Add an Exa search-to-classification example

[RogueTex]

Exa retrieval results need a small adapter before they can be passed to jev_classify and joined back to their source URLs. This adds a runnable source-checkout example that accepts a saved Exa response or performs a direct search, deduplicates URLs, bounds excerpts, and classifies in batches of 64.

The application preserves the full Jev judgment and explicitly routes empty excerpts, invalid responses, and the manual_review class to review. It includes a synthetic offline fixture, four adapter tests, and instructions for evaluating a fixed labeled sample. No new MCP tools or dependencies are added.

Validation:

• 243 offline tests pass; TypeScript checking and git diff --check pass.
• All 19 live TypeSafe integration tests pass.
• Three live runs on ten saved Exa excerpts completed in 267–292 ms per classification batch, with 8–9/10 agreement against agent-assigned reference labels. Every disagreement was routed to review. This is a smoke test, not a validated accuracy benchmark; jev-latest was used.
• Synthetic live controls exercised library, tutorial, and empty-excerpt review routing.
• Exa retrieval was performed through the connected Exa app; the standalone REST search branch has not been live-tested with an Exa API key.

[jkudish]

Thank you for this — a clean search → classify pipeline: URL-preserving, fragment-deduped, batched to the tool's limits, review-routed, with offline adapter tests. Merged with one small amendment: the example pointer is folded into the classify bullets.