v0.16.1 — Audit hardening

A maintenance release from a full functional audit of the repo (3 parallel reviewers + live smoke tests across all three tiers). The audit found the core solid — 1295 tests, mypy fully green, chimera maturity 37/37 GA, and run/fuse/solve verified end-to-end on real models — and closed the loose ends it surfaced.

Added

• chimera hierarchy-bench — first-class CLI command for the hierarchy paired A/B (single-agent all-docs-inline vs orchestrator-worker hierarchy, same model both arms to isolate the orchestration). Closes the asymmetry where every peer bench had a command but the hierarchy's eval module was reachable only from the external bench/ scripts.

Changed

• The aggregate cross-agent collusion monitor now runs for the fan-out commands (solve-batch, crew-isolated). The v0.16.0 monitor was built + tested but connected to no orchestrator, so its split-exfiltration defense was unreachable in a real run. It is now always on (pure observability; escalates a review note, never blocks) — each parallel worker gets its own capability ledger and the monitor runs over all of them. --taint now controls only the stronger per-worker adaptive allowlist.
• Collective cross-model skill evolution now fires under --cascade, not just --fuse — both routes reach fusion over the same panel, so a cascade run now keeps the most transferable skill proposal too.

Fixed

• Honest feature catalog — dropped computer_use (no impl/dep), renamed voice_mode→speech_io (the STT/TTS tools that exist), and labelled x_search/spotify as pluggable-via-OpenAPI.
• Graceful missing-key handling for skills-evolve, playbook curate, rubric-grade (were tracebacks).
• mypy chimera fully green (219 files) — 3 scrape type-narrowing fixes.
• Test added for the messaging missing-extra install hint.

Still alpha. Everything defaults to prior behaviour. pip install -U chimera-agent.

Full changelog: https://github.com/brcampidelli/chimera-agent/blob/main/CHANGELOG.md