Audit and harden cancellable connection attempts

Context

PR #848 propagated its task AbortSignal through source and target ClustersClient.getClient calls. ClustersClient already attempts to interrupt a pending MongoClient.connect by closing the underlying client and maps an aborted connection to UserCancelledError.

That shared cancellation behavior has little direct test coverage, and cancellable callers are not systematically audited.

Follow-up

• Add focused tests for cancellation before a new connection, during a new connection, and during cached-client reconnection.
• Verify abort-listener cleanup and the connect/abort race around a connection that has just succeeded.
• Audit interactive ClustersClient.getClient call sites and pass an AbortSignal where the UI or task exposes cancellation.
• Define consistent cancellation classification across UserCancelledError, AbortError, telemetry, and user-visible errors.
• Confirm aborting a cached reconnection does not leave the cached client unusable.

Acceptance criteria

• Shared connection cancellation paths have deterministic unit coverage.
• Cancellable call sites stop promptly without being reported as failures.
• Any intentionally non-cancellable call sites are documented.

Start by locating ClustersClient.getClient and its source and target callers, then read PR #848 and the UserCancelledError handling around MongoClient.connect. Add deterministic coverage for pre-connect, in-flight, cached reconnection, listener cleanup, and the connect/abort race; audit interactive callers and document intentionally non-cancellable ones. Done means cancellation is prompt, consistently classified, and does not corrupt cached clients.