[tarekgh]

Summary

Extend the test-only discovery timeout protection from #1943 to the remaining affected controlled-server connections.

• Add a shared TestMcpClient helper that disables the separate discovery probe timer while retaining the 60-second overall initialization deadline. Explicitly supplied options remain unchanged.
• Apply the defaults across HTTP/MRTR, Tasks, OAuth, hybrid session-mode, response-driven HTTP fallback, direct pipe/stdio, and AOT test connections. Apply defaults before test configuration callbacks so explicit overrides still work.
• Pin the legacy-only elicitation defaults mock to its supported protocol version and reject unexpected requests instead of silently waiting for discovery to time out.
• Preserve dedicated timeout tests, silent-server fallback coverage, external conformance-client behavior, and production SDK defaults. Document the helper's scope in the contributor guide.

Why

A controlled server can take longer than the default five-second discovery probe budget to respond on a busy CI runner. The client then attempts legacy fallback even though the server supports the modern protocol. For tests pinned to the modern version, this surfaces as a misleading unsupported-protocol error before the behavior under test executes.

This is the connection failure observed in MapMcpStatelessTests.Mrtr_ConcurrentThreeInputs_ResolvedSimultaneously on #1934. The existing fix covered only selected helpers; other direct and shared connection paths retained the short probe budget.

Validation

• Release solution build: zero warnings and zero errors.
• Full non-manual Release test suite: 11,416 passed, 397 skipped, zero failures across all configured target frameworks.
• Added fake-clock coverage for delayed discovery, the overall initialization deadline, and explicit probe-timeout overrides. Confirmed the delayed-discovery regression fails with the original defaults and passes with the fix.
• Managed AOT compatibility app: passed. Native AOT publishing could not be verified locally because NuGet TLS failures prevented downloading the required runtime packs; CI must cover native publishing.

[jeffhandley]

Nice! Looks good to me. The only other thing I would suggest is setting TestConstants.DefaultTimeout to be a property getter that would respect an environment variable to override the 60s hard-coded value, but I don't think we need to introduce that unless we see any further flakiness where 60s isn't always enough in CI, but gosh I hope it is.

[tarekgh]

Nice! Looks good to me. The only other thing I would suggest is setting TestConstants.DefaultTimeout to be a property getter that would respect an environment variable to override the 60s hard-coded value, but I don't think we need to introduce that unless we see any further flakiness where 60s isn't always enough in CI, but gosh I hope it is.

Thanks, this is a good idea, let's wait to see any scenario need it and I'll be happy to add it.