-fulfilment loop for free. AudioWorklet code now lives entirely upstream.

• WebSocket fallback for free. @moq/net races WebTransport and WebSocket — Firefox (no WebTransport yet) just works without the demo knowing.

• Catalog-driven track discovery. The bot writes a catalog from publish_audio's AudioEncoderInput; the browser reads it from subscribe(\"catalog.json\") and uses whatever rate/channel count it advertises. Track names aren't pinned in the config — adding a video track or a screen-share later is a one-liner.

• Better resampling. moq-rs uses rubato inside its publish_audio / subscribe_audio so you can hand it any sample rate and it does sinc-interpolation to/from libopus's native rates. The earlier draft used a hand-rolled 4-point linear interpolator that was fine for speech and dubious for anything else.

• Single dependency. pyproject.toml [moq] extra is now just moq-rs>=0.2.13 and cryptography. aioquic and opuslib are both gone; libopus is statically bundled into the moq-rs wheel so there's no brew install opus / apt-get install libopus0.

• Clean separation of concerns. transport.py is down to ~700 lines — basically a thin glue layer between the pipecat Frame pipeline and the moq library. No varint codec, no stream-state machine, no codec helpers.

What's still hand-rolled

The browser-side playback loop — @moq/hang's Container.Consumer + WebCodecs AudioDecoder + Web Audio scheduling, about 60 lines. Could swap for @moq/watch.Audio.Decoder later, but that pulls in Sync/Source/jitter-buffer machinery that's overkill for a single-track playback path.

Things to know before merging

1. The pipecat Client SDK transport plugin (@pipecat-ai/moq-transport) is still a future item, per moq_prebuilt/PLAN-moq-transport-package.md. The browser demo here is still ad-hoc.

2. End-to-end testing requires moq-rs 0.2.13 installed (auto via uv sync --extra moq). I haven't run the live mic-to-speaker round-trip in this branch — the smoke test is uv run python examples/transports/transports-moq.py -t moq --moq-serve and opening localhost:7860.

3. The browser loads @moq/net, @moq/publish, @moq/hang, @moq/signals from esm.sh. If that bothers you for production, the path is a tiny Vite build dropped into moq_prebuilt/client/. For dev / demo it works as-is.