Redacted excerpt; full evidence withheld.

$ pwd; git status --short --branch; git rev-parse HEAD; python3 --version; sed -n '1,240p' README.md; rg --files -g 'AGENTS.md' -g 'pyproject.toml' -g 'requirements*.txt' -g 'package.json' -g 'Dockerfile*' -g '*.yml' -g '*.yaml' | head -80
exit: 0
 |
| `HF_HOME` / `HF_HUB_CACHE` | server | Standard Hugging Face cache location, honoured by both `serve` and `prefetch` (default `~/.cache/huggingface/hub`). |

Keepalive notes: the `websockets` library default (`ping_interval=20`,
`ping_timeout=20`) closes a live connection with `1011 keepalive ping timeout`
if a pong is late by 20s. During a heavy or cold-start generation, GIL-holding
Metal compute can starve the asyncio loop past that window and truncate the
in-flight utterance. So the server (and `TTSClient`) keep the periodic ping but
use a **large finite pong timeout (120s) by default** — long enough that a
briefly-starved loop never drops a live connection, yet bounded so a dead *idle*
[excerpt truncated]