bug: MCP search returns internal error after lock -- possible cold-start race

[robotrocketscience]

Bug

Immediately after running agentmemory lock via CLI, two consecutive mcp__agentmemory__search calls returned [Tool result missing due to internal error]. A subsequent mcp__agentmemory__status call succeeded, and search calls after that also succeeded.

Reproduction

1. Lock a new belief: uv run agentmemory lock "some text"
2. Immediately call mcp__agentmemory__search via MCP
3. First 1-2 calls may fail with internal error
4. Subsequent calls succeed

Hypothesis

Possible causes:

• WAL checkpoint race: CLI lock writes to DB, MCP server reads stale WAL state
• Connection pool: MCP server connection doesn't see new data until next checkpoint
• Transient: may be unrelated to the lock operation (network/IPC hiccup)

Evidence

Only 2 data points (single occurrence). Could be transient. Filing to track in case it recurs.

Impact

Low -- workaround is to retry. But if this happens during hook injection, beliefs could be silently missed.