lab: wfbench — validation as judge evidence, namespace as run input

[Evanfeenstra]

Two small pieces toward the Hive "run in vein" toggle.

• Validation as evidence. wfbench-run now runs meta/validate-workflow on the produced YAML and wfbench/build-materials inlines the result as a VALIDATION material (ok, errors, warnings, summary). Structural criteria ("workflow is structurally valid") get judged on the engine's own check plus the rerun status instead of on the judge's reading of the YAML. Not a gate: a candidate that validates poorly is still launched and judged.
• Namespace as input. input.namespace || params.namespace on every roster/record write. Hive already upserts the EvalSet/EvalRequirement roster through jarvis before dispatch; passing that partition lets vein's writes merge onto the same nodes by id rather than creating a second roster under wfbench.

Verification: mcp tsc clean; npx tsx src/lab/wfbench/smoke.ts green (both workflows static-validate, build-materials asserts the VALIDATION section).

[Evanfeenstra]

Live run 1788554443025 on this branch with input.namespace: wfbench-ns-test: 4/4, candidate v4. Namespace registered and the whole roster + record chain landed under wfbench-ns-test (1 EvalSet, 4 EvalRequirement, 1 EvalTrigger on HAS_BASELINE_TRIGGER since that partition had none, 1 EvalTriggerOutput, 4 CriterionResult) — the wfbench partition from earlier runs untouched. The judge materials carried ### VALIDATION: static validation: ok (0 error(s), 0 warning(s)) from meta/validate-workflow.