[JacobStephens2]

Summary

Closes #1895.

 scope command ends
   sweep surviving processes
   inspect OOM evidence
+  release owner → stop its empty systemd unit

-thirdshift-<kind>-<pid>-<counter>-<detail>.scope
+thirdshift-<kind>-<pid>-<counter>-<random128>-<detail>.scope

Session, Trial and Grader owners unload retained empty units without relying on manager notifications. Already-collected units are silent, including collection racing with stop. Borrowed Trial membership retains its owner; survivor handling, memory caps and --collect are preserved.

Evidence

Execution evidence

Reproduced at the supplied merge base, 02c96545737cbd35efb69dd1ef93594f218ba9f2, in a detached worktree. Only the branch's unchanged regression file and its test-module declaration were added there; production code stayed at the merge base. The fake manager successfully runs the command, requires --collect, reports no processes, and retains the empty unit until explicitly stopped. No host systemd state was changed.

Commands used CARGO_BUILD_JOBS=2 and CARGO_TARGET_DIR=/tmp/thirdshift-1895-repro.ZTQmtu/target. The baseline package was cleaned before rebuilding the Issue branch to prevent reuse of baseline build artifacts. After results are from branch head f12ba6fba17bbdc11fe8bb0efbd364455b7a3e45.

• Before: cargo test --manifest-path /tmp/thirdshift-1895-repro.ZTQmtu/baseline/Cargo.toml --bin thirdshift scope::lifecycle_tests::finished_scopes_are_unloaded -- --exact --nocapture
  failed with finished Session scope remains loaded: thirdshift-session-1078883-0-harness.scope and 0 passed; 1 failed.
  After: the same test, cargo test --bin thirdshift scope::lifecycle_tests::finished_scopes_are_unloaded -- --exact --nocapture, passed: 1 passed; 0 failed, checking Session, Trial and Grader disposal.
• Before: cargo test --bin thirdshift scope::lifecycle_tests::reused_pid_and_counter_do_not_reuse_unit_names -- --exact --nocapture in the baseline worktree failed with PID reuse must not collide with a retained unit; both names were thirdshift-session-1079746-0-harness.scope (0 passed; 1 failed).
  After: the identical command on the Issue branch passed: 1 passed; 0 failed.
• cargo test --bin thirdshift scope:: -- --nocapture: 9 passed; 0 failed, including already-gone units, collection racing stop, cap arguments and OOM evidence retained until owner disposal.
• cargo test --bin thirdshift evals::experiment::execution::tests::scope_cap_evidence_preserves_explicit_voids_and_unexplained_kills_fail -- --exact --nocapture: 1 passed; 0 failed. Scope cap-kill evidence still produces a Void Trial with memory_cap; explicit Void causes and unexplained-kill failures remain intact.

Confirmed cause: scope owners left disposal entirely to manager collection; PID plus a restarting counter could then collide with a retained unit. The regressions exercise session_command, trial_command and grader_command through shared process supervision against isolated fake systemd commands.

The exact head's Rust, Console and site CI passed. This Continuation needed no source changes; successful exact-head full-suite CI was reused per docs/testing.md rather than repeated locally.

Review coverage

Independent thirdshift-code-review against refs/remotes/origin/main, pinned at 84c5856e6846f5c1f6cd8c1c80722341e2e267bf: Standards 0 findings, Spec 0 findings. Both reviewers fully read both changed files, src/scope.rs and src/scope/lifecycle_tests.rs. Unread changed files: none. Generated-output exclusions: none. Reports, each axis's final files-read list, finding counts and reproduction notes are retained in this Session's .thirdshift-review-vXiIWx/ directory.

Unaddressed findings

Standards

None.

Spec

None.

Security review outcome

No unaddressed introduced findings.

Merge Danger

Door: two-way

Reverting restores the earlier behavior. Existing leftover host units require separate operator cleanup.

Blast Radius: Execution

Scope disposal affects Session, Trial and Grader completion. It stops empty owned scopes after existing survivor handling and OOM inspection. Disposal errors produce warnings without changing command outcomes.

Built with codex codex-cli 0.164.0 · gpt-6.1-sol · xhigh