fix(search): bound expansion fanout and honor cancellation

[rszrszrsz]

Dependency

Draft: depends on #912. Until #912 merges, GitHub will also show its runtime commits in this PR. After it merges, this diff reduces to the search-policy and cancellation changes below.

Summary

• bound query-expansion fanout once at the store boundary
• apply identical sanitization to fresh model output and legacy cached expansions
• exclude normalized restatements of the original query before consuming caps
• propagate caller and shutdown cancellation through expansion, embedding, vector fanout, and reranking
• make searches reject after abort instead of returning empty or partial success

Expansion policy

The policy preserves source order and keeps:

• at most 6 variants total
• at most 2 variants per type (lex, vec, hyde)
• unique (type, normalized text) pairs

Normalization trims, collapses whitespace, and compares case-insensitively. Unknown types and empty text are rejected. The same policy runs after both cache reads and fresh generation, before any search fanout.

Cancellation

Search observes the existing PR1 session abort signal plus an optional caller signal. Checks sit around the expensive boundaries:

• before and after expansion
• before and after embedding
• before every vector lookup and after the final lookup
• before and after reranking
• before returning RRF-only, empty, or partial results

vectorSearchQuery resolves the effective LLM, so stores without an attached LLM still observe shutdown from the default instance.

Verification

Before rebasing onto the latest main:

• full Vitest: 1,255 passed, 78 skipped
• full Bun: 1,255 passed, 87 skipped
• normal build and CLI help/version passed

After stacking on the rebased #912 head:

• npm run lint
• npm run test:types
• 276 focused expansion-policy, cache, and cancellation tests passed; 13 existing tests skipped

No production QMD collection or index is used by these tests.