feat: remove the release-blocker concept

[feng-shiplight]

Summary

Removes the "Release blockers" metric end to end. It was structurally always 0: the blocked gap category it counted could only be produced by a standalone-word match on "blocked" in a narrow text blob (evidence type/path/url/command, residual risk, proof-gap next step), and its other trigger — assessmentStatus === "blocked" — was unreachable, since deriveExpectationAssessment only ever returns NOT COVERED / MANUAL / IMPLICIT / COVERED / PARTIAL. Nothing in the pipeline writes the word, so the tile read as "we checked and found none" while never having checked.

Measured against the two projects consuming this engine: 0 of 326 real gap records across shipyard (54 targets, 297 records) and monots (4 targets, 29 records) would have been classified blocked.

It also had teeth where it should not have:

• rollupStatusFor set an entire project's status to BLOCKED off that keyword match.
• riskIndicatorsFor added gapCounts.blocked to releaseRiskCounts.blockers — the same records counted twice, so one blocked gap rendered as "2 blockers".

Removed: the blocked gap category and both classifier branches, the blocked-gaps analytics metric, ReleaseRiskSummary.blockers, WorkspaceReleaseRiskCounts.blockers, WorkspaceAttentionCounts.blocked, the BLOCKED rollup status, the "Blocked" owner risk badge, the dashboard tile, and the feature-table column. The task status blocked is a separate lifecycle vocabulary and is untouched.

Regression found in review and fixed. Deleting the status.includes("BLOCK") branch from riskBadgeFor looked safe because the structured path can never produce that status — but the markdown-fallback path passes the Result cell verbatim from the scanned project, so a row reading BLOCKED | HIGH | None badged as Covered. The root cause was broader than the missing branch: an evidenceConfidence === "HIGH" claim promoted any unrecognized status to Covered, and the audited project writes that text. Confidence now settles only a row that states no result of its own; a stated non-passing result maps to Gap.

Breaking change / versioning

Exported surface narrowed, so the packages that changed are bumped per CONTRIBUTING.md and the fada5bd precedent:

• @shiplightai/quality-map 0.1.0 → 0.2.0 (GAP_CATEGORIES, GapCategory, the schema enum)
• @shiplightai/quality-core 0.2.0 → 0.3.0 (removed read-model fields and badge member)

quality-tools stays at 0.3.2 by design: it re-exports GAP_CATEGORIES, but moving it outside ^0.3.0 requires migrating the agent-skill pin in the same change, and scripts/check-quality-skill.sh npx-fetches the published package — so that pin cannot move before a release exists. Flagging for the release step.

Impact on consumers is bounded: accepted_gaps: ["blocked"] still parses and becomes an inert no-op (validate.ts checks array-of-strings only, not membership), so no stored artifact breaks. The one real consumer is shipyard's z.enum(GAP_CATEGORIES) at apps/web/src/app/api/quality-center/feature/quality-map/schema.ts, and narrowing it matches that file's stated purpose.

Guards added

• Pin the literal published gap-category vocabulary.
• Assert gapCategoryOrder covers GAP_CATEGORIES — an omission there silently drops a category from every gap record and is not a type error.
• Pin the riskBadgeFor ladder over the statuses markdown reports actually carry.

Test plan

• pnpm build clean
• pnpm typecheck clean (5 projects)
• pnpm test — 382 passing, 69 files (6 new)
• pnpm --filter @shiplightai/quality-tools check:size — packed −0.04%, unpacked −0.01%
• pnpm --filter @shiplightai/quality-ui check:size — packed −0.16%, unpacked −0.23%
• scripts/check-quality-skill.sh passes
• Five-lens local review (correctness, security, performance, test coverage, conventions); all MEDIUM+ findings resolved
• docs/assets/quality-explorer-overview.png still shows the removed tile — needs a fresh screenshot (caption already updated)

🤖 Generated with Claude Code