fix(search): rank English term boundaries and guide direct API searches

[humandebri]

Purpose

SSE receives strong title/path scores inside unrelated words such as precompressed and WebAssembly, displacing relevant body evidence. Direct ICP callers also need guidance for turning English questions into useful existing API searches.

Changes

• Require case-insensitive ASCII word, camelCase or acronym boundaries for strong title/path boosts; keep dotted identifiers intact.
• Retain incidental substring candidates at weak existing path scores, including when FTS title weighting would otherwise promote them. Preserve exact-path/basename/prefix priorities, non-ASCII matching, short-term rules, candidate queries, caps and score constants.
• Add English keyword-search guidance to the existing shared JSON rendered by get_api_doc: up to three focused searches, scoped candidates, deduplication and body verification. MCP search code and tool descriptors are unchanged.
• Reuse the existing 1,154-document / 46-case MCP fixture in real store tests, with search/retrieval hit parity and per-case English recall guards.

Verification

• Confirmed the 32-distractor SSE regression fails before the fix and passes after it.
• Paired real-store replay: English positive hit@20 remains 34/35, with no previously successful English case lost. Bare sse finds the timeout gold at rank 4, previously outside the top 20. The six documented keyword examples find gold at ranks 1 / 1 / 3 / 6 / 2 / 1; original labels are unchanged.
• Rust store/canister suites, MCP Node/Vitest/workerd suites, shared-core tests, typecheck, generated Candid consistency, formatting and diff checks passed. Canister build and MCP dry-run build passed.
• Existing recorded MCP/Jev replay also passes (holdout hit@10 19/20); this frozen-response replay is separate from the changed-store measurements above.
• Independent review found no blocking issues. Added individual-case recall and short-term/Unicode boundary checks from review.

Impact

Ranking affects existing MCP, CLI, browser and direct ICP callers. Relevant bodies with incidental filenames can also receive weaker scores; frozen-corpus hit@20 is preserved, while the original general-SSE question changes from rank 1 to 3. Natural-language parsing and Japanese short-word handling are unchanged.

No public API/Candid/authentication/authorization/billing/schema changes, Worker changes or new operational document. No deployment performed. Live ICP/CLI comparison requires a separate canister upgrade; no Worker deployment is needed.