fix(core): remove pending response entries on request timeout and cancellation

[Lubaoshuai]

Fixes #1133

McpClientSession.sendRequest and McpServerSession.sendRequest left their pendingResponses entry in place when the downstream .timeout(...) fired or the caller cancelled — the entry was only removed on a response or a send failure. Since request IDs are unique per request, a timed-out request leaked its entry forever; KeepAliveScheduler pings amplify this into one leaked entry per interval on sessions with a dead peer.

Both methods now mirror the cleanup pattern already used by McpStreamableServerSession.McpStreamableServerSessionStream.sendRequest: doOnError + doOnCancel remove the entry after the timeout boundary. Removal is idempotent, and a late response for a removed id already takes the existing "unexpected response" path.

How verified

• New testRequestTimeoutRemovesPendingResponse in McpClientSessionTests: a request against a 50 ms-timeout session is let expire, then the pendingResponses map is asserted empty.
• ./mvnw -pl mcp-core test passes locally.

[soyeladice-svg]

The cleanup placement looks consistent with the streamable-session pattern, and the idempotent remove makes the duplicate send-error cleanup harmless.

One regression-coverage gap seems worth closing before merge: this PR changes both McpClientSession.sendRequest and McpServerSession.sendRequest, and the stated fix covers timeout + downstream cancellation, but the added test exercises only client timeout.

I would add at least:

1. a server-side timeout regression asserting McpServerSession.pendingResponses is empty after the request timeout; and
2. a cancellation regression where the subscriber cancels a request before a response arrives, then asserts the pending entry is gone.

If inexpensive, exercising cancellation on both client and server would pin the exact contract this PR is adding. The existing reflection-based assertion style is sufficient for this focused regression; no public test hook seems necessary.

That would protect against a future refactor preserving timeout cleanup while accidentally dropping the doOnCancel behavior, and would also prove the server-side change rather than relying on symmetry with the client.

[Lubaoshuai]

Thanks for the careful review — agreed on both points. Pushed in 6be0c53:

1. Server-side timeout regression — new McpServerSessionTests (there was no dedicated spec-level server session test class yet): a server session with a 50 ms request timeout against a transport that accepts the message but never responds; after the TimeoutException surfaces, pendingResponses is asserted empty.
2. Cancellation regression, both sides — McpServerSessionTests.testRequestCancellationRemovesPendingResponse and McpClientSessionTests.testRequestCancellationRemovesPendingResponse: subscribe while the response is still pending (entry present, asserted hasSize(1)), dispose, and assert the pending entry is removed by the doOnCancel handler on each session type.

All three use the same reflection-based pendingResponses assertion as the existing client timeout test. spring-javaformat:apply run on the new file, and the targeted mcp-core suite is green (15/13+2 tests).

[soyeladice-svg]

Verified the updated head. The new server-side timeout regression is present, and cancellation cleanup is now covered on both client and server with the pending map asserted non-empty before cancellation and empty after disposal.

That closes the coverage gap I raised. Thanks for adding the symmetric regressions.