audiobench 0.1.1

First public release on PyPI. (Version 0.1.0 was tagged in git as part
of pre-release work but was never published to PyPI; 0.1.1 is the first
version users can pip install.)

Added

• ab/asr-robust suite: speech recognition under clean, noise-cafe-10db,
  noise-pink-5db, bandlimited-8k, and reverb-medium conditions. Reports
  per-condition WER and a weighted mean. Default model: Whisper.
• ab/sound-id suite: sound-event identification on labeled mixtures, with
  solo / pair / triple / quad mixture-size conditions. Reports
  recall, precision, F1, and false-positive rate per (pack, condition),
  plus a headline "components understood: X / Y" number.
• Bundled demo pack runs end-to-end on a fresh laptop with no GPU, no
  weight downloads, and no network. Five additional packs (core, home,
  cabin, security, health) reference user-supplied datasets at
  ~/.cache/audiobench/sound_id/.
• Four model adapters for ab/sound-id:
  • heuristic-v0 and heuristic-weak: deterministic spectral matchers
    bundled with the package; no GPU, no downloads.
  • clap-base: LAION-CLAP zero-shot. Lazy-imported.
  • qwen2-audio-7b: Qwen2-Audio-Instruct via local transformers or a
    remote endpoint via AUDIOBENCH_QWEN_ENDPOINT. Lazy-imported.
• Versioned prompt protocol pinned in run_hash (prompt_version,
  parser_version, prompt_ensemble, plus a SHA-256 over the canonicalized
  paraphrase list). audiobench compare refuses mismatched prompts unless
  --allow-mismatched-prompt is passed.
• Recipe-driven mixture authoring: bundled defaults, inline --mix "label1+label2", and YAML/JSON recipe files with per-source dB levels.
• Mixture preview command: audiobench mix preview writes a WAV without
  running probes.
• Forensic per-mixture view: audiobench inspect <run.json> --mixture N.
• audiobench compare for both suites; audiobench push writes a signed
  local payload (no network in MVP mode).
• Reproducibility guarantees: fixed manifest/mixture/probe seeds,
  deterministic mixer, and a run_hash SHA-256 over the canonicalized run
  payload (mixture spec + prompt config) written into every run JSON.
• Optional installs:
  • pip install "audiobench[clap]" — adds the LAION-CLAP zero-shot adapter.
  • pip install "audiobench[qwen]" — adds local Qwen2-Audio via
    transformers + torch.
  • pip install "audiobench[docs]" — MkDocs Material toolchain.
  • pip install "audiobench[dev]" — pytest, build, twine.

Notes

• Python 3.10+ supported; tested on 3.13.
• macOS users may hit ModuleNotFoundError immediately after
  pip install -e . due to a known macOS + pip + Python 3.13 site.py
  interaction (Python issue
  #127012 /
  pip issue #13153). Fix:
  chflags -R nohidden .venv/lib/python3.13/site-packages.