Add real-time stdio streaming runtime with framed PCM protocol

[nsalminen]

This PR adds a real-time stdin/stdout runtime for Moshi (moshi-stdio) so audio can be streamed in/out of other processes without the web UI/WebSocket stack. It includes EOF drain handling to avoid truncated tails and a concise stdio test runner that always executes the end-to-end path using existing test assets.

Approach

The runtime uses a framed PCM protocol ([u32 len][payload]) with explicit message kinds and buffers PCM16 frames to the model frame size. On EOF, it feeds configurable silence frames to flush delayed outputs. A single test runner script validates framing and then runs moshi.stdio end-to-end.

How to test:

./.venv/bin/python moshi/test/stdio_realtime_check.py \
  --moshi-args --voice-prompt NATM1.pt --device cuda

Optional alternate input:

./.venv/bin/python moshi/test/stdio_realtime_check.py \
  --input-wav assets/test/input_service.wav \
  --moshi-args --voice-prompt NATM1.pt --device cuda

Key Changes

• moshi/moshi/stdio.py
  • Framed protocol: [u32 len][payload]
  • Kinds: 0x00 handshake, 0x01 audio (PCM16), 0x02 text, 0x05 error, 0x06 ping
  • PCM16 buffering and EOF drain frames (default 32)
  • stderr-only logging to keep stdout binary-safe
• moshi/pyproject.toml
  • new entrypoint moshi-stdio = moshi.stdio:main
• moshi/test/stdio_realtime_check.py
  • framing roundtrip + always runs stdio E2E
  • uses existing assets by default and writes outputs to repo root