[ranjianghong526-hash]

Change Summary

Cancelling async with MCPClient(...) during parallel startup can leave servers that have already initialized connected. CancelledError bypasses the except Exception in __aenter__, and Python does not call __aexit__ when entry fails. An initialized stdio server continues answering tool calls after the cancelled entry.

• Run the existing cleanup on BaseException during entry, then re-raise the original exception. Transports still close in their existing owner tasks.
• Add real stdio MCP regressions for cancellation with one initialized/one initializing server and for normal two-server context exit.
• Include the existing configured YAPF formatter's adjustments in the touched client file; these do not change behavior.

Related issue number

Independently discovered and reproduced on main at 9b0244a43c6a15b7d3f52d0f1c50617ba00e7268; no existing issue is being closed.

Validation

• Windows, Python 3.12.14, and Linux, Python 3.12.15: the final related suite (tests/mcp and tests/config/test_mcp_resolver.py) passes all 36 tests on each platform.
• Pre-fix regression/control: 1 cancellation failure and 1 normal-exit pass. This was checked on untouched Windows upstream source, and rechecked with the final tests on Windows Python 3.10.19 and Linux (retaining formatting-only changes while restoring the original exception handler).
• Python 3.10.19: both final new regressions pass. The full related suite is 34 passed / 2 failed: the existing test_classify_mcp_failure and test_is_connection_error expect builtin TimeoutError to be the same as asyncio.TimeoutError. Both failures reproduce on the unmodified upstream client; their runtime implementation and tests match upstream byte-for-byte. They are outside this resource-ownership change. The new tests use the already-installed AnyIO timeout context to support Python 3.10.
• Independent real stdio probe: three runs before/after on each platform. Before, the initialized server still responds pong after failed entry. After, both owner tasks are finished and all session/task/shutdown registries are empty; cancellation still propagates.
• Applicable pre-commit hooks for the two changed files and git diff --check pass.
• pre-commit run --all-files was run, but is not green: the clean upstream control fails the same flake8/isort/yapf/trailing-whitespace/string/line-ending hooks. All 69 flake8 diagnostics are identical. Unrelated automatic fixes are excluded from this PR.

This is scoped validation, not a full repository or live model-provider test run. The existing ready-future cancellation diagnostic is reproducible before and after; this PR addresses connected-resource ownership only.

AI assistance

Codex assisted with source/duplication investigation, real-process reproduction, the minimal fix, tests, validation and this draft.

Checklist

• The pull request title is a good summary of the changes - it will be used in the changelog
• Unit tests for the changes exist
• Run pre-commit install and pre-commit run --all-files before git commit, and passed lint check. (Run; existing upstream failures described above.)
• Documentation reflects the changes where applicable (no API/configuration change)