v0.15.0 — Trustworthy Verifier & Write Security

The M18 cycle, grounded in an 8-category arXiv sweep — five Tier-1 items that harden the weakest link in the "prove a weak model performs like a frontier one" thesis (the verifier), plus write-side security and a measured skill-lifecycle loop. Every item is an integration (never a reimplementation), opt-in, degrades safely, and is fake-tested. The headline shipped with a measured paired A/B and an honest retraction.

Added

• solve --gen-tests — spec-grounded test generation (M18-1). With no --verify command, the fitness gate falls back to an LLM judging coverage — a proxy that rubber-stamps wrong code. --gen-tests generates executable pytest grounded in the task's atomic requirements and runs it as the gate. Measured (paired A/B, goldilocks mistral-small-24b, n=6): gen-tests 3/6 vs 0/6 for the coverage grade (Δ +50pp, 3-0 discordant, CI [-6.1%,+50%] — not significant at n=6, reported as-is). Honest caveat: the win is a resolve-rate win, not the false-positive reduction I pre-registered — that prediction was retracted.
• Cross-provider, decomposed envelope auditing (M18-2). The spot-check auditor grades three named failure classes separately (invented / dropped / contradiction) and can run on a distinct model (orchestrate --verify-model) so a model doesn't grade its own family's output.
• Declared write-region for the file-writers (M18-3). solve --write-region "src/**,*.py" makes write_file/edit_file/apply_patch refuse a write outside the declared globs — blocking an injected instruction from rewriting an unrelated file.
• Measured skill-lifecycle loop (M18-4). A provisional tier: new skills run on probation, then chimera skills-lifecycle [--apply] auto-promotes the proven ones and demotes regressions — from measured stats, never self-report.
• PROBE — best-arm identification with a cheap-proxy control variate (M18-5). chimera probe-select picks the best model/config with fewer expensive grades the better a cheap proxy correlates, and stays unbiased when it doesn't.

Gate: ruff clean, mypy clean, 1250+ tests. Still alpha.

PyPI: `pip install chimera-agent==0.15.0`