test(gateway): compose mid-loop compaction queue regression

[agent-farnsworth]

Summary

Add the remaining composed regression for #4121. The test drives the real conversation-scoped inbound queue, real DefaultAgentSupervisor, real in-process agent handle, real mid-loop hook, and real SessionCompactionCoordinator together.

Closes #4121

Coverage

• pauses an applied compaction inside the active run;
• enqueues a second message on the same conversation while compaction is in flight;
• proves both queue items reach Accepted within a signal-bounded harness;
• proves both runs emit RunEnded;
• proves the compacted summary and queued message persist exactly once;
• proves the next provider context contains the compacted summary;
• proves the active handle remains cached rather than synchronously disposing itself.

Restoring unconditional post-compaction eviction makes the bounded composed test fail instead of hanging the test project indefinitely.

Validation

• dotnet build tests/gateway/BotNexus.Gateway.Tests/BotNexus.Gateway.Tests.csproj --nologo --no-restore: 0 warnings, 0 errors.
• git diff --check origin/main...HEAD: clean.
• Authoritative remote gate: blocked before Azure execution by the known [CI] remote snapshots resurrect staged deletions from bundled HEAD #3918 sender-side exact-source packaging stall. Repeated bounded attempts produced no output directory, payload, run ID, execution name, or numeric result contract. No remote pass is claimed.
• Local test execution was not used because this workstation hosts the live gateway.

Risk

Test-only. No production, schema, configuration, or UI change. GitHub CI is expected to execute the composed regression while #3918 remains the validation-plane blocker.

[agent-farnsworth]

Validation update: the first GitHub core run had one unrelated BundledTrailguideBringUpJourney registration timeout. A bounded unchanged job retry passed at the same head e13dd9e74c0808c7dbf8ce7321d7b7e7f80267f4: 18,918 total, 18,880 executed/passed, 0 failed, 38 not executed. All 17 PR checks are now terminal with 14 successes and 3 expected skips.