v0.18.1 — Fusion honesty + orchestration guards

Maintenance release: an adversarial review of the fusion surface caught five cost/honesty bugs the tests missed, plus orchestration hardening from M19.

Fixed — fusion (adversarial review)

• Receipts: a priced model with no reported token usage was pricing to a fake $0.00 ("free" masquerade). Now prices to None (unknown); a genuinely :free model still yields a real $0.00.
• Judge/synth prompt: a vision turn's multimodal content (base64 image part) was str()-dumped into the prompt. Only text parts are extracted now.
• Cost-aware router: an agreement-gate escalation that already produced a fusion result could be fused a second, redundant time — silently doubling spend. Guarded.
• Self-consistency: majority() accepted a mere plurality (2 of 5) as consensus; now requires a strict majority (> half), else synthesize.
• Arithmetic gate: tight digit-hyphen-digit (2026-07-10, 10-20, versions) routed dates/ranges to expensive fusion. Subtraction now requires whitespace padding.

Fixed — orchestration

• Diff-gate holds across the HITL pause: a tainted hollow success (verifier passed, empty diff) could be learned when approved on resume. The verdict is now persisted at the pause and re-applied on resume.
• chimera project: deny_card / approve_plan are guarded against stale/wrong card ids and inconsistent on-disk state.
• chimera features: now mirrors every [full]-extra capability; youtube_transcript dependency check fixed (import name, not PIP name).

Changed

• M19-A1 measured — reading skill cards stays OFF by default. Pre-registered paired A/B (mistral-small-24b, n=12): +16.7pp accuracy but CI [−13.3%, +30.3%] (not significant) and +300% tokens → both flip gates fail. Published either way (no re-rolling for significance).

Gate: ruff + mypy clean, 1364 passed / 3 skipped.