fix(cli): speed up recall and separate hook deadlines

[orga-agent]

Recall could exceed its one-second deadline and misleadingly report that a healthy daemon was unavailable. Use an extraction-first SQLite join to avoid reading large note rows for nonmatches, give human recall five seconds and hook/MCP recall three seconds, and report response timeouts with a distinct timeout code. Remove the competing one-second recall IPC response limit.

Existing matching, scope, ordering, deduplication, output bounds and old-hook preservation remain intact. Installer output and documentation use the new deadline; updated daemon execution and explicit hook reinstall are needed to activate changes on an existing machine.

Validation: workspace tests/build/check, all-target/all-feature Clippy, format and cargo deny passed. Eighteen focused recall tests include deterministic delayed MCP success and 3/5-second deadline cases. An isolated real-executable/fake-daemon smoke confirmed hook success at 2s, human success at 4s, and timeout at 3s/5s. Standards and spec reviews passed.

Synthetic actual PowerSync fixture (2,200 notes, 22,000 extractions, 100KB bodies): sparse/no-hit optimized backend ~67–70ms; dense ~306ms. These are comparative measurements, not a production latency SLA. No live installation, daemon restart or hook changes were performed.