[triggeredcode]

Summary

• add an offline, candidate-neutral routing and context experiment harness
• freeze development, calibration, and holdout cases by intent family
• record safety, routing latency, target recall, unrelated selection, and context reduction
• write local JSON and Markdown reports without changing Lucid runtime behavior

Baseline result

• current explicit-state routing: 34/34, zero high-confidence wrong direct actions, sub-millisecond p95
• current targeted context: 100% required-target recall and about 82% reduction on eligible large documents
• three small structured-page cases over-select neighbouring components, so the baseline does not pass the adoption gate yet

Validation

• typecheck
• lint
• formatting
• eval package: 52 passed
• full unit suite: 1,525 passed, 22 skipped
• import-impact diagnostics: clean

Closes #51