fix(review): make review_cad's fitness and mechanism verdicts agree on joint-primitive assemblies

[w1ne]

The defect (KC-04)

review_cad on a sound joint-primitive assembly returned this, in a single response:

"ok": false,  "mechanism": "broken",
"fitness": { "functional": true, "repairMode": "none",
             "repairDirective": "No repair needed. Preserve the current topology…" }
"mechanismFailures": [ mechanism.orphan-part —
   "Part 'arm' is not reachable from the mate graph (no mate edge connects it to 'base'…)" ]

An agent trusting fitness ships a broken mechanism; one trusting mechanism abandons a sound one. Both cannot be right.

Root cause — three places knew about only one of the two conventions

1. checkOrphanParts (src/modeling/runtime/mechanismTruth.ts) built adjacency from arm.__mates() alone. A joint-primitive assembly has zero mates by construction, so every part after parts[0] was reported orphaned and the verdict was broken.
2. summarizeMechanismFitness never saw mechanism-truth failures at all — it reads the validator/envelope streams only, and independently concluded functional: true. reviewPipeline folded the verdict into ok but left fitness untouched. That is why one response disagreed with itself.
3. validateAssemblyWithMates early-returned at arm.__mates().length === 0, reading "no mates" as "nothing left to validate" and disabling every gate below for joint-primitive assemblies.

After: "ok": true, "mechanism": "real", "mechanismFailures": [], fitness unchanged and now consistent.

The change

The connectivity walk adds joint-primitive edges (joints address parts by FeatureId, so a name map is built) and connect: { to } placement edges, which the v0.5 validator has always treated as structural. Fitness now receives mechanism-truth failures, so the contradiction is structurally impossible rather than fixed for one case — error-severity truth failures become blocking reasons, while warn carriers (mechanism.sweep-budget-exceeded → unverified) are deliberately not folded, so "couldn't verify" never reads as "broken". The early return became a solver-only skip; every mate-driven gate already loops arm.__mates() and is inert on an empty set, so the no-mates cost is unchanged, while validateWorkspaceReachability now actually fires instead of being dropped.

Registry hints no longer tell an agent to "add a mate" when a joint primitive is the right fix. No new diagnostic code, so the 248 catalogue counts are untouched.

Safety

Joint-primitive FK semantics are untouched — origin remains the parent→child frame offset and forwardKinematics still composes T(o)·M. git diff against forwardKinematics.ts and solver.ts is empty. The robotics tests that caught an earlier bad attempt at this area stay green: tests/unit/assemblies/forwardKinematics.test.ts and tests/unit/kinematic/checkReachableMultiDof.test.ts → 5 passed.

Verification (re-run independently in a checkout with generated assets)

• tsc --noEmit → clean.
• src/modeling src/kinematic src/agent → 1183 passed, 0 failed.
• tests/ → 3370 passed, 0 test failures (one file fails to collect: the known lightningcss.darwin-arm64.node darwin-only issue).
• Negative control: reverting mechanismTruth.ts alone → 6 fail, including the orphan diagnostic reappearing verbatim and expected 'broken' to be 'real'; restored → 16 pass. The "genuinely broken" test fails differently under revert (expected 2 orphans, wanted 1), confirming it discriminates rather than merely passing.

Deliberately left undone

• No joint-primitive analogue of assembly.mate.limit-missing — a .revolute() with no limitsDeg is still invisible to envelope review. Needs a new diagnostic code and the 248→249 bumps; larger than the contradiction fix. The structural blocker is removed, so future gates apply automatically.
• mechanicalPlausibility.ts has a third connectivity walk (assembly.mechanical.part-disconnected, reading arm.__joints()). Not implicated in the reported output, but worth consolidating so three walks don't drift apart again.
• The 'fixed' joint kind is handled generically but untested — arm.fixed() was removed in G0, so there is no public API to create one.