test: measure first-attempt workflow conformance - #142

Stack

Depends on #140; targets feat/lifecycle-inspection. Review after the testing-contract and lifecycle layers. No merge or auto-merge requested.

Why this layer

The historical routing evaluation documented repaired workflow-schema attempts despite final routing success. The current unchanged-copy baseline did not reproduce that problem: one requested model/configuration, six existing fixtures, all four checklist items passing and zero repairs. Compact tool copy, validator, and scenario prompts remain unchanged; this PR does not claim a guidance fix or model-performance improvement.

Changes

• Report firstAttemptValid and firstAttemptSemantic separately from the existing four routing judgments. Repaired first calls stay first-pass failures; accepted-but-wrong arguments remain distinct from schema validity.
• Preserve critical-success rules, accuracy, retries, timing, competing-controller rejection, and the full three-task first batch.
• Extract a pure evaluator and add deterministic expectations/non-mutation controls. Enable discovery and lint for the new MJS test/helper and existing affected harness.
• Capture bounded response provider/model/API metadata separately from the requested model string, plus Pi/Node versions. Unknown metadata stays unknown; omitted observations are counted.
• Validate the existing published workflow example with the production parser and LoopStore; test cadence/rework misuse and cadence's inability to choose an outcome.
• Document measurement boundaries, the bounded current baseline, and the evidence required before future copy tuning.