fix(mcp): enforce session pool capacity during promotion

[XIIRUAN]

Why

MCPSessionPool checks its LRU capacity before starting session creation. Because initialization awaits, different keys can all observe spare capacity and enter _inflight, then each promote into _entries without another capacity check. This can leave the persistent-session registry above MAX_SESSIONS, retaining extra subprocesses and connections.

What changed

• Recheck and enforce LRU capacity atomically when an initialized session is promoted.
• Close promotion-time eviction victims outside the registry lock through their owner-task lifecycle.
• Add a deterministic regression test for two distinct keys initializing concurrently with a one-session capacity.
• Document the promotion-time capacity invariant.

Surface area

• Frontend UI — page / component / setting / interaction under frontend/
• Backend API — endpoint / SSE event / request-response shape under backend/app
• Agents / LangGraph — agent node, graph wiring, langgraph.json, or prompt change
• Sandbox — docker/ or sandboxed execution
• Skills — change under skills/
• Dependencies — new/upgraded entry in backend/pyproject.toml or frontend/package.json
• Default behavior change — the configured session-pool capacity is now enforced during concurrent promotion
• Docs / tests / CI only — no runtime behavior change

Screenshots / Recording

N/A — backend-only concurrency fix.

Bug fix verification

• Test path: backend/tests/test_mcp_session_pool.py::test_concurrent_distinct_sessions_respect_capacity
• Red on main, green on this branch: yes.
• On main, two distinct concurrent initializations with MAX_SESSIONS = 1 leave two registered sessions. This branch keeps one registered session and closes the evicted owner session.

Validation

• cd backend && uv run pytest tests/test_mcp_session_pool.py::test_concurrent_distinct_sessions_respect_capacity -q
• cd backend && uv run pytest tests/test_mcp_session_pool.py -q -k "not test_session_pool_tool_pins_cwd_and_temp_env and not test_session_pool_tool_does_not_override_explicit_tmpdir" — 40 passed, 2 deselected (existing Windows path/permission assertions)
• cd backend && uv run ruff check .
• cd backend && uv run ruff format --check .
• python scripts/check_agent_guidance.py
• Full non-live backend suite was also run on Windows: 11,794 passed, 117 skipped; 140 platform-dependent failures were unrelated to this change.

AI assistance

Tool(s) used: Codex

How you used it: Used to inspect the concurrency path, draft the regression test and implementation, and run validation. I reviewed and verified the final diff.

• I've read and understand every line of this change and take responsibility for it — it's not unreviewed AI output.

[CLAassistant]

All committers have signed the CLA.