[linnnn89]

Change

Multi-keyword local-text context requests previously reread the workspace once per keyword. Version 0.17.1 shares one bounded scan for up to four extracted keywords by default, while retaining query order, diagnostics, file/line scopes and per-query limits. Any capped scan or query reaching 500 matches falls back to sequential queries under the existing request deadline. Single-keyword requests and providers without batching retain their existing path.

Add optional internal phase/work counters and a model-free paired MCP benchmark with an explicit sequential baseline. Update the managed Skill with multi-name request guidance, fallback costs and incomplete-result handling, and align package/native component versions. No new MCP arguments, dependencies, resident service or persistent source snapshot.

Validation

• 64 focused local checks passed, including batching defaults, edits/deletion, capped fallback, encoding gaps, cancellation and skill synchronization.
• Full local npm run check: 504 passed, 0 failed, 1 optional fixed-fixture skip; 0.17.1 delivery matched. Fresh production stdio acceptance on the real project passed all eight existing source-context scenarios. Final-default paired synthetic benchmark passed all 12 pairs (24 measured calls), with unchanged input/source hashes.
• Prior read-only real-project comparison: five rounds, 140 measured calls, 70 response pairs all identical. Warm four-name median 1429 ms to 439 ms; a deliberately broad Async Save query hit 500 matches and regressed from 589 ms to 900 ms. Eight existing lexical gaps were preserved by both variants. These fixed-task measurements are not an end-to-end model or P95 claim.
• No paid model calls or application UI launch. Raw source responses, personal paths, benchmark artifacts and screenshots are not committed.

Merge only after all required checks pass for this head. Current desktop-client connections have not been reloaded; disk build/Skill deployment does not hot-update a running MCP process.