fix(mcp): cleanup cancelled servers during connect failure

[hsusul]

Summary

When MCPServerManager is configured with suppress_cancelled_error=False, a CancelledError raised during connect() was re-raised without recording the server as failed. Sequential connect_all() failure cleanup only walks connected_servers + failed_servers, so the cancelled server was skipped. Under the recommended async with MCPServerManager(...) pattern, __aexit__ never runs if __aenter__ raises, so there is no later chance to call cleanup_all() — leaving transports / subprocesses / sessions open.

This aligns CancelledError handling with the existing BaseException path: always _record_failure(...) first, then re-raise when suppression is disabled.

Root cause

In _attempt_connect:

except asyncio.CancelledError as exc:
    if not self.suppress_cancelled_error:
        raise  # never recorded → never cleaned by connect_all failure path
    self._record_failure(server, exc, phase="connect")

Neighboring BaseException handling already records before re-raising, and the lifecycle reference requires partial connection failures to close every context already entered.

Reproduction

async with MCPServerManager([ok_server, cancelled_server], suppress_cancelled_error=False):
    ...
# CancelledError propagates; cancelled_server.cleanup() was never called

Before this change: cancelled server stays resource_open=True, cleanup_calls=0.
After this change: connect_all() failure cleanup runs cleanup() on the cancelled server.

Solution

Record the cancelled server via _record_failure before optionally re-raising, so connect_all()'s sequential failure cleanup includes it (same contract as fatal BaseException connect failures).

Test plan

• Strengthened test_manager_cleanup_runs_on_cancelled_error_during_connect to assert the cancelled server is recorded and cleaned by connect_all() itself
• Added test_manager_async_with_cleans_cancelled_server_when_unsuppressed covering the async with leak path
• Ran .agents/skills/code-change-verification/scripts/run.sh (format, lint, typecheck, full tests) — all passed
• git diff --check clean

Issue number

N/A — found via code-path review; no existing issue/PR covering this exact gap (nearby MCP cleanup work targets stdio semaphore / exit-stack reconnect issues, not this manager cancellation recording path).

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

Compatibility and risk

• Default suppress_cancelled_error=True behavior is unchanged (still records and does not re-raise).
• Only the unsuppressed path gains recording-before-raise, matching documented cleanup expectations.
• No public API shape changes.

Validation

make format / make lint — pass
make typecheck — pass
make tests — pass
code-change-verification — all commands passed