fix: add timeout + fail-open to recall search path

[kagura-agent]

Summary

Fixes #1452 — auto-recall can block gateway startup / first-turn path long enough to fail health checks when embedding or LLM calls are slow.

Problem

With auto-recall enabled and existing memories, the recall search path can block for 30-40 seconds when:

• Embedding model is slow to respond
• LLM skill relevance judgment hangs
• Hub memory search times out

This is enough to trip health checks and cause restart loops.

Solution

Add configurable timeout + fail-open semantics at three layers:

1. Recall engine (recall/engine.ts)

• withTimeout() helper: races any promise against a deadline, returns fallback on timeout
• embedder.embedQuery() wrapped with timeout → falls back to FTS-only search (no vector candidates)
• Hub memory embedding wrapped with timeout → skipped on timeout
• judgeSkillRelevance() (LLM call) wrapped with timeout → returns all candidates on timeout

2. Tool handler (tools/memory-search.ts)

• Top-level timeout on the entire memory_search handler
• Returns empty results with timedOut: true in meta on timeout
• Never throws — always returns a valid response shape

3. Configuration (types.ts, config.ts)

• New recall.timeoutMs option (default: 10000ms, 0 = no timeout)
• Operators can tune based on their model latency

Key Principles

• Fail-open: timeout/error → partial or empty results, never block
• No propagation: recall exceptions never reach gateway top level
• Startup independence: ready state doesn't depend on slow recall

Testing

• TypeScript typecheck passes (tsc --noEmit)
• No behavioral changes when recall completes within timeout
• Timeout only activates when operations exceed configured timeoutMs