Persist predict_physics claims: physics_specs + receipt re-verification

[ecto]

Mirrors the clearance-claim persistence pattern (check_clearance) for predict_physics, so structural limits re-verify via build_receipt / verify_receipt instead of rotting as one-off checks.

Stacked on #524 (needs ClaimBasis + predict_physics). Base will be retargeted to main once #524 merges.

What

• IR: PhysicsSpec (label, part/domain_box, loads, supports, material props, limits, fidelity) + Document.physics_specs, ts-rs exported (npm run ir:gen regenerated; gen-ir-types.mjs marks the field wire-optional like clearance_specs).
• predict_physics label arg: upserts the spec on the document; the part is persisted by resolved root id so it survives renames. Requires document_id and at least one limit.
• physicsReceiptClaims: re-solves each spec at its stored fidelity, emitting physics.static.<label>.{displacement,stress} claims with basis: predicted|verified (ClaimBasis from Verified design loop, Acts I–III: claim basis, trust boundary, predict_physics, macro library #524) and a re-runnable StoredPhysicsClaim payload in details.
• verifyPhysicsClaims: re-solves per claim and classifies Holds (identical result), Stale (geometry changed, limit still holds), Violated (limit exceeded). One solve per spec via cache when a spec carries both limits.
• Fail-closed: unresolvable part, load region off the structure, non-converged solve, or a tampered/missing details payload → unverifiable claim at build time, Violated at verify time. Never a silent pass.
• Wiring: build_receipt emits physics claims in both the mechanical-only and PCB branches; verify_receipt re-verifies them and folds into the worst-status rollup. Drive-by fix: the verify_receipt tool handler never passed the engine, so clearance re-verification via the live tool always errored — now passes ctx.engine.

Tests

• New physics-claims.test.ts: persistence/upsert, labeled claim emission with basis, Holds/Stale/Violated transitions, fail-closed paths (10 tests).
• Full @vcad/mcp suite: 805 passed. cargo test -p vcad-ir, workspace clippy -D warnings clean. Tool-surface fixture regenerated.
• No WASM artifacts committed (rebuilt locally for tests only).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vcad-mcp	Building	Preview, Comment	Jul 12, 2026 7:15pm

3 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
mecheval	Ignored		Jul 12, 2026 7:15pm
vcad	Ignored		Jul 12, 2026 7:15pm
vcad-docs	Ignored		Jul 12, 2026 7:15pm

[chojiai]

Choji review — Looks good

This delta addresses both prior findings: STALE_REL_EPS is now 1e-6 with an explanatory comment, and the epsilon-based stale classification is no longer a concern. The new changes wire physicsReceiptClaims into the place_order gate alongside clearanceReceiptClaims, add two gate tests (violated and holds paths), and fix the no-specs guard to check both clearance_specs and physics_specs. The logic is correct: the gate correctly refuses on failing or unverifiable physics claims, the no-specs path is properly updated, and the test fixtures look sound. No blockers found in the delta.

No findings — looks good.

---

Verified by Choji

Choji ran your change live — checks failed.

3 checks.

Claim	Result	Evidence
npm run test	Unverified — environment	sandbox failure, not your PR
npm run typecheck	Unverified — environment	sandbox failure, not your PR
cargo check	Unverified — environment	sandbox failure, not your PR

Checks marked environment failed in Choji's sandbox, not against your PR. They never block.

The ticket asked for persisted predict_physics claims (physics_specs + receipt re-verification); the PR implements exactly that — PhysicsSpec IR types, label-based upsert on the document, physicsReceiptClaims for build_receipt, verifyPhysicsClaims for verify_receipt, and the verify_receipt engine-pass fix. No live preview was run (this is an MCP/server-side change with no browser UI). Both npm test and npm typecheck failed solely because wasm-pack is absent in the environment (vcad-kernel-wasm#build: spawnSync wasm-pack ENOENT), which cascades through Turborepo and prevents downstream packages from running. The cargo check failure is also environmental: the tang crate path is missing from the sandbox. Neither failure reflects a defect in this PR's code. Tool-surface fixture is updated, changelog entry is present, and the gen-ir-types.mjs FORCE_OPTIONAL list is updated for physics_specs — all consistent with the stated intent.

No issues found in the running app.

---

Was this review useful? Rate the findings → — your thumbs up or down trains Choji on what's worth flagging.

Reviewed 215fcb8 · Choji keeps this comment up to date as you push.

[ecto]

Added e0eec02: the place_order M4 receipt gate now also re-solves persisted physics specs — failing/unverifiable physics.static claims refuse the debit fail-closed, and the no-specs 'unverified' path requires neither clearance nor physics specs. Two new gate tests; full mcp suite 807 passed.

[ecto]

Pushed 215fcb8: the two new order-gate tests timed out at vitest's 5 s default on CI (two mesh FEA solves each) — they now solve at resolution 12 with a 30 s allowance. Also took Choji's epsilon finding: STALE_REL_EPS raised 1e-9 → 1e-6 with a comment explaining that PCG converges to a tolerance, not bit-identical results.

🤖 Addressed by Claude Code

Dismissing prior approval to re-evaluate 215fcb8.