fix(recall): accept superseded claim history

[hjqcan]

Summary

• distinguish all canonical facts from active canonical facts during projection validation
• require claim status coverage only for active facts
• retain claim/status history for superseded facts without reporting it as orphaned
• keep truly deleted canonical facts subject to orphan cleanup

Runtime evidence

A Hermes GoodMemory v0.1.1 live smoke against published goodmemory@0.7.2 completed remember, new-session recall, workspace isolation, trace routing, revise, and forget, but emitted orphan_claim_status / orphan_claim validation warnings after the revised fact was superseded and the replacement was deleted.

The same live smoke against this branch completed the full loop with no projection-validation warning.

Verification

• bun test tests/unit/recall.claim-revision-visibility.test.ts — 13 pass
• bun test tests/unit/recall.claim-projection.test.ts — 32 pass
• bun run typecheck — pass
• Hermes live smoke against this branch — pass; new-session recall and workspace isolation true; revise/forget pass; warning absent
• git diff --check — pass
• bun test — 6296 pass, 62 skip, 1 fail across 6359 tests. The sole failure is the unrelated existing C3 physical-path assertion because this isolated clone lives under /tmp; the fixed-temporary-root guard fires before the overlap guard expected by tests/unit/codex-coding-effect.c3-pilot-options.test.ts.

[hjqcan]

CI comparison after the first PR run:

• PR run: https://github.com/hjqcan/GoodMemory/actions/runs/31239726523 — 6116 pass, 61 skip, 20 fail; Node 20/22/24 package boundaries all pass.
• latest main run: https://github.com/hjqcan/GoodMemory/actions/runs/31239231653 — 6115 pass, 61 skip, the same 20 failures; Node 20/22/24 package boundaries all pass.

Both CI gates fail only in tests/unit/run-v0-7-3-lifecycle-protection-gate.test.ts, beginning with claim recipe benchmark root must match the protected root. The PR adds one passing regression test and does not add a new failing test. I am leaving the check red and recording the baseline comparison rather than presenting the PR as fully green.