fix(testinghost): dispose stopped in-process silo hosts without leaking or hanging

Merged
RE
3mo ago
+408-6

Problem

InProcessTestClusterBuilder/InProcessTestCluster leaks memory: stopped silos are never fully released (#10258). InProcessSiloHandle.Dispose/DisposeAsync began with if (!IsActive) return;, and stopping a silo sets IsActive to false, so a stopped silo's SiloHost was never disposed. That left the entire silo object graph (the IHost, its ServiceProvider, watchdog threads, and everything they root) alive for the lifetime of the test process. Building many in-process clusters in a single run accumulated stopped silo graphs and leaked memory.

Solution

Decouple disposal from IsActive using a one-shot interlocked guard, so the SiloHost is disposed exactly once regardless of the silo's active state, and await async disposal directly.

Adds regression tests that assert a stopped/killed silo's host is disposed and its ServiceProvider is collectible by the GC, plus a repro that kills a silo with an in-flight outbound call and asserts host disposal completes promptly instead of hanging.

Stacked on #10275

This PR is stacked on top of #10275 (fault outstanding callbacks when a host shuts down). That base change is what allows an ungracefully-killed silo with an in-flight outbound call to dispose without hanging. Until #10275 merges, this PR's diff will also show those runtime changes; once it merges, this branch will show only the in-process test cluster disposal/leak fix.