fix(search): scope rerank cache fallback to the scoring query

[andrewloux]

QMD prepends caller-supplied intent to the query before reranking. The compatibility fallback can reuse a score for the bare query, bypassing scoring with that intent.

For example, a caller reranks a passage from doc.md with query cache and intent configuration commands. QMD's scoring input is:

configuration commands

cache

Assume a previous file-scoped entry holds a score for doc.md, this passage, the same model, and bare cache. The current key has no entry for the input above.

• Before: QMD returns the bare-query score. The model receives zero calls, so the supplied intent never reaches scoring.
• After: QMD skips that fallback, scores the passage with the input above, and caches the new score. Repeating the request reuses that score.

The fallback now requires the formatted query to equal the bare query. Missing and empty intents retain compatibility reuse; current-key entries retain priority. formatRerankQuery() extracts the existing formatting; production rerank() calls it to build the scoring input.

The regression installs LegacyCacheReranker, a model fake that records the already-formatted query and returns 0.99. The test seeds a file-scoped score of 0.11, then checks one intent-aware model call, repeat reuse, and bare/empty-intent compatibility. All 324 store tests pass on Node and Bun (13 integration tests skipped per runtime). Types, lint, and compiled-package smoke pass.

bun run test:node test/store.test.ts
bun run test:bun test/store.test.ts
bun run test:types
bun run lint
bun run test:package