feat(stores): run-scoped knowledge stores with lineage-chain reads, and the claim-evidence contract

[drewstone]

What

Two additive modules, no changes to existing surfaces:

• run-scoped.ts — one knowledge store per run, isolated by construction, inheriting only through ancestry declared at init. Chain reads return every visible page labeled here / inherited:<runId> / shared, because "this run established it", "an earlier attempt believed it", and "the lab curates it" are three different provenance claims a reader must weigh differently. Lineage cycles fail loud; absent ancestry ends the chain quietly.
• claim-evidence.ts — evidence rungs (1–5) as a first-class type; a record-time refusal of rung ≥4 claims that carry no re-executable check; and the calibrated verdict lattice verified / silent-check / contradicted / unrunnable / uncheckable as a pure function, so every grader shares one semantics while execution stays with the caller.

Why here

Grown in the discovery-lab campaign and recorded there (docs/15-adoption-decisions.md) as debt owed to this package — the lab's standing rule is that a capability its tools grow must migrate upstream rather than live as a local fork.

Each rule was paid for before it was written:

• Two false "measured" claims crossed run boundaries through a shared store and became standing instructions in live agent profiles before anyone re-derived them. Isolation-with-labeled-inheritance is the countermeasure.
• A run produced 28 true claims that scored zero because nothing carried a check a grader could re-execute; a sibling lab's cells once produced 6 false certifications in 17 deliveries with no agent lying. assertGradeableEvidence refuses the shape at record time.
• The verdict lattice is calibrated, not designed: its naive predecessor graded three TRUE claims as refuted on first contact (silent asserts print nothing, so an expected-output match can never succeed), and could not tell a missing input file from a failed assertion. The tests encode exactly those cases.

Verification

• pnpm lint clean, pnpm typecheck clean, pnpm build clean
• 13 new tests: isolation, origin labeling, sibling non-inheritance, shared-store reads, absent-lineage termination, cycle refusal, and the four miscalibration cases

🤖 Generated with Claude Code

[tangletools]

✅ Auto-approved drewstone PR — 45669a41

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-06T09:43:34Z}