[elezar]

Stacked on #5455, which is stacked on #5450. This diff is against #5455's branch.

Summary

The testkit program wrote junit.xml only after its last test, so a run-level timeout (45 minutes) lost the report and only the log survived.

• Rewrite the report after each test, so a run that is stopped from outside still leaves the results so far. A write error during the run is logged; the final write still decides the exit status.
• Write to <path>.partial and rename, so a kill during a write leaves the previous report, not a truncated one.

The test that is running when the run is stopped is not in the report, since it has no result yet.

Related Issue

Part of #5416 (this item came from the review of #5332; the issue stays open). Also resolves the junit.xml note on #5333.

Testing

• cargo fmt --check, cargo clippy --all-targets -- -D warnings and cargo test pass for tests/testkit.
• A new unit test checks that writing replaces the report and leaves no partial file.
• No e2e lane was run.

[Karthik-Chowdary]

The Disposition::Skip branch pushes an Outcome::Skipped and then continues before the new incremental JUnit checkpoint. That leaves a completed skip out of the surviving report if the supervisor is killed during the next executed test.

On head 31cd8bd488b10997f4dd3a2c52b46a233ba64dee, I exercised the unchanged program/run and registry modules with a pass → deliberate failure → ignored test → later blocking test. SIGKILL after the later test started left parseable XML with tests="2", failures="1", skipped="0": the pass and failure survived, including failure output, but the already-processed ignored test did not. This repeated three times. A skip before a completed test was retained; --include-ignored, normal completion, and graceful SIGTERM behaved as controls.

Could skipped and executed results both reach the incremental checkpoint, with a process-level regression for a skipped result immediately before the in-flight test? The in-flight test itself can remain absent as documented. Checkpointing the skip branch in a scratch-only correction retained tests="3", failures="1", skipped="1" in the same cases.

The reproduction used an isolated fixture over the upstream runner modules; it does not qualify gateway readiness or an e2e lane.