Added
- A project with no browser-observable surface can declare its QA method once, so the QA step proceeds by that method automatically instead of asking the same override question on every feature.
- A release for such a project now describes QA as having run by that declared method — a settled project fact — rather than reporting a one-off exception granted for that release.
Changed
- A later review or QA pass can now cite a fact an earlier pass in the same feature already established, instead of re-deriving it from scratch — except when verifying a fix to that exact fact, verifying a required acceptance criterion, when the earlier pass flagged the fact as unverified, or when there's a concrete reason to doubt it; those cases still get a fresh check.
- On a project with no such declaration (every project's default today, unchanged), or where the declaration is missing a piece or names a method the session genuinely can't perform, the QA step behaves exactly as before — same question, nothing silently skipped.
Fixed
- The QA-method override that had to be renegotiated by hand on four consecutive features no longer needs repeating once declared.
Stated plainly, not hidden: four of the files carrying this release's QA-declaration wording — skills/demo-day/SKILL.md, skills/spark/SKILL.md, skills/go-live/SKILL.md, agents/release-manager.md — remain not-verified-live in this loop's own testing; the nested session needed to exercise them could not authenticate in this environment, reproduced in all 8 QA rounds. Recorded as an accepted, environment-caused gap, not a defect, and still true as of this release. This change makes no claim about line, token or cost savings, and no claim that the result generalizes beyond this project.
Full gate trail and evidence: .spark/right-sizing/ (spec, plan, review, qa, release, evidence). Merged via #31.