v2.22.0

Two review-quality disciplines: trace behavior claims to real code before writing them, and check what was built against what was actually required.

Added

• "Trace, Don't Assume" in the verification-before-completion guideline (so it applies everywhere, via CLAUDE.md / AGENTS.md): a claim about how the code currently behaves — a root cause, an existing mechanism, present behavior — must be traced to real code or a runtime observation before it's written into a spec, PR, commit, review, issue, or comment, and no illustrative example may be invented. Intended behavior a spec proposes as a requirement is exempt. Stops one unverified guess from seeding a whole ticket's context and tests on a false premise.
• Conformance & scope check in code-review — fetch the real requirement (the spec's goals, technical sections, and edge cases; un-superseded linked-issue criteria; or the task itself), then verdict each requirement requirement-down (Met / Partial / Unmet), and scope-check for implied requirements, unrequested extras, silent interpretations, and side effects. The diff shows what was built, never what was forgotten.

Changed

• implement-spec now walks the requirements before final verification — task checkboxes track tasks done, not requirements met, so a required behaviour no task mapped to is otherwise never caught. It verifies each requirement against real code, then runs the full quality gate last so that gate covers anything the walk changed.

Full Changelog: 2.21.0...2.22.0