[P0] Expose and fix trusted-binding replay CI race #181

New issue

Closed

opened 2026-08-24 16:46:55 +00:00 by lost-rob0t · 2 comments

lost-rob0t commented

2026-08-24 16:46:55 +00:00

(Migrated from github.com)

Parent/runtime area: #79

Observed failure

Post-merge main CI run 32750635555 at 0d8e1008ba525f0ec7e51cba2de34ede8a63cf6f completed the deterministic aggregate with:

aggregate_plunit_summary suites=75 discovered=783 planned=783 completed=783 passed=782 failed=1 timeout=0 blocked=0 fixme=0

The failed case was:

rlm_tool_effect:trusted_binding_identity_prevents_cross_binding_replay

The PR head and merge commit have the identical Git tree 3c7b41f8b4bc02bedac94fb5d4ad01ef1572cada. The PR aggregate passed; the post-merge aggregate failed. Fresh-process reproduction attempts passed 50/50 for the focused case, 50/50 for the full effect suite, and 12/12 for the complete aggregate.

Problem

The test contains unasserted asynchronous transition and outcome unifications. On failure, PL-Unit reports only test ...: failed, discarding the actual approval, Future, effect, and mutation state needed to identify the violated boundary. Rerunning CI cannot substitute for exposing and fixing the cause.

Required fix

deterministically reproduce or force the failing interleaving using barriers/queues/mutexes, never sleeps as correctness synchronization;
expose the exact failed state/transition in test evidence;
fix the underlying authority/effect/async lifecycle defect, or if executable evidence proves only the oracle is wrong, repair the oracle without weakening the invariant;
preserve stable trusted binding identity across registry allocation and prevent cross-binding replay;
keep the existing #53 authority, #54 async, and #57 durable-effect boundaries;
do not skip, retry-away, or mark the deterministic check informational.

Acceptance

a focused red regression identifies the violated invariant and fails for the intended reason;
the focused effect suite is green under deterministic adversarial scheduling;
the complete aggregate reports all 783+ discovered/planned/completed/passed with zero failure/timeout/blocked/fixme;
runner integrity, benchmarks, CLI demo, Nix/package checks, and exact-head GitHub Actions are green;
the post-merge main workflow is green after an explicitly authorized merge.

Parent/runtime area: #79 ## Observed failure Post-merge `main` CI run 32750635555 at `0d8e1008ba525f0ec7e51cba2de34ede8a63cf6f` completed the deterministic aggregate with: ```text aggregate_plunit_summary suites=75 discovered=783 planned=783 completed=783 passed=782 failed=1 timeout=0 blocked=0 fixme=0 ``` The failed case was: ```text rlm_tool_effect:trusted_binding_identity_prevents_cross_binding_replay ``` The PR head and merge commit have the identical Git tree `3c7b41f8b4bc02bedac94fb5d4ad01ef1572cada`. The PR aggregate passed; the post-merge aggregate failed. Fresh-process reproduction attempts passed 50/50 for the focused case, 50/50 for the full effect suite, and 12/12 for the complete aggregate. ## Problem The test contains unasserted asynchronous transition and outcome unifications. On failure, PL-Unit reports only `test ...: failed`, discarding the actual approval, Future, effect, and mutation state needed to identify the violated boundary. Rerunning CI cannot substitute for exposing and fixing the cause. ## Required fix - deterministically reproduce or force the failing interleaving using barriers/queues/mutexes, never sleeps as correctness synchronization; - expose the exact failed state/transition in test evidence; - fix the underlying authority/effect/async lifecycle defect, or if executable evidence proves only the oracle is wrong, repair the oracle without weakening the invariant; - preserve stable trusted binding identity across registry allocation and prevent cross-binding replay; - keep the existing #53 authority, #54 async, and #57 durable-effect boundaries; - do not skip, retry-away, or mark the deterministic check informational. ## Acceptance - [ ] a focused red regression identifies the violated invariant and fails for the intended reason; - [ ] the focused effect suite is green under deterministic adversarial scheduling; - [ ] the complete aggregate reports all 783+ discovered/planned/completed/passed with zero failure/timeout/blocked/fixme; - [ ] runner integrity, benchmarks, CLI demo, Nix/package checks, and exact-head GitHub Actions are green; - [ ] the post-merge `main` workflow is green after an explicitly authorized merge.

lost-rob0t commented

2026-08-24 17:02:27 +00:00

(Migrated from github.com)

Fix is ready in PR #182 at exact head 94737163ed7980b60a4d559920cbc04d8fc287d0. The previously failing deterministic job is green; clean-pack, Nix, Tree-sitter FFI, and REAL OpenRouter are also green. Local evidence: forced cross-binding interleaving 800/800, effect suite 14/14, runner integrity 19/19, aggregate 783/783, deterministic benchmark 16/16, deep experiment 15/15, JSON demo and three Nix checks passed. The PR remains unmerged pending explicit merge authorization.

lost-rob0t commented

2026-08-24 23:56:26 +00:00

(Migrated from github.com)

Merged by rebase via PR #182. Reviewed head d672a0fe7d became main commit e6f04bb641. The post-merge main CI, clean-pack, Nix, Tree-sitter FFI, and pinned paid OpenRouter workflows all completed successfully.

Merged by rebase via PR #182. Reviewed head d672a0fe7d53fc38d2728dda0a711834ded3de97 became main commit e6f04bb641d5f15d2b806eca4bd00d2718615034. The post-merge main CI, clean-pack, Nix, Tree-sitter FFI, and pinned paid OpenRouter workflows all completed successfully.