voice tests

Stdlib unittest only. No pytest, no extra deps beyond what voice.py needs (gguf, numpy). Run from the repo root, next to voice.py.

# offline: pure logic + fake-server download engine (always safe, <2 min)
python3 -m unittest discover -s tests

# live: real Hugging Face fetches (~1.6 GB first run, cached after) plus
# a full end-to-end cast and CPU inference run
VOICE_TEST_LIVE=1 python3 -m unittest discover -s tests

Live knobs (all optional):

VOICE_TEST_CACHE — model download cache (default ~/.cache/voice-tests).
LLAMA_CLI — inference binary (default: llama-cli on PATH).
No HF_TOKEN needed: every live fixture is a public repo.

Layout: test_units.py (pure logic) · test_download.py (fake-server engine) · test_live.py (cheap real reads) · test_voices.py (real head fetches + delta) · test_e2e.py (cast + llama-cli inference). Shared fixtures live in _helpers.py.