Fix .NET in-process callback lifetime

[minyitang]

Summary

• open tracked runtime connections and drain them before releasing managed callback state
• retain callback state and surface an error if graceful draining fails
• preserve ForceStopAsync behavior with an immediate close that safely retains callback state when necessary
• root the netstandard callback host with a GCHandle, matching the modern .NET lifetime guarantee

Compatibility

This is stacked on #2390 because copilot_runtime_connection_open_tracked and copilot_runtime_connection_close_and_wait are available in the bundled runtime starting with CLI 1.0.81-7; #2390 updates the SDK to 1.0.81-9.

Validation

• built net8.0, net10.0, and netstandard2.0 with zero warnings
• 217 .NET unit tests passed
• graceful and force-stop in-process FFI E2Es passed
• dotnet format --verify-no-changes passed

[github-actions]

Cross-SDK Consistency Review ✅

This PR makes internal implementation changes to the .NET FFI layer (FfiRuntimeHost, which is internal sealed). The changes address:

• Switching to copilot_runtime_connection_open_tracked / copilot_runtime_connection_close_and_wait native entry points
• Proper GCHandle lifetime management for the outbound callback
• Graceful vs. force-close code paths (Dispose vs ForceDispose)

No cross-SDK consistency issues found. Specifically:

• The public API (ForceStopAsync) already exists and is unchanged — no new surface area is being added
• The FFI/GCHandle mechanics are .NET-specific (P/Invoke, GCHandle, LibraryImport) and have no equivalent in other SDKs' runtimes
• Other SDKs (Node.js, Python, Go, Java, Rust) use different runtime bridging mechanisms that are unaffected by these native ABI changes

This is a valid language-specific internal fix.

Generated by SDK Consistency Review Agent for #2397 · sonnet46 11 AIC · ⌖ 4.06 AIC · ⊞ 6.6K · ◷