perf: replace Mutex with RwLock for concurrent read requests

[ajianaz]

Summary

HTTP server serializes ALL requests via Mutex<Uteke> (main.rs:571-578). Read-only requests (recall, search, list, stats) blocked while writes process.

With 17K+ memories, a single /recall can take 50-100ms. During this time no other requests can be processed.

Proposed Fix

Replace Mutex<Uteke> with RwLock<Uteke>:

• Read endpoints (recall, search, list, stats, health, namespaces, graph, room/, doc/, context) = read()
• Write endpoints (remember, forget, room/create, room/delete, doc/create, doc/move, doc/delete) = write()

Acceptance Criteria

Functional

• Replace Mutex<Uteke> with RwLock<Uteke>
• Read endpoints use read() — concurrent access
• Write endpoints use write() — exclusive access

Verification (CMO can run without reading code)

• Start server → curl GET /health returns 200 (basic sanity)
• 10 simultaneous curl POST /recall requests → all return 200 within 500ms each (no serialization)
• During a POST /recall, another POST /recall completes without waiting (concurrent reads)
• During a POST /remember (write), a POST /recall (read) can still complete (read doesn't block on write)
• During a POST /remember, a second POST /remember waits until first completes (writes are exclusive)
• No crashes or panics under 50 concurrent mixed read/write requests
• cargo test passes (no data race warnings from #[warn(dead_code)])

Tests

• Unit: verify RwLock compiles and passes existing tests
• Integration: concurrent recall benchmark (10 goroutines/threads)

Priority

Medium — performance bottleneck under load.

[ajianaz]

Duplicate of #674 (same scope, newer issue with cleaner AC)