feat(atlas): release-truth doctor + error-atlas acceptance harness + loader hardening

[templetwo]

Dogfood of T2Helix v0.10.0 against its own repo. The mission premise held: the README lagged the implementation, and probing the atlas loader surfaced a real robustness bug. Both are now fixed, detectable, and locked by tests.

What changed

Loader hardening (lib/atlas.js, scripts/import-atlas.js)

• importAtlas() now wraps record() per-entry: a throw (e.g. content over the 100KB MAX_CONTENT_BYTES cap) is counted as a dropped row (fingerprint only, no content leak) instead of aborting the whole batch. Found by dogfood — a single 300KB line previously killed the entire load and stranded every valid entry after it. Restores the loader's "load the valid subset, report the rest" contract.
• New pure detectConflicts() + a conflicts field on the return: same-pattern / different-resolution entries are surfaced (CLI warns on stderr) instead of silently buried. Both still load (append-only — an error can have more than one valid fix); the resolution text is never exposed (secret-bearing), only pattern + fingerprints.

Acceptance matrix (test/atlas-acceptance.js, +15 tests)

• Explicit 12-case matrix: valid minimal / dense-with-extra-fields / duplicate-same-resolution / duplicate-conflicting-resolution / missing pattern / missing resolution / non-string pattern / non-string resolution / huge-entry-does-not-abort-the-batch / malformed JSON / empty-array+empty-file / mixed valid+invalid, plus two CLI end-to-end cases (huge → exit 1 partial, conflict → exit 0 + stderr).

Release-truth doctor (scripts/release-doctor.js, npm run release:doctor)

• Verifies README/docs prose against machine sources: package↔CHANGELOG version, README version currency, tool count (static TOOLS-registry parse — not a server spawn, which can truncate a one-shot tools/list on stdin-close), hook count, test count (live or non-brittle), Node range vs engines, atlas-documented, manifest-version. Exits nonzero on drift. First run caught: stale "203 tests total", README changelog lagging 6 versions, manifest has no version field.

Doc truth (README.md, package.json)

• Reworded the brittle hardcoded test count, fixed the npm test suite list, added a "Current release: v0.10.0 → CHANGELOG" pointer. release:doctor now reports STATUS: TRUE.

Receipts

• npm test — 268 passed, 0 failed (was 253; +15 acceptance). No existing behavior changed; the importAtlas return-shape addition is non-breaking (existing 18 import-atlas tests untouched).
• npm run release:doctor — STATUS: TRUE, exit 0.
• Adversarial diff review (18 agents across 4 dimensions, each finding independently verified) — 0 confirmed findings.

Open follow-up (not in this PR)

Give .claude-plugin/plugin.json a version field synced from package.json — the first concrete surface of "should there be a single machine-readable release-truth source?", and it upgrades the doctor's manifest↔package check from informational to a hard cross-check.

🤖 Generated with Claude Code