fix(mcp): fail RPCs when transport sessions restart

[embwl0x]

Summary

• bind each managed MCP RPC to the ClientSession generation that started it
• signal that generation before reconnect/shutdown teardown, cancelling stale tool, resource, prompt, and dynamic-refresh calls immediately
• reject new calls during the narrow teardown window while allowing the replacement session to accept fresh calls

Root cause

A keepalive or lifecycle signal could rebuild an MCP transport while session.call_tool() was still awaiting the old SDK session. The server's session reference changed, but the in-flight coroutine had no ownership signal, so it remained attached to the old transport until the configured tool timeout (300 seconds by default).

The lifecycle now owns a per-session invalidation event. RPCs race their SDK operation against that event, cancel the stale operation when the transport is invalidated, and return a retryable error. The session reference remains available until the transport context exits, but its set invalidation event prevents new work from entering the stale generation.

Fixes #81995

Coverage

The sibling-path audit covers:

• tools/call
• resources/list and resources/read
• prompts/list and prompts/get
• notification-driven tools/list refresh
• stdio and HTTP/SSE lifecycle exits, including reconnect, recycle, shutdown, exception, and cancellation paths

Validation

• current-main reproduction: replacing a hung session did not wake its in-flight handler; it consumed the full configured timeout
• scripts/run_tests.sh tests/tools/test_mcp_session_invalidation.py — 8 passed
• scripts/run_tests.sh tests/tools/test_mcp_*.py — 459 passed across 51 files
• remaining MCP suites under tests/ — 240 passed across 24 files (with the lockfile-pinned acp extra for ACP suites)
• uv run ruff check . — passed
• python scripts/check-windows-footguns.py --all — passed, 939 files scanned
• python -m compileall -q tools/mcp_tool.py tests/tools/test_mcp_session_invalidation.py — passed
• ty diagnostics improved from 10 on exact base to 4 on this branch; no new diagnostics
• contribution publish gate and gitleaks — passed

No live MCP server, credentials, or user Hermes state was used; all lifecycle behavior is exercised with isolated fake transports.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #48069. Both changes address in-flight MCP RPCs that outlive a reconnect or shutdown and otherwise wait for the full tool timeout. #48069 is the earlier, broader repair.

[embwl0x]

Closing as a duplicate of #48069 after collaborator triage. That earlier PR owns the in-flight MCP request cancellation/reconnect repair. Our current-main implementation and broader resource, prompt, and tool-list regression coverage remain available in commit 684ebb6 if maintainers want to salvage any of those tests. Thanks for catching the overlap.