hotato 1.12.0

Score voice and chat agents, exercise them harder, decide with layered deterministic checks, and gate every pull request.

Catch a failure from a recording, a timestamped transcript, or your traces; acoustic health signals and cross-run failure clustering (hotato diagnose --fleet) find the pattern.
Exercise the agent: seeded personas, chat simulation over local HTTP (hotato simulate --chat), robustness batteries at staged SNR/clip/dropout, scenario variables and branch enumeration.
Decide with deterministic assertions: composite formula and conditional when:, policy check packs (hotato assert run --pack), the new order kind.
Gate with content-addressed contracts, per-dimension baseline drift tolerances (hotato baseline check), PR step summaries, and fleet latency percentiles.

Same input, same verdict, byte for byte. Offline. The scoring engine is unchanged.

Full changelog: https://github.com/attenlabs/hotato/blob/main/CHANGELOG.md