[triggeredcode]

Closes #55

This expands the offline experiment to dataset v2 and records a local embedding comparison. It does not add a model or routing dependency to Lucid runtime.

Measured result on Apple M4 Pro:

• Current: 37/46; semantic scope 0%; vague near-miss scope 100%
• Nomic Embed Text, calibrated margin 0.04: 42/46; semantic 55.6%; near-miss 100%; challenge p50/p95 9.6/11.5 ms; cold max 175 ms
• Qwen3 Embedding 0.6B, same margin: 39/46; semantic 22.2%; near-miss 100%; challenge p50/p95 24.7/34.3 ms; cold max 415 ms
• All candidates kept lane/action accuracy at 100% and produced zero high-confidence wrong direct actions

Decision: Nomic is the leading local candidate, but 55.6% semantic scope accuracy is not enough to justify product integration. Keep it experiment-only.

Validation:

• pnpm format:check
• pnpm lint
• pnpm typecheck
• pnpm test (113 files passed, 5 skipped; 1529 tests passed, 22 skipped)
• Docdex import diagnostics clean