v0.6.0

Added

• ai-forge-create: Phase 0 dedup check and Phase 1b baseline probe — measure what the agent does without the artifact before drafting, so observed failures drive the body instead of asserted ones. New references/baseline-probe.md carries the protocol and a failure-form matching table (guidance shape follows failure type).
• ai-forge-create: references/drafting-craft.md for leading words, degrees of freedom, completion criteria, and the constraint format.
• ai-forge-eval: Phase 0 loads or writes a persisted evals/evals.json, so a suite survives the conversation and can be re-run as a regression test. New references/eval-suite.md documents the format, trigger protocol, and pressure escalation.
• ai-forge-eval: Phase 3b human review step with a self-contained assets/eval-review.html template; Phase 6 trigger check reporting recall and precision against a 60/40 train/held-out split.
• ai-forge-eval: assertion discrimination — classifies each assertion by whether it distinguishes the with-artifact arm from baseline, so a suite of assertions that pass either way stops reading as a healthy pass rate.
• ai-forge-audit: Phase 0 ecosystem coherence pass and scripts/check-ecosystem.cjs — trigger collisions, duplicate and shadowed names, stale links, orphaned references, dated model pins, and word budgets across a whole roster.
• ai-forge-review: refutation pass with a severity-ordered charge and blocker/should-fix/note labels.
• ai-forge-update: references/iteration-guide.md with iterate-vs-redesign sizing and a signal-to-fix table; conversion now runs a field-disposition report and routes judgment calls through ai-forge-apply.
• ai-forge-create: validate-metadata.cjs gains --target (per-platform disposition report), --when-to-use, a dated-model-ID check, and the 1,536-char combined listing cap.
• Documentation for the full Claude Code SKILL.md frontmatter field set, skill-scoped hooks (including the exit-code-1-does-not-block trap), the script-vs-prose decision, and domain-variant reference organization.

Changed

• ai-forge-judge: U3 renamed to Constraint Quality and rescored. It now grades whether constraints are justified, explained, and load-bearing rather than whether a NEVER list exists — a constraint carried by explained reasoning scores higher than the same constraint asserted, and prohibition bloat is penalized.
• Reference-file TOC threshold raised from 100 to 300 lines, matching the spec.
• Skill body budget stated as 200 lines and ~1,800 words, set to bind at the same point.

Fixed

• Three shipped skills failed this pack's own metadata validator (unescaped colon in description).
• ai-forge-review/assets/AI-SPEC-TEMPLATE.md carried source-project specifics; generalized, and given requirement traceability and a decisions table.
• AGENTS.md leakage check widened from a single project-name grep to cover ticket IDs, absolute home paths, and brand names, plus self-conformance checks.