fix(memory): coordinate concurrent sandbox sync

[yanpgwang]

Summary

• coordinate Session-shared tool operations with Docker resource refreshes using provider-owned shared/exclusive file locks
• let parallel Threads join the current filesystem wave while making Memory snapshot, optimistic PostgreSQL sync, and baseline refresh atomic with respect to active tools
• avoid taking the synchronization barrier for Sessions without Memory attachments and document per-Thread concurrency accurately

Why

Managed Agents allows Threads in one Session to run concurrently while sharing a sandbox. Previously, Memory reconciliation protected each individual snapshot/replace call, but not the complete snapshot -> PostgreSQL sync -> refresh transaction. One Thread could therefore refresh /mnt/memory from an old snapshot while another tool was still modifying the shared mount.

The PostgreSQL Memory repository already rejects stale per-file SHA baselines. This change closes the filesystem-side race without serializing independent tool execution.

Verification

• make verify
• make test-service
• make docs-check
• real Docker + PostgreSQL regression covering an active tool wave, skipped pre-tool refresh, blocking post-tool writeback, and persisted Memory contents

References

• https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration
• https://platform.claude.com/docs/en/api/beta/memory_stores/memories/update