test(sessions): drive run_with_semaphore in the detach test

[ScriptedAlchemy]

This PR fixes one test that never exercised its subject. It also closes three load-flake issues whose listed tests are already fixed on master, with evidence for each.

#2809 deadline_detaches_without_leaking_the_permit

Root cause. BoundedGitControl uses a wall-clock Instant deadline (25 ms in the test). On a loaded host, run_with_semaphore can hit that deadline at its pre-spawn control.check() before it spawns the closure. The closure's started_tx was then dropped, and started_rx.await.unwrap() panicked with RecvError. #2852 removed that panic by reimplementing the semaphore and poll loop inline in the test. After that change the test never called run_with_semaphore, so it passed even when the product released the permit early.

Fix. The test calls run_with_semaphore again. If the deadline expires before the closure starts, the test asserts that the attempt returned CommandTimedOut and leaked no permit, then retries. Once the closure is running, the test asserts that the detached closure holds the permit until it is released.

Regression test. deadline_detaches_without_leaking_the_permit.

• I made the product release the permit early by changing let _permit = permit; to drop(permit); in blocking.rs. Master's test still passed. This test fails with left: 1 right: 0 at tests.rs:168.
• With a 50 ms delay injected before started_rx.await, the pre-fix(hosts,sessions): Hermes pycache cleanup and session test flakes #2852 test fails with RecvError(()). This test passes, taking the retry path once.
• It passed in 48 of 48 runs of the tracedecay-sessions lib suite at load 36 to 678.

Fixes #2809

#2811 prepared_generation_uses_bounded_parallelism_and_retained_bytes

Already fixed on master. 0f4bd88 (#2852) replaced the peak() > 8 assertion, which depended on scheduling, with a path_count-wide barrier that every page build must pass before it builds. fa6678c (#2929) made those gates one-shot so a rebuild cannot park in a barrier with no partners. The build guard is entered before the barrier, so a passing barrier already proves that all path_count builds overlapped.

Evidence. The test passed in 48 of 48 runs of the sessions lib suite at load 36 to 678, and passes again on this branch.

Fixes #2811

#2667 six tests that fail under host load

All six are fixed on master. The evidence comes from debug builds of master at 08f8fac.

• production_hook_ingest_reads_only_the_pinned_transcript_home was fixed by fix: unwedge refresh joins and scope rollback under load #2677. It passed 64 of 64 runs.
• maintenance_reclaims_a_removed_linked_worktree_and_the_text_artifact_only_it_named was fixed by fix: unwedge refresh joins and scope rollback under load #2677. It passed in 11 of 11 full lib-suite runs at load 50 to 355.
• an_identifier_named_in_the_task_leads_plan_and_explore_context was fixed by test: wait on readiness instead of racing host load #2678. It passed 64 of 64 runs.
• unenrolled_directory_session_serves_user_settings was fixed by test: wait on readiness instead of racing host load #2678. It passed 16 of 16 runs with 8 copies in parallel.
• failed_cold_mount_graph_replay_preserves_retained_text_generation was fixed by test: wait on readiness instead of racing host load #2678. It passed 11 of 11 full-suite runs.
• resume_attempts_refuses_a_live_holder_and_reports_the_lost_attempt was fixed by test: wait on readiness instead of racing host load #2678 and test(work): resume only after the restart's startup fence lands #2688. It passed 64 of 64 runs.

The 64 runs of each mcp_suite test ran with 8 to 24 copies in parallel, some of them alongside full lib-suite copies.

Fixes #2667

#2650 lib tests that fail under full-suite load

All ten are fixed on master or no longer reproduce. Each passed in all 11 full lib-suite runs (2 or 3 concurrent copies, load 50 to 355).

• many_slow_initialize_roots_share_one_discovery_budget, client_identity_startup_replays_retained_profile_receipts, owned_project_replay_worker_continues_past_one_bounded_batch, one_shot_tool_call_allows_long_response_while_daemon_stays_live and mounted_code_generation_retention_continues_capped_segment_reclamation were fixed by test(daemon): wait on readiness in load-sensitive lib tests #2674.
• first_advisory_cycle_after_a_reopen_answers_without_a_retry was fixed by fix(feedback): wait for a reopened generation's source proof #2668.
• The two session_temporal_benchmark tests were fixed by test(sessions): drop invented wall clocks from session tests #2676.
• preserved_profile_lcm_discovery_converges_without_blocking_retrieval was fixed by c0fe494 (feat(sessions): publish typed LCM convergence in status #2717). The fleet analysis on this issue left it open because it needed a truthful quiescence signal in place of the test's 2-minute CONVERGENCE_WAIT. feat(sessions): publish typed LCM convergence in status #2717 published a typed converged state, and the test now waits on it with no deadline. It passed 12 of 12 targeted runs alongside two full-suite copies at load 50 to 88.
• a_deferred_discovery_converges_on_the_next_request no longer reproduces. It failed once in the original 50 runs. It passed 12 of 12 targeted runs under full-suite load, plus 6 runs pinned to one or two cores shared with 6 to 12 busy-loop processes.

Fixes #2650

Related

Looping the lib suite surfaced four other load flakes that neither umbrella lists. They are tracked in #2969. The most frequent is cancelled_large_candidate_search_stops_before_the_next_batch, which failed in 5 of 11 runs.

Landing verification

Rebased on 05fe6b9.

• Mutation: let _permit = permit; changed to drop(permit); in blocking.rs. deadline_detaches_without_leaking_the_permit fails with left: 1 right: 0 at tests.rs:168. So does detached_task_holds_capacity_until_its_closure_finishes.
• cargo test -p tracedecay-sessions --lib: 542 passed.
• clippy -D warnings --all-targets, Windows gnu cargo check --all-targets on tracedecay-sessions, and cargo fmt --check are clean.

The test still loops when the 25 ms wall-clock deadline expires before the closure spawns. Each pass of the loop asserts the outcome of the branch it took: CommandTimedOut with the permit returned, or a detached closure that holds the permit until it is released. So the loop cannot hide a leak. BoundedGitControl reads std::time::Instant, and no paused clock can pin that branch.

[changeset-bot]

⚠️ No Changeset found

Latest commit: e4dbb0c

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[devin-ai-integration]

🔍 Devin Review: 1 flag

Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)