fix: avoid replaying ambiguous stale stdio MCP calls

[9s5bz2jvd2-lang]

Summary

• Stop automatically replaying an MCP tool call after a stale stdio transport failure, because the first remote outcome may already have committed.
• Restart the transport for a future independent call, but return status="error", outcome="ambiguous", and retryable=false for the in-flight call.
• Preserve one explicit retry_policy="safe" opt-in for callers that have independently verified replay safety. No production caller opts in in this patch.

Validation

• git diff --check
• Focused stale-resource tests on current main: 21 passed
• MCP-focused tests on current main: 177 passed
• Direct-caller and structured-result consumer tests on current main: 659 passed
• Fake-session harness: one caller invocation, one side effect, zero replacement-session calls, one transport restart, FIX_CONFIRMED

Post-review timeout completion

• Added commit 3b51e21386f3c003a1405f4c0616e1600456fb28 (fix(mcp): mark post-submit timeouts ambiguous).
• A TimeoutError raised after asyncio.run_coroutine_threadsafe() has returned a Future now returns outcome="ambiguous" and retryable=false; the local Future is cancelled for cleanup only, with no claim that the remote call was cancelled and no blind replay.
• A same-class error raised before a Future exists remains an ordinary non-ambiguous error.
• Added regression coverage for both sides of that submission boundary.
• Additional validation: git diff --check, python -m compileall -q src/lingtai/services/mcp.py, focused test module 23 passed, and relevant MCP suite 156 passed.
• A second independent, read-only review found APPROVE / no material issue and confirmed the submission boundary plus unchanged stale-resource/safe-retry behavior.

Notes

• Follow-up to fix: restart stale MCP stdio sessions #222, which introduced stale stdio recovery by restarting the session and retrying once; this change narrows replay behavior when the remote outcome is ambiguous.
• Scope is limited to the stdio MCPClient and its regression tests; HTTPMCPClient is unchanged.
• This does not claim exactly-once execution. Shared-client restart synchronization and the existing broad stale-error classifier remain separate follow-ups.
• Validation used fake sessions only; it did not start a real MCP process, access credentials, contact the network, or perform an external delivery.
• One post-open broad-suite attempt failed during fixture setup because the test environment and pytest basetemp shared a parent that disappeared; there were no assertion failures. The rerun used independent parents and passed 177 + 659 tests.