harness-everything 0.30.9 → 0.31.0

package/benchmarks/results/pr-303-review-2026-10-13/gates/heavy/test-mechanism.log

@@ -0,0 +1,1737 @@
+
+> harness-everything@0.30.10 test:mechanism
+> node ci/mechanism-test.js
+
+=================================================
+=== 🔧 Harness Mechanism Test Suite (VERIFICATION.md §2) ===
+=================================================
+
+Discovered 76 mechanism test suite(s):
+  - mechanism-2a-rule-of-3.test.js
+  - mechanism-2aa-codex-action-gate-policy.test.js
+  - mechanism-2ab-openai-plugin-sync-integrity.test.js
+  - mechanism-2ac-atomic-commit-check.test.js
+  - mechanism-2ad-action-gate-classifier-hardening.test.js
+  - mechanism-2af-worktree-isolation.test.js
+  - mechanism-2ag-router-worktree-contract.test.js
+  - mechanism-2ah-workflow-lifecycle.test.js
+  - mechanism-2ai-workspace-fingerprint-batch.test.js
+  - mechanism-2aj-waza-reference-adapter.test.js
+  - mechanism-2ak-contract-probe-adapter.test.js
+  - mechanism-2al-codex-hooks-compat.test.js
+  - mechanism-2am-observations-session-import.test.js
+  - mechanism-2am-openai-packaged-helper-runtime.test.js
+  - mechanism-2an-environment-detection-readonly.test.js
+  - mechanism-2an-workflow-obligations.test.js
+  - mechanism-2ao-plugin-sync.test.js
+  - mechanism-2ap-sbom.test.js
+  - mechanism-2aq-fable-stage-bindings.test.js
+  - mechanism-2ar-binding-resolution.test.js
+  - mechanism-2b-boundary-guard.test.js
+  - mechanism-2c-state-persist.test.js
+  - mechanism-2d-fact-audit.test.js
+  - mechanism-2e-scope-guard.test.js
+  - mechanism-2f-stop-gate.test.js
+  - mechanism-2g-platform-ignore.test.js
+  - mechanism-2h-installer-manifest.test.js
+  - mechanism-2i-crlf-frontmatter.test.js
+  - mechanism-2i-install-verification.test.js
+  - mechanism-2j-behavioral-runner.test.js
+  - mechanism-2j-fixture-entrypoints.test.js
+  - mechanism-2k-reference-check.test.js
+  - mechanism-2l-multi-agent-workspace.test.js
+  - mechanism-2m-self-repo-guard.test.js
+  - mechanism-2n-opencode-plugin.test.js
+  - mechanism-2o-state-home-isolation.test.js
+  - mechanism-2p-installer-links.test.js
+  - mechanism-2q-installer-global-shared-store.test.js
+  - mechanism-2r-uninstall-flags.test.js
+  - mechanism-2s-tdd-quality.test.js
+  - mechanism-2t-disclosure.test.js
+  - mechanism-2t-evidence-triage.test.js
+  - mechanism-2t-issue-20.test.js
+  - mechanism-2t-platform-skill-targets.test.js
+  - mechanism-2t-transcript-parser.test.js
+  - mechanism-2t-workspace-identity.test.js
+  - mechanism-2u-issue-42-regressions.test.js
+  - mechanism-2u-issue-44.test.js
+  - mechanism-2v-verifier-boundaries.test.js
+  - mechanism-2w-platform-compatibility.test.js
+  - mechanism-2x-release-automation.test.js
+  - mechanism-2x-release-commit-message.test.js
+  - mechanism-2y-release-tooling-compat.test.js
+  - mechanism-2z-claude-plugin-manifest.test.js
+  - mechanism-30-opencode-plugin-loadability.test.js
+  - mechanism-31-opencode-hardlock-evidence.test.js
+  - mechanism-32-paired-benchmark.test.js
+  - mechanism-33-issues-205-207.test.js
+  - mechanism-34-observations.test.js
+  - mechanism-34-worktree-submodule-reachability.test.js
+  - mechanism-35-verification-contract.test.js
+  - mechanism-packaging-bytecode.test.js
+  - mechanism-system-one-collect.test.js
+  - mechanism-system-one-contract.test.js
+  - mechanism-system-one-corpus.test.js
+  - mechanism-system-one-evaluation.test.js
+  - mechanism-system-one-install.test.js
+  - mechanism-system-one-intent.test.js
+  - mechanism-system-one-label.test.js
+  - mechanism-system-one-laya.test.js
+  - mechanism-system-one-ngram.test.js
+  - mechanism-system-one-provider.test.js
+  - mechanism-system-one-relevance.test.js
+  - mechanism-system-one-resident.test.js
+  - mechanism-system-one-router.test.js
+  - mechanism-system-one-train.test.js
+
+-------------------------------------------------
+
+▶️  Running test suite: mechanism-2a-rule-of-3.test.js
+
+[2a] Rule of 3 circuit breaker...
+✅ 2a-wiring. PostToolUseFailure invokes the Rule-of-3 tracker
+✅ 2a-wiring. apply_patch is covered before and after execution
+✅ 2a-three. exactly three matching failures arm zoom-out
+npm.cmd : [CRITICAL] RULE OF 3 CIRCUIT BREAKER TRIGGERED!
+位於 線路:4 字元:1
++ & npm.cmd run test:mechanism *> $log
++ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    + CategoryInfo          : NotSpecified: ([CRITICAL] RULE...AKER TRIGGERED!:String) [], RemoteException
+    + FullyQualifiedErrorId : NativeCommandError
+
+The same failure signature has now occurred 3 times.
+Error Signature: 84edbe719770818783c799972ef28e77
+
+Mutating tools (Bash/PowerShell/Edit/Write/apply_patch) are locked - not because you
+must wait for a human, but because you must stop drilling and rebuild the
+full picture first. Do NOT ask the Human Partner yet.
+
+ACTION REQUIRED (reflect first):
+1. CEASE FIRE. No new fix attempts, no variations of the last one.
+2. Load the 'zoom-out' skill and follow its Reflect & Fact-Check protocol.
+3. Using READ-ONLY tools (Read/Grep/Glob - still available), re-verify the
+   assumption behind each failed attempt against the actual code/logs/docs.
+4. Write your findings to '<temp>\harness-mechanism-state-V9H7Y9\workspaces\harness-e
+verything-issue-297-6eb1109f6a2a\state\sessions\__mechanism_test__\zoom-out-report.md' (the ONLY write
+   allowed right now) with sections: ## Goal, ## Failed Attempts,
+   ## Verified Facts, ## Diagnosis, ## Decision (ending RESUME: or ESCALATE:).
+
+A valid report releases this breaker automatically. Escalate to the Human
+Partner ONLY if your Decision concludes a genuinely human call is needed
+(requirement conflict, architecture trade-off, destructive action, missing
+access) - and then present options with a recommendation, not apologies.
+✅ 2a-trip. third matching failure blocks mutation for reflection
+✅ 2a-release. valid reflection releases breaker and resets count
+[CRITICAL] RULE OF 3 CIRCUIT BREAKER TRIGGERED!
+The same failure signature has now occurred 3 times.
+Error Signature: mech-test
+
+Mutating tools (Bash/PowerShell/Edit/Write/apply_patch) are locked - not because you
+must wait for a human, but because you must stop drilling and rebuild the
+full picture first. Do NOT ask the Human Partner yet.
+
+ACTION REQUIRED (reflect first):
+1. CEASE FIRE. No new fix attempts, no variations of the last one.
+2. Load the 'zoom-out' skill and follow its Reflect & Fact-Check protocol.
+3. Using READ-ONLY tools (Read/Grep/Glob - still available), re-verify the
+   assumption behind each failed attempt against the actual code/logs/docs.
+4. Write your findings to '<temp>\harness-mechanism-state-V9H7Y9\workspaces\harness-e
+verything-issue-297-6eb1109f6a2a\state\sessions\__mechanism_test__\zoom-out-report.md' (the ONLY write
+   allowed right now) with sections: ## Goal, ## Failed Attempts,
+   ## Verified Facts, ## Diagnosis, ## Decision (ending RESUME: or ESCALATE:).
+
+A valid report releases this breaker automatically. Escalate to the Human
+Partner ONLY if your Decision concludes a genuinely human call is needed
+(requirement conflict, architecture trade-off, destructive action, missing
+access) - and then present options with a recommendation, not apologies.
+✅ 2a-repeat. later three-failure cycle requests another zoom-out, not permanent hard lock
+✅ 2a-category. permission failures still use the same threshold of three
+✅ 2a-category-two. two permission failures do not trip early
+[CRITICAL] RULE OF 3 CIRCUIT BREAKER TRIGGERED!
+The same failure signature has now occurred 3 times.
+Error Signature: ca808a9df9b7d8c531929ca6ea04a324
+
+Mutating tools (Bash/PowerShell/Edit/Write/apply_patch) are locked - not because you
+must wait for a human, but because you must stop drilling and rebuild the
+full picture first. Do NOT ask the Human Partner yet.
+
+ACTION REQUIRED (reflect first):
+1. CEASE FIRE. No new fix attempts, no variations of the last one.
+2. Load the 'zoom-out' skill and follow its Reflect & Fact-Check protocol.
+3. Using READ-ONLY tools (Read/Grep/Glob - still available), re-verify the
+   assumption behind each failed attempt against the actual code/logs/docs.
+4. Write your findings to '<temp>\harness-mechanism-state-V9H7Y9\workspaces\harness-e
+verything-issue-297-6eb1109f6a2a\state\sessions\__mechanism_test__\zoom-out-report.md' (the ONLY write
+   allowed right now) with sections: ## Goal, ## Failed Attempts,
+   ## Verified Facts, ## Diagnosis, ## Decision (ending RESUME: or ESCALATE:).
+
+A valid report releases this breaker automatically. Escalate to the Human
+Partner ONLY if your Decision concludes a genuinely human call is needed
+(requirement conflict, architecture trade-off, destructive action, missing
+access) - and then present options with a recommendation, not apologies.
+✅ 2a-category-three. third permission failure trips normally
+✅ 2a-interleaved. unrelated success preserves the matching failure count
+✅ 2a-interleaved-three. third matching failure still arms zoom-out after unrelated success
+✅ 2a-resolved. same-operation success clears the resolved signature
+
+▶️  Running test suite: mechanism-2aa-codex-action-gate-policy.test.js
+PASS: Codex action-gate uses supported hard blocking, attributable PreToolUse entry evidence, and non-authorizing PermissionRequest observation.
+
+▶️  Running test suite: mechanism-2ab-openai-plugin-sync-integrity.test.js
+PASS: OpenAI/Codex plugin sync is idempotent, uses portable plugin placeholders, and preserves every referenced hook runtime file.
+
+▶️  Running test suite: mechanism-2ac-atomic-commit-check.test.js
+PASS: atomic commit check scopes edits to the workspace and suppresses clean-tree false nudges
+
+▶️  Running test suite: mechanism-2ad-action-gate-classifier-hardening.test.js
+PASS: destructive-command classifier covers confirmed bypasses while suppressing quoted-text false positives
+
+▶️  Running test suite: mechanism-2af-worktree-isolation.test.js
+PASS fixture repository initializes
+PASS Tier-3 primary-tree mutation is not hard-blocked
+PASS Tier-3 primary-tree mutation emits mandatory isolation disposition without blocking
+PASS mutation-shaped shell command remains available
+PASS mutation-shaped shell command receives the same isolation contract
+PASS read-only discovery remains available
+PASS linked worktree fixture creates
+PASS linked-worktree mutation remains available
+PASS cross-worktree target is reminder-only
+PASS cross-worktree target emits scope/isolation guidance
+PASS legacy blocked workflow state does not trap mutation
+PASS legacy blocked state is surfaced as reminder
+PASS corrupt workflow state fails open
+PASS: worktree isolation is a semantic MUST disposition, not a cognitive lock
+
+▶️  Running test suite: mechanism-2ag-router-worktree-contract.test.js
+PASS: router major-workflow plan carries worktree isolation independently of multi-agent workspace state.
+
+▶️  Running test suite: mechanism-2ah-workflow-lifecycle.test.js
+PASS workflow guidance fixture is created
+PASS hooks/scripts/workflow-gate.js malformed payload fails open
+PASS hooks/scripts/workflow-stop-gate.js malformed payload fails open
+PASS legacy exhausted state no longer blocks mutation
+PASS legacy state is a reminder
+PASS unproven shell is not iteration-blocked
+PASS no mutation-probe state is created
+PASS failed shell mutation tracking is fail-open
+PASS failed-but-mutating shell advances edit freshness without fabricating verification
+PASS missing verification does not block Stop
+PASS missing verification emits reminder
+PASS exec_command verification is fail-open
+PASS #193 exec_command records lastVerifyAt
+PASS resolved evidence never rewrites an explicit blocked workflow to satisfied
+PASS: workflow guidance lifecycle (14 assertions)
+PASS: packaged OpenAI workflow guidance mirrors source
+
+▶️  Running test suite: mechanism-2ai-workspace-fingerprint-batch.test.js
+PASS #169 batched fingerprint matches the legacy clean-worktree fingerprint
+PASS #169 packaged plugin fingerprint matches canonical/legacy on a clean worktree
+PASS #169 batched fingerprint matches legacy for dirty tracked + untracked content
+PASS #169 packaged plugin batched fingerprint matches legacy dirty semantics
+PASS #169 staging visible content does not change the workspace fingerprint
+PASS #169 tracked deletion keeps legacy fingerprint semantics
+PASS #169 1000-untracked-file fingerprint returns a valid digest
+PASS #169 1000-untracked-file fingerprint stays below 5s regression ceiling (actual 501.9ms)
+PASS #169 packaged plugin fingerprint remains byte-identical for 1000 untracked files
+PASS #169 explicit fingerprint cap reports a structured conservative-fallback signal
+PASS #169 large-workspace Tier-2 fixture routes normally
+PASS #169 large-workspace Tier-2 fixture resolves requirements/workflow planning
+PASS #169 large-workspace Tier-2 fixture starts after planning
+PASS #169 >5000 observed paths degrade conservatively instead of blocking the shell call
+PASS #190 shell admission no longer creates conservative budget/probe state
+PASS: batched workspace fingerprint (15 assertions, 1000-file 501.9ms)
+
+▶️  Running test suite: mechanism-2aj-waza-reference-adapter.test.js
+PASS discovers the canonical Harness skill set
+PASS adapter invokes Waza exactly once per canonical top-level skill
+PASS Waza receives explicit canonical skill paths instead of recursive workspace discovery
+PASS environment-detection placeholder reference is reachable: references/shell-syntax-rules.md
+PASS environment-detection removes Waza orphan advisories for Harness-reachable references
+PASS environment-detection reports the suppressed advisories explicitly
+PASS environment-detection recomputes readiness without false local-orphan failures
+PASS eval-harness placeholder reference is reachable: references/scoring-rubric.md
+PASS eval-harness removes Waza orphan advisories for Harness-reachable references
+PASS eval-harness reports the suppressed advisories explicitly
+PASS eval-harness recomputes readiness without false local-orphan failures
+PASS fable-discipline placeholder reference is reachable: references/discipline-rules.md
+PASS fable-discipline removes Waza orphan advisories for Harness-reachable references
+PASS fable-discipline reports the suppressed advisories explicitly
+PASS fable-discipline recomputes readiness without false local-orphan failures
+PASS fable-mode placeholder reference is reachable: references/profile-lookup.md
+PASS fable-mode placeholder reference is reachable: references/model-matrix.md
+PASS fable-mode removes Waza orphan advisories for Harness-reachable references
+PASS fable-mode reports the suppressed advisories explicitly
+PASS fable-mode recomputes readiness without false local-orphan failures
+PASS find-skills placeholder reference is reachable: references/discovery-flow.md
+PASS find-skills removes Waza orphan advisories for Harness-reachable references
+PASS find-skills reports the suppressed advisories explicitly
+PASS find-skills recomputes readiness without false local-orphan failures
+PASS git-commit placeholder reference is reachable: references/commit-flow.md
+PASS git-commit removes Waza orphan advisories for Harness-reachable references
+PASS git-commit reports the suppressed advisories explicitly
+PASS git-commit recomputes readiness without false local-orphan failures
+PASS grill-me placeholder reference is reachable: references/grilling-playbook.md
+PASS grill-me removes Waza orphan advisories for Harness-reachable references
+PASS grill-me reports the suppressed advisories explicitly
+PASS grill-me recomputes readiness without false local-orphan failures
+PASS grill-with-docs placeholder reference is reachable: references/session-playbook.md
+PASS grill-with-docs removes Waza orphan advisories for Harness-reachable references
+PASS grill-with-docs reports the suppressed advisories explicitly
+PASS grill-with-docs recomputes readiness without false local-orphan failures
+PASS harness-everything placeholder reference is reachable: references/triage-and-tiers.md
+PASS harness-everything removes Waza orphan advisories for Harness-reachable references
+PASS harness-everything reports the suppressed advisories explicitly
+PASS harness-everything recomputes readiness without false local-orphan failures
+PASS improve-codebase-architecture placeholder reference is reachable: references/README.md
+PASS improve-codebase-architecture removes Waza orphan advisories for Harness-reachable references
+PASS improve-codebase-architecture reports the suppressed advisories explicitly
+PASS improve-codebase-architecture recomputes readiness without false local-orphan failures
+PASS install-cognitive-os placeholder reference is reachable: references/cognitive-loop.md
+PASS install-cognitive-os removes Waza orphan advisories for Harness-reachable references
+PASS install-cognitive-os reports the suppressed advisories explicitly
+PASS install-cognitive-os recomputes readiness without false local-orphan failures
+PASS multi-agent-workspace placeholder reference is reachable: references/orchestration.md
+PASS multi-agent-workspace placeholder reference is reachable: references/architecture-guide.md
+PASS multi-agent-workspace removes Waza orphan advisories for Harness-reachable references
+PASS multi-agent-workspace reports the suppressed advisories explicitly
+PASS multi-agent-workspace recomputes readiness without false local-orphan failures
+PASS repo-docs placeholder reference is reachable: references/process-guide.md
+PASS repo-docs removes Waza orphan advisories for Harness-reachable references
+PASS repo-docs reports the suppressed advisories explicitly
+PASS repo-docs recomputes readiness without false local-orphan failures
+PASS rewrite-commits placeholder reference is reachable: references/workflow.md
+PASS rewrite-commits removes Waza orphan advisories for Harness-reachable references
+PASS rewrite-commits reports the suppressed advisories explicitly
+PASS rewrite-commits recomputes readiness without false local-orphan failures
+PASS security-review placeholder reference is reachable: references/security-checklist.md
+PASS security-review removes Waza orphan advisories for Harness-reachable references
+PASS security-review reports the suppressed advisories explicitly
+PASS security-review recomputes readiness without false local-orphan failures
+PASS self-evolve placeholder reference is reachable: references/memory-resolution.md
+PASS self-evolve removes Waza orphan advisories for Harness-reachable references
+PASS self-evolve reports the suppressed advisories explicitly
+PASS self-evolve recomputes readiness without false local-orphan failures
+PASS skill-creator placeholder reference is reachable: references/quality-checklist.md
+PASS skill-creator removes Waza orphan advisories for Harness-reachable references
+PASS skill-creator reports the suppressed advisories explicitly
+PASS skill-creator recomputes readiness without false local-orphan failures
+PASS skill-style placeholder reference is reachable: references/style-guide.md
+PASS skill-style removes Waza orphan advisories for Harness-reachable references
+PASS skill-style reports the suppressed advisories explicitly
+PASS skill-style recomputes readiness without false local-orphan failures
+PASS tdd placeholder reference is reachable: references/core-discipline.md
+PASS tdd removes Waza orphan advisories for Harness-reachable references
+PASS tdd reports the suppressed advisories explicitly
+PASS tdd recomputes readiness without false local-orphan failures
+PASS to-spec placeholder reference is reachable: references/process.md
+PASS to-spec removes Waza orphan advisories for Harness-reachable references
+PASS to-spec reports the suppressed advisories explicitly
+PASS to-spec recomputes readiness without false local-orphan failures
+PASS to-tickets placeholder reference is reachable: references/publishing-and-templates.md
+PASS to-tickets removes Waza orphan advisories for Harness-reachable references
+PASS to-tickets reports the suppressed advisories explicitly
+PASS to-tickets recomputes readiness without false local-orphan failures
+PASS todo-driven-workflow placeholder reference is reachable: references/execution-guide.md
+PASS todo-driven-workflow removes Waza orphan advisories for Harness-reachable references
+PASS todo-driven-workflow reports the suppressed advisories explicitly
+PASS todo-driven-workflow recomputes readiness without false local-orphan failures
+PASS using-git-worktrees placeholder reference is reachable: references/submodules-in-worktrees.md
+PASS using-git-worktrees placeholder reference is reachable: references/workflow-details.md
+PASS using-git-worktrees removes Waza orphan advisories for Harness-reachable references
+PASS using-git-worktrees reports the suppressed advisories explicitly
+PASS using-git-worktrees recomputes readiness without false local-orphan failures
+PASS verification-loop placeholder reference is reachable: references/project-verification-contract.md
+PASS verification-loop placeholder reference is reachable: references/verification-phases.md
+PASS verification-loop removes Waza orphan advisories for Harness-reachable references
+PASS verification-loop reports the suppressed advisories explicitly
+PASS verification-loop recomputes readiness without false local-orphan failures
+PASS verify-before-claim placeholder reference is reachable: references/verification-guide.md
+PASS verify-before-claim removes Waza orphan advisories for Harness-reachable references
+PASS verify-before-claim reports the suppressed advisories explicitly
+PASS verify-before-claim recomputes readiness without false local-orphan failures
+PASS zoom-out placeholder reference is reachable: references/circuit-breaker.md
+PASS zoom-out removes Waza orphan advisories for Harness-reachable references
+PASS zoom-out reports the suppressed advisories explicitly
+PASS zoom-out recomputes readiness without false local-orphan failures
+PASS at least one canonical skill uses the Harness placeholder reference contract
+PASS adapter resolves canonical placeholder-orphan advisories
+PASS fable-mode nested reference remains reachable: references/execution-phases.md
+PASS fable-mode nested reference advisories are resolved through the reachable reference graph
+PASS harness-everything nested reference remains reachable: references/router-workflow-plan.md
+PASS harness-everything nested reference remains reachable: references/skill-registry.md
+PASS harness-everything nested reference advisories are resolved through the reachable reference graph
+PASS multi-agent-workspace nested reference remains reachable: references/agency-agents.md
+PASS multi-agent-workspace nested reference advisories are resolved through the reachable reference graph
+PASS skill-creator nested reference remains reachable: references/authoring-workflow.md
+PASS skill-creator nested reference remains reachable: references/testing-workflow.md
+PASS skill-creator nested reference remains reachable: references/quality-principles.md
+PASS skill-creator nested reference remains reachable: references/dynamic-generation-contract.md
+PASS skill-creator nested reference advisories are resolved through the reachable reference graph
+PASS tdd nested reference remains reachable: references/quality-model.md
+PASS tdd nested reference remains reachable: references/unit-testing.md
+PASS tdd nested reference remains reachable: references/integration-testing.md
+PASS tdd nested reference advisories are resolved through the reachable reference graph
+PASS <skills-repo-root> existing target is a resolved Harness contract link
+PASS generic runtime placeholder is classified as a template advisory
+PASS unknown placeholder remains a real local-link problem
+PASS existing repo-scope relative link is resolved by the Harness repository contract
+PASS missing repo-scope relative link remains a hard local-link problem
+PASS links inside generated-document templates remain visible template advisories
+PASS network and generic-template findings do not masquerade as broken local links
+PASS network and template findings are retained in separate advisory categories
+PASS normalized readiness remains green when only false-orphan/network/template advisories remain
+PASS non-link readiness failures remain explicit machine-readable blockers
+PASS: Waza/Harness reference adapter (139 assertions across 26 skills)
+
+▶️  Running test suite: mechanism-2ak-contract-probe-adapter.test.js
+=== Contract Integrity Phase 3 Node Probe Adapter (#84) ===
+  PASS declared node-npm-v1 probe plan validates
+  PASS green baseline plus both explicit probes produces PROTECTED evidence
+  PASS env and exact-replace probes are killed only by expected contract failures
+  PASS adapter never mutates the source workspace
+  PASS probe report retains output hashes/evidence refs instead of project test output
+  PASS derived adapter evidence feeds the existing #84/#58 strict integrity gate
+  PASS green test run under a required probe is explicitly SURVIVED, never PASS
+  PASS green but weak suite cannot compensate 100 completeness with a surviving contract probe
+  PASS unrelated failing test output is INVALID and cannot be credited as a kill
+  PASS ambiguous exact-replace fallback fails closed without mutating source
+  PASS replace probe rejects symlink/junction escape without mutating outside the isolated workspace
+  PASS workspace copy rejects source symlinks/junctions before project code can mutate an outside target
+  PASS v1 adapter refuses to infer/install external dependencies
+  PASS non-green baseline blocks all protection claims
+  PASS probe plan rejects replace targets that escape the isolated workspace
+  PASS CLI returns exit 0 only for fully killed declared probes
+  PASS adapter documentation states exact supported scope and explicit non-support
+
+PASS: #84 contract protection node adapter (0 failures)
+
+▶️  Running test suite: mechanism-2al-codex-hooks-compat.test.js
+
+[2al] Codex user-hook compatibility fallback (#122)...
+  PASS explicit install materializes compatibility entries
+  PASS install never changes Codex trust/config.toml state
+  PASS status verifies exact installed entry hashes and runtime tree
+  PASS status distinguishes installed hooks from Codex-trusted live capture
+  PASS existing unrelated user hooks and top-level metadata are preserved semantically
+  PASS every compatibility hook command resolves to the Harness-owned runtime copy
+  PASS compatibility entries use literal absolute runtime paths without shell-specific plugin expansion
+  PASS compatibility runtime is self-contained for hook and router dependencies
+  PASS reinstall/update is idempotent and does not duplicate hook entries
+  PASS update preserves user hooks added after the initial compatibility install
+  PASS modified Harness compatibility entry makes update fail closed without rewriting user config
+  PASS modified Harness compatibility entry makes uninstall fail closed rather than deleting user edits
+  PASS explicit uninstall removes exact Harness-owned entries without changing trust state
+  PASS uninstall preserves unrelated pre-existing and later-added user hooks
+  PASS uninstall leaves config.toml/trust state byte-identical
+  PASS uninstall removes Harness compatibility ownership manifest
+  PASS compatibility install can create a new hooks.json when none exists
+  PASS uninstall removes hooks.json only when Harness created it and no user entries remain
+  PASS compatibility runtime is copied from the verified installed marketplace source
+  PASS extra identical Harness hook is treated as user config drift, not silently deleted
+  PASS uninstall never recreates a user hooks.json that was removed after compatibility install
+  PASS byte-identical pre-existing user hooks are reused but never recorded as Harness-owned
+  PASS uninstall preserves byte-identical hooks that predated compatibility install
+  PASS forged manifest ownership fails closed without deleting a user hook
+  PASS failed manifest commit rolls hooks.json back byte-identically
+  PASS linked compatibility owner path cannot redirect writes outside Codex home
+  PASS malformed user hooks.json fails closed without replacement
+  PASS status clearly reports inactive observation capture when hooks are absent
+  PASS fallback derives from the current packaged hook manifest instead of maintaining a second manual hook inventory
+
+PASS: #122 Codex compatibility fallback (0 failures)
+
+▶️  Running test suite: mechanism-2am-observations-session-import.test.js
+PASS: post-revision Codex session labels import as redacted, prompt-paired, idempotent observations.
+
+▶️  Running test suite: mechanism-2am-openai-packaged-helper-runtime.test.js
+OpenAI packaged helper runtime smoke test passed
+
+▶️  Running test suite: mechanism-2an-environment-detection-readonly.test.js
+Issue #213 verified: detection is read-only across six adapters, repair is explicit/idempotent, and OpenCode stays separate.
+
+▶️  Running test suite: mechanism-2an-workflow-obligations.test.js
+=== Issue #85 requirements-first workflow contract ===
+  PASS Tier 2 route persists successfully
+  PASS Tier 2 receives a preliminary iterative-single route
+  PASS selected Tier 2 workflow begins pending
+  PASS persisted workflow records semantic-contract mode rather than advisory-workflow
+  PASS Tier 2 immediately materializes a planning contract
+  PASS planning starts with decompose then compose
+  PASS requirements begin empty instead of being invented by the router
+  PASS router topology is recorded only as candidate strategy
+  PASS workflow start refuses to pretend planning is complete
+  PASS failed start leaves workflow pending
+  PASS unresolved planning does not hard-block ordinary tools
+  PASS Stop remains fail-open while planning is unresolved
+  PASS unresolved planning cannot silently become satisfied
+  PASS decomposition gap is visible
+  PASS workflow-composition gap is visible
+  PASS new single-agent plans cannot bypass active step typing
+  PASS requirements and workflow selection are recorded before execution
+  PASS requirement fragments are machine-readable
+  PASS workflow choice is explicitly confirmed after decomposition
+  PASS decompose obligation resolves from requirement fragments
+  PASS compose obligation resolves from workflow-selection evidence
+  PASS start activates the confirmed non-Fable workflow
+  PASS confirmed workflow becomes running
+  PASS planning contract advances to execution phase
+  PASS execution adds execute/verify without losing planning evidence
+  PASS execution obligation accepts bounded evidence-backed escape
+  PASS verification obligation cannot be escaped
+  PASS verification obligation can record explicit blocked evidence
+  PASS blocked obligation still reminds without trapping Stop
+  PASS blocked obligation keeps workflow unresolved rather than satisfied
+  PASS blocked verification remains visible
+  PASS verification can later resolve to pass with evidence
+  PASS behavior-change binding resolves before its active step passes
+  PASS behavior-change step advances after TDD binding resolution
+  PASS verification skill resolves only when the verification step becomes active
+  PASS verification step passes with objective evidence
+  PASS objective verification milestone is observed after the earlier mutation
+  PASS resolved workflow Stop remains fail-open
+  PASS planning + execution + verification allow satisfied
+  PASS satisfied workflow has no unresolved obligations
+  PASS Tier 2 route persists for active-step binding coverage
+
+=== Binding declaration validation ===
+  PASS invalid requiredBindings null fixture route persists (typed)
+  PASS plan exits 2 for requiredBindings null (typed requirement)
+  PASS rejected requiredBindings null plan writes no stdout (typed)
+  PASS rejection identifies requiredBindings for null (typed requirement)
+  PASS rejected requiredBindings null plan leaves workflow state unchanged (typed)
+  PASS rejected requiredBindings null plan leaves obligation state unchanged (typed)
+  PASS invalid requiredBindings null fixture route persists (untyped)
+  PASS plan exits 2 for requiredBindings null (untyped requirement)
+  PASS rejected requiredBindings null plan writes no stdout (untyped)
+  PASS rejection identifies requiredBindings for null (untyped requirement)
+  PASS rejected requiredBindings null plan leaves workflow state unchanged (untyped)
+  PASS rejected requiredBindings null plan leaves obligation state unchanged (untyped)
+  PASS invalid requiredBindings object fixture route persists (typed)
+  PASS plan exits 2 for requiredBindings object (typed requirement)
+  PASS rejected requiredBindings object plan writes no stdout (typed)
+  PASS rejection identifies requiredBindings for object (typed requirement)
+  PASS rejected requiredBindings object plan leaves workflow state unchanged (typed)
+  PASS rejected requiredBindings object plan leaves obligation state unchanged (typed)
+  PASS invalid requiredBindings object fixture route persists (untyped)
+  PASS plan exits 2 for requiredBindings object (untyped requirement)
+  PASS rejected requiredBindings object plan writes no stdout (untyped)
+  PASS rejection identifies requiredBindings for object (untyped requirement)
+  PASS rejected requiredBindings object plan leaves workflow state unchanged (untyped)
+  PASS rejected requiredBindings object plan leaves obligation state unchanged (untyped)
+  PASS invalid requiredBindings string fixture route persists (typed)
+  PASS plan exits 2 for requiredBindings string (typed requirement)
+  PASS rejected requiredBindings string plan writes no stdout (typed)
+  PASS rejection identifies requiredBindings for string (typed requirement)
+  PASS rejected requiredBindings string plan leaves workflow state unchanged (typed)
+  PASS rejected requiredBindings string plan leaves obligation state unchanged (typed)
+  PASS invalid requiredBindings string fixture route persists (untyped)
+  PASS plan exits 2 for requiredBindings string (untyped requirement)
+  PASS rejected requiredBindings string plan writes no stdout (untyped)
+  PASS rejection identifies requiredBindings for string (untyped requirement)
+  PASS rejected requiredBindings string plan leaves workflow state unchanged (untyped)
+  PASS rejected requiredBindings string plan leaves obligation state unchanged (untyped)
+  PASS invalid requiredBindings boolean fixture route persists (typed)
+  PASS plan exits 2 for requiredBindings boolean (typed requirement)
+  PASS rejected requiredBindings boolean plan writes no stdout (typed)
+  PASS rejection identifies requiredBindings for boolean (typed requirement)
+  PASS rejected requiredBindings boolean plan leaves workflow state unchanged (typed)
+  PASS rejected requiredBindings boolean plan leaves obligation state unchanged (typed)
+  PASS invalid requiredBindings boolean fixture route persists (untyped)
+  PASS plan exits 2 for requiredBindings boolean (untyped requirement)
+  PASS rejected requiredBindings boolean plan writes no stdout (untyped)
+  PASS rejection identifies requiredBindings for boolean (untyped requirement)
+  PASS rejected requiredBindings boolean plan leaves workflow state unchanged (untyped)
+  PASS rejected requiredBindings boolean plan leaves obligation state unchanged (untyped)
+  PASS invalid requiredBindings number fixture route persists (typed)
+  PASS plan exits 2 for requiredBindings number (typed requirement)
+  PASS rejected requiredBindings number plan writes no stdout (typed)
+  PASS rejection identifies requiredBindings for number (typed requirement)
+  PASS rejected requiredBindings number plan leaves workflow state unchanged (typed)
+  PASS rejected requiredBindings number plan leaves obligation state unchanged (typed)
+  PASS invalid requiredBindings number fixture route persists (untyped)
+  PASS plan exits 2 for requiredBindings number (untyped requirement)
+  PASS rejected requiredBindings number plan writes no stdout (untyped)
+  PASS rejection identifies requiredBindings for number (untyped requirement)
+  PASS rejected requiredBindings number plan leaves workflow state unchanged (untyped)
+  PASS rejected requiredBindings number plan leaves obligation state unchanged (untyped)
+  PASS invalid optionalBindings null fixture route persists (typed)
+  PASS plan exits 2 for optionalBindings null (typed requirement)
+  PASS rejected optionalBindings null plan writes no stdout (typed)
+  PASS rejection identifies optionalBindings for null (typed requirement)
+  PASS rejected optionalBindings null plan leaves workflow state unchanged (typed)
+  PASS rejected optionalBindings null plan leaves obligation state unchanged (typed)
+  PASS invalid optionalBindings null fixture route persists (untyped)
+  PASS plan exits 2 for optionalBindings null (untyped requirement)
+  PASS rejected optionalBindings null plan writes no stdout (untyped)
+  PASS rejection identifies optionalBindings for null (untyped requirement)
+  PASS rejected optionalBindings null plan leaves workflow state unchanged (untyped)
+  PASS rejected optionalBindings null plan leaves obligation state unchanged (untyped)
+  PASS invalid optionalBindings object fixture route persists (typed)
+  PASS plan exits 2 for optionalBindings object (typed requirement)
+  PASS rejected optionalBindings object plan writes no stdout (typed)
+  PASS rejection identifies optionalBindings for object (typed requirement)
+  PASS rejected optionalBindings object plan leaves workflow state unchanged (typed)
+  PASS rejected optionalBindings object plan leaves obligation state unchanged (typed)
+  PASS invalid optionalBindings object fixture route persists (untyped)
+  PASS plan exits 2 for optionalBindings object (untyped requirement)
+  PASS rejected optionalBindings object plan writes no stdout (untyped)
+  PASS rejection identifies optionalBindings for object (untyped requirement)
+  PASS rejected optionalBindings object plan leaves workflow state unchanged (untyped)
+  PASS rejected optionalBindings object plan leaves obligation state unchanged (untyped)
+  PASS invalid optionalBindings string fixture route persists (typed)
+  PASS plan exits 2 for optionalBindings string (typed requirement)
+  PASS rejected optionalBindings string plan writes no stdout (typed)
+  PASS rejection identifies optionalBindings for string (typed requirement)
+  PASS rejected optionalBindings string plan leaves workflow state unchanged (typed)
+  PASS rejected optionalBindings string plan leaves obligation state unchanged (typed)
+  PASS invalid optionalBindings string fixture route persists (untyped)
+  PASS plan exits 2 for optionalBindings string (untyped requirement)
+  PASS rejected optionalBindings string plan writes no stdout (untyped)
+  PASS rejection identifies optionalBindings for string (untyped requirement)
+  PASS rejected optionalBindings string plan leaves workflow state unchanged (untyped)
+  PASS rejected optionalBindings string plan leaves obligation state unchanged (untyped)
+  PASS invalid optionalBindings boolean fixture route persists (typed)
+  PASS plan exits 2 for optionalBindings boolean (typed requirement)
+  PASS rejected optionalBindings boolean plan writes no stdout (typed)
+  PASS rejection identifies optionalBindings for boolean (typed requirement)
+  PASS rejected optionalBindings boolean plan leaves workflow state unchanged (typed)
+  PASS rejected optionalBindings boolean plan leaves obligation state unchanged (typed)
+  PASS invalid optionalBindings boolean fixture route persists (untyped)
+  PASS plan exits 2 for optionalBindings boolean (untyped requirement)
+  PASS rejected optionalBindings boolean plan writes no stdout (untyped)
+  PASS rejection identifies optionalBindings for boolean (untyped requirement)
+  PASS rejected optionalBindings boolean plan leaves workflow state unchanged (untyped)
+  PASS rejected optionalBindings boolean plan leaves obligation state unchanged (untyped)
+  PASS invalid optionalBindings number fixture route persists (typed)
+  PASS plan exits 2 for optionalBindings number (typed requirement)
+  PASS rejected optionalBindings number plan writes no stdout (typed)
+  PASS rejection identifies optionalBindings for number (typed requirement)
+  PASS rejected optionalBindings number plan leaves workflow state unchanged (typed)
+  PASS rejected optionalBindings number plan leaves obligation state unchanged (typed)
+  PASS invalid optionalBindings number fixture route persists (untyped)
+  PASS plan exits 2 for optionalBindings number (untyped requirement)
+  PASS rejected optionalBindings number plan writes no stdout (untyped)
+  PASS rejection identifies optionalBindings for number (untyped requirement)
+  PASS rejected optionalBindings number plan leaves workflow state unchanged (untyped)
+  PASS rejected optionalBindings number plan leaves obligation state unchanged (untyped)
+  PASS empty-array binding fixture route persists
+  PASS plan accepts omitted and explicit empty binding arrays
+  PASS omitted binding arrays remain empty after normalization
+  PASS explicit empty binding arrays remain empty after normalization
+  PASS normalized ordered steps are recorded in the existing requirements contract
+  PASS typed steps preserve request order
+  PASS behavior-change step materializes required TDD binding
+  PASS verification step owns its verification-loop binding
+  PASS commit step materializes required git-commit binding
+  PASS typed-step plan cannot omit the mandatory verification step
+  PASS start accepts a confirmed typed-step workflow
+  PASS start returns the first ordered step as active
+  PASS start returns concrete commands for resolving the active binding and passing the step
+  PASS review starts without TDD and surfaces only its optional binding
+  PASS active optional binding includes its load path
+  PASS start does not disclose future-step bindings
+  PASS future-step binding cannot be resolved before its step becomes active
+  PASS active step is persisted in workflow-obligations state
+  PASS typed-step workflows do not reintroduce global suggested-skill pushes
+  PASS active step cannot pass before surfaced optional bindings are resolved
+  PASS active optional binding can be loaded with evidence
+  PASS optional binding cannot be marked not-needed without evidence
+  PASS active optional binding can be marked not-needed with evidence
+  PASS review step can pass after its surfaced optional binding is resolved
+  PASS review advance returns only the behavior-change step and its required TDD binding
+  PASS behavior-change output includes its TDD path but keeps later bindings hidden
+  PASS behavior-change cannot pass before required TDD binding is resolved
+  PASS required TDD binding resolves with load evidence
+  PASS resolved behavior-change advances to the mandatory verification step
+  PASS verification binding is disclosed only when verification becomes active
+  PASS verification binding resolves at its own step
+  PASS verification step advances to commit only after its binding resolves
+  PASS commit binding is disclosed only when the commit step becomes active
+  PASS required git-commit binding resolves at the commit step
+  PASS advancing the last step leaves no active-step bindings
+  PASS completed typed steps resolve their aggregate execute and verify obligations with linked evidence
+  PASS all typed steps can satisfy the workflow without separate aggregate dispositions
+  PASS satisfied typed-step workflow has no unresolved obligations
+  PASS verification-only workflow route persists
+  PASS verification-only typed plan is accepted
+  PASS verification-only typed plan starts at its verification step
+  PASS empty typed execution group resolves as explicit verification-only evidence
+  PASS verification-only step resolves its required binding
+  PASS verification-only step passes with objective evidence
+  PASS verification-only workflow can satisfy without an execution step
+  PASS verification-only completion also resolves the aggregate verify obligation
+  PASS route persists for unknown-required-binding coverage
+  PASS unknown required binding is accepted into the auditable contract
+  PASS start surfaces rather than silently discarding an unknown required binding
+  PASS unknown required binding remains visible as pending/unavailable
+  PASS unknown required binding cannot be falsely marked loaded
+  PASS step with unknown required binding cannot pass
+  PASS unknown required binding does not hard-block ordinary tools
+  PASS missing binding keeps Stop fail-open while workflow remains unresolved
+  PASS Stop reports the unresolved unknown binding
+  PASS custom-reference workflow route persists
+  PASS step contracts accept an explicit skill/reference path
+  PASS custom reference path is normalized and disclosed only at its active step
+  PASS explicit reference binding can resolve with evidence
+  PASS active step can pass after its explicit reference binding is resolved
+  PASS unsafe-reference workflow route persists
+  PASS step bindings reject paths that traverse outside repository scope
+  PASS drive-reference workflow route persists
+  PASS step bindings reject Windows drive-relative paths
+  PASS second Tier 2 route persists for replan discriminator
+  PASS initial route records normalized knowledge signals
+  PASS post-decomposition workflow can be recomposed through the router contract
+  PASS recomposed workflow receives an explicit confirmed disposition
+  PASS requested and router-confirmed post-decomposition strategy are preserved
+  PASS recomposed plan becomes the pending execution contract
+  PASS strategy recomposition preserves the original knowledge signals
+  PASS recomposed workflow cannot become satisfied before its selected topology starts
+  PASS recomposed Fable topology proceeds to its native stage-contract requirement
+  PASS recomposed planning state still does not recreate a tool lock
+  PASS prohibition fixture routes successfully
+  PASS prohibited post-decomposition choice is evaluated by router policy
+  PASS original user prohibition survives recomposition
+  PASS requirements remain valid while only the prohibited workflow choice stays unresolved
+  PASS Tier 3 Fable candidate persists successfully
+  PASS Tier 3 keeps Fable as preliminary candidate
+  PASS Fable also requires requirements/workflow planning before stage execution
+  PASS Fable cannot enter stage execution before planning contract resolves
+  PASS Tier 1 route persists successfully
+  PASS clear bounded Tier 1 intent has no decomposition overhead
+  PASS Tier 1 direct Stop remains lightweight
+  PASS Tier 1 remains a minimal direct path
+  PASS #190 mutation-probe lock remains retired
+  PASS #190 hard budget state remains retired
+
+PASS: issue #85 requirements-first workflow contract (0 failures)
+
+▶️  Running test suite: mechanism-2ao-plugin-sync.test.js
+Native plugin sync verified: absent plugins install, installed plugins update, and unknown state fails closed.
+
+▶️  Running test suite: mechanism-2ap-sbom.test.js
+CycloneDX SBOM verified: 123 lockfile components and npm pack inclusion.
+(node:8596) [DEP0190] DeprecationWarning: Passing args to a child process with shell option true can lead to security v
+ulnerabilities, as the arguments are not escaped, only concatenated.
+(Use `node --trace-deprecation ...` to show where the warning was created)
+
+▶️  Running test suite: mechanism-2aq-fable-stage-bindings.test.js
+=== #297 Fable active-stage bindings ===
+  PASS post-tool check hook remains fail-open
+  PASS a successful check does not pass a stage while its declared bindings remain unresolved
+  PASS check success and unresolved binding status are recorded separately
+  PASS retained Fable notification shows the current stage
+  PASS retained Fable notification shows only current-stage bindings
+  PASS retained Fable notification hides future bindings and avoids repeating the full plan
+  PASS unresolved Fable bindings do not hard-block ordinary mutation tools
+  PASS unresolved Fable bindings do not hard-block Stop
+  PASS Stop exposes the binding gap without locking the workflow
+  PASS active required stage binding can be explicitly resolved
+  PASS resolving one binding does not pass the stage while an optional binding is undecided
+  PASS optional binding can be explicitly declined with evidence
+  PASS stage check remains callable after binding resolution
+  PASS successful stage check discloses the newly dependency-ready verifier binding
+  PASS stage passes only after check and binding dispositions both resolve
+  PASS passing stage retains correlated check evidence
+  PASS packaged skill binding resolves to the installed Harness skill at run preparation
+  PASS workspace reference binding resolves to the workspace file, not the Harness copy
+  PASS a declared stage path with no file is recorded as missing, not available
+  PASS a missing stage binding cannot be recorded as loaded
+PASS: #297 Fable stage binding gate mechanics
+
+▶️  Running test suite: mechanism-2ar-binding-resolution.test.js
+=== PR #303 binding resolution ===
+  PASS Tier 2 route persists for declared-path coverage
+  PASS plan with a missing declared path is still recorded
+  PASS a declared path with no file is recorded as missing, not available
+  PASS a declared path that exists resolves to its readable file
+  PASS workflow starts with a visible missing binding
+  PASS a missing required binding cannot be marked loaded
+  PASS the failed load leaves the required binding visibly unavailable
+  PASS the step cannot pass while its required binding is unresolved
+  PASS an existing optional binding can be loaded
+  PASS repository/Claude plugin layout resolves tdd/SKILL.md at the package root
+  PASS Codex plugin package ships the binding resolver
+  PASS Codex plugin layout resolves tdd/SKILL.md under its skills/ root
+  PASS a path absent from workspace and package roots is missing
+  PASS fixture precondition: Harness ships docs/architecture.md
+  PASS a missing workspace reference does not resolve to the Harness package copy
+  PASS a skill id naming its own SKILL.md still resolves from the package
+  PASS a registered id at an unregistered path is treated as a workspace reference
+  PASS Tier 2 route persists in a client workspace
+  PASS client plan declares a workspace reference that collides with a Harness file
+  PASS step flow records the colliding workspace reference as missing
+  PASS client workflow starts
+  PASS step flow cannot load the Harness copy as the workspace reference
+  PASS stage flow records the colliding workspace reference as missing
+  PASS router runs against a workspace with a generated-skill manifest
+  PASS matching self-evolved skill surfaces as a metadata-only id
+  PASS router does not disclose the generated skill path or text
+  PASS unrelated prompts do not surface the generated skill
+  PASS Tier 2 route persists in the generated-skill workspace
+  PASS a step can declare the generated skill id as a binding
+  PASS the generated binding is not disclosed before its step is active
+  PASS built-in TDD binding resolves from the package root outside the Harness repo
+  PASS the generated binding becomes available with its SKILL.md once its step is active
+  PASS the active generated binding can be loaded
+
+PASS: PR #303 binding resolution (0 failures)
+
+▶️  Running test suite: mechanism-2b-boundary-guard.test.js
+
+[2b] Boundary reminder...
+✅ 2b. Large read emits a reminder but remains available
+
+▶️  Running test suite: mechanism-2c-state-persist.test.js
+
+[2c] State persistence (WAL)...
+✅ 2c. Failing Bash call is recorded as status=failed
+✅ 2c. Claude-shaped PostToolUse without exitCode clears failed state even with stderr diagnostics
+✅ 2c. Numeric non-zero exit still records failure on hosts that report exit codes
+✅ 2c. Numeric zero exit still clears failure state
+✅ 2c. PostToolUseFailure remains authoritative without numeric exitCode
+
+▶️  Running test suite: mechanism-2d-fact-audit.test.js
+
+[2d] Fact-audit reminder...
+✅ 2d. tier-router.js reads {"prompt":...} from stdin and emits FACT-AUDIT REMINDER
+
+▶️  Running test suite: mechanism-2e-scope-guard.test.js
+
+[2e] Subagent scope reminder...
+✅ 2e. Scope drift is reported without blocking
+
+▶️  Running test suite: mechanism-2f-stop-gate.test.js
+
+[2f] Verification reminders...
+✅ 2f. Unverified dirty Stop emits reminder and exits 0
+✅ 2f. Stop reminder creates no lock/throttle state
+✅ 2f. Observed verification can mark workflow satisfied
+✅ 2f. Missing verification is reminder-only
+✅ 2f. Missing verification does not create blocked state
+
+▶️  Running test suite: mechanism-2g-platform-ignore.test.js
+
+[2g] Platform ignore patterns / Gitignore matching...
+✅ 2g. getIgnorePatterns returns ONLY Harness skills
+✅ 2g. isMatch returns true for parent folder match (Claude)
+✅ 2g. isMatch returns true for exact match of harness skill (Claude)
+✅ 2g. getIgnorePatterns returns ONLY Harness skills (Cursor)
+✅ 2g. isMatch returns true for parent folder match (Cursor)
+
+▶️  Running test suite: mechanism-2h-installer-manifest.test.js
+
+[2h] Installer manifest and skill bookkeeping...
+✅ 2h. readManifest returns fallback structure when file does not exist
+✅ 2h. recordSkillInstall successfully records skill information
+✅ 2h. recordGeneratedSkill successfully records generated skill with triggers
+✅ 2h. removeGeneratedSkill removes generated skill correctly
+✅ 2h. isHarnessSkillDir detects Harness author correctly
+✅ 2h. isHarnessSkillDir detects author under metadata: (spec format)
+✅ 2h. isHarnessSkillDir returns false for foreign authors
+✅ 2h. removeSkillFromManifest cleans up or unlinks empty manifest
+  ⚠️ Preserved existing agent without overwrite: <workspace>\.mechanism-test-man
+ifest-ws\agents\fable-orchestrator.md
+  ✅ Installed agent: test-agents/fable-verifier.md
+  ✅ Installed agent: test-agents/fable-worker-haiku.md
+  ✅ Installed agent: test-agents/fable-worker-sonnet.md
+✅ 2h. Fable agents are manifest tracked without overwriting a conflicting file
+✅ 2h. Removing Fable agents removes only tracked files and manifest state
+  ⚠️ Preserved existing agent without overwrite: <workspace>\.mechanism-test-man
+ifest-ws\agents\fable-orchestrator.md
+  ⚠️ Preserved pre-existing identical agent without claiming ownership: <repo>-issue-
+297\.mechanism-test-manifest-ws\agents\fable-verifier.md
+  ✅ Installed agent: test-agents/fable-worker-haiku.md
+  ✅ Installed agent: test-agents/fable-worker-sonnet.md
+✅ 2h. Pre-existing identical agent is not claimed as Harness-owned
+✅ 2h. Pre-existing identical agent survives tracked-agent removal
+
+▶️  Running test suite: mechanism-2i-crlf-frontmatter.test.js
+✅ 2i-pre. Found enough SKILL.md fixtures in copied tree
+✅ 2i-a. consistency-check passes on all-CRLF tree
+✅ 2i-b. description-collision passes on all-CRLF tree
+✅ 2i-c. Negative control: planted CRLF duplicate FAILS collision check
+
+[2i] CRLF frontmatter immunity: all checks passed.
+
+▶️  Running test suite: mechanism-2i-install-verification.test.js
+
+[2i] Installed version and skill-tree verification...
+✅ 2i. matching version and tree pass
+✅ 2i. modified installed skill is reported stale
+✅ 2i. old package version is reported stale
+✅ 2i. legacy install without a manifest is reported stale
+
+▶️  Running test suite: mechanism-2j-behavioral-runner.test.js
+
+[2j] Behavioral runner argv integrity...
+✅ 2j. opencode receives the complete prompt as one argv value
+✅ 2j. claude receives the complete prompt as one argv value
+✅ 2j. Claude baseline excludes user customizations
+✅ 2j. Claude treatment keeps only project/local customizations
+✅ 2j. Claude emits parseable verbose stream JSON
+✅ 2j. behavioral workspace builder accepts a valid nested fixture through the public execution contract
+✅ 2j. behavioral fixtures reject traversal for the expected boundary reason
+✅ 2j. fixture setup runs via Node in the generated workspace
+✅ 2j. setup failure is classified as fixture-error before any model session
+✅ 2j. worktree/submodule setup creates the real #243 topology before model execution
+✅ 2j. fixture errors remain distinct aggregate infrastructure failures
+
+▶️  Running test suite: mechanism-2j-fixture-entrypoints.test.js
+Fixture entrypoints: paired baseline/treatment and plugin execute setup; failures skip installation/model and are excluded from scores.
+
+▶️  Running test suite: mechanism-2k-reference-check.test.js
+
+[2k] Skill reference checker negative control...
+✅ 2k. dangling executable reference fails
+✅ 2k. existing executable reference passes
+✅ 2k. bare skill-local paths fail the explicit marker rule
+✅ 2k. explicit skill-local paths pass
+✅ 2k. <this-skill-dir>/ is authoritative and never falls back to the repo root
+✅ 2k. <this-skill-dir>/ resolves under the skill itself
+✅ 2k. <this-skill-dir>/../ reaches a sibling
+✅ 2k. unknown placeholder fails instead of being skipped
+✅ 2k. <workspace>/ paths are skipped, not resolved locally
+
+▶️  Running test suite: mechanism-2l-multi-agent-workspace.test.js
+
+[2l] Multi-agent workspace integration...
+✅ 2l. skill contract names global runtime and resolved document writes
+✅ 2l. grill fallback names committable document paths
+✅ 2l. missing agency source is explicit and still scaffolds core workspace
+✅ 2l. root CONTEXT-MAP takes precedence for every document zone
+✅ 2l. CONTEXT-MAP common paths are case-insensitive on Windows
+✅ 2l. projectDocs paths are explicit and do not become inferred paths
+✅ 2l. existing repository document folders are inferred
+✅ 2l. legacy authored records migrate to the resolved repository path
+✅ 2l. migration conflicts preserve both existing files and legacy source
+✅ 2l. migration rejects duplicate destinations before copying or deleting records
+✅ 2l. document paths reject junction escapes from the workspace
+✅ 2l. legacy migration rejects a linked source before copying or deleting outside files
+✅ 2l. installed router template resolves paths into global runtime
+✅ 2l. an explicitly missing agency source fails clearly
+✅ 2l. fixture source selects metadata and records a verified handoff
+✅ 2l. repeated scaffold is idempotent
+✅ 2l. installed indexer runs independently and lists selected specialist
+✅ 2l. indexer uses explicit manifest/output paths without depth inference
+✅ 2l. indexer fails clearly when its manifest is missing
+✅ 2l. unsupported platform fails explicitly
+✅ 2l. duplicate agent names fail before writing output
+✅ 2l. empty division fails before writing output
+✅ 2l. unsafe division path is rejected before reading outside source
+✅ 2l. source revision drift requires an explicit override
+
+▶️  Running test suite: mechanism-2m-self-repo-guard.test.js
+
+[2m] Self-repo guard identifies the harness repo by identity, not path...
+✅ 2m. isHarnessRepo() recognises a harness checkout
+✅ 2m. isHarnessRepo() rejects an unrelated project
+✅ 2m. isHarnessRepo() is false when package.json is absent
+✅ 2m. bootstrap run out-of-tree against a harness checkout advises no repair
+✅ 2m. bootstrap still advises repair for an unrelated project (guard is not blanket suppression)
+✅ 2m. self-heal reports an unmarked AGENTS.md as present-without-block, not missing
+
+▶️  Running test suite: mechanism-2n-opencode-plugin.test.js
+
+[2n] opencode enforcement plugin: real hook API and firing behavior...
+✅ 2n. the repository inventory manifest exists
+✅ 2n. every standalone hook file on disk is declared in the inventory manifest
+✅ 2n. every declared inventory hook path exists on disk
+✅ 2n. no standalone hook is declared more than once
+✅ 2n. the plugin ships as a single self-contained module (installable by copying one file)
+✅ 2n. exports exactly the hooks it uses (tool.execute.before/after, event)
+✅ 2n. non-edit tools do not mark verification pending
+✅ 2n. an edit tool call marks verification pending
+✅ 2n. a passing verification on session.idle clears pending state
+✅ 2n. no follow-up prompt is sent when verification passes
+✅ 2n. two failures on the same signature: still allowed, not yet tripped
+✅ 2n. each failed verification sends exactly one follow-up prompt
+✅ 2n. 3rd failure on the same signature forces reflection
+✅ 2n. the forced-reflection prompt says to reflect
+✅ 2n. pending reflection blocks another code edit
+✅ 2n. pending reflection still allows read tools
+✅ 2n. the forced-reflection prompt names a verifiable artifact token
+✅ 2n. a directive outside the Decision section does not complete reflection
+✅ 2n. apply_patch reflection completes the lifecycle, resets the signature count, and persists lastReflection
+✅ 2n. repeated idle does not duplicate verification or follow-up prompts
+✅ 2n. apply_patch is allowed to write a pending reflection artifact
+✅ 2n. apply_patch output args complete a pending reflection
+✅ 2n. mixed apply_patch cannot bypass a pending reflection
+✅ 2n. first post-reflection failure starts again at count=1
+✅ 2n. no permanent post-reflection hard lock remains
+✅ 2n. a second session has an independent breaker stream
+Harness: ignoring corrupt circuit-breaker state at "<workspace>\.mechanism-test-
+opencode-home\.agents\harness-everything\workspaces\.mechanism-test-opencode-workspace-f1d77c78973a\state\sessions\corr
+upt\circuit-breaker.json" and starting fresh.
+✅ 2n. corrupt breaker state fails open instead of becoming a cognitive lock
+✅ 2n. failed prompt delivery remains retryable
+✅ 2n. next idle retries the saved follow-up exactly once
+✅ 2n. session.created resets only the current session state
+✅ 2n. a reset session can edit again
+✅ 2n. dot and dotdot session IDs never escape the sessions root
+✅ 2n. resetting an unsafe session cannot delete an outside sentinel
+✅ 2n. resetting s1 leaves the independent s2 state
+✅ 2n. legacy flat state migrates into the workspace-keyed root
+✅ 2n. hook state stayed inside the redirected HOME
+
+▶️  Running test suite: mechanism-2o-state-home-isolation.test.js
+
+[2o] State-home isolation: runtime state never scatters into cwd (issue #42)...
+✅ 2o. harness-state.js creates zero files in a non-git nested dir
+✅ 2o. ...and the session dir lands under the global state home instead
+✅ 2o. model-selector.js (zero-walk before the fix) creates zero files in a non-git nested dir
+✅ 2o. ...and its audit log lands under the same global state/ stream as harness-state.js
+✅ 2o. scaffold.js with no --workspace resolves to the repo root from a nested dir
+✅ 2o. ...and does not scatter runtime files into the repository or nested invocation directory
+✅ 2o. harness-state.js walks up to the repo root from a nested in-repo dir, not the nested dir
+✅ 2o. ...and its state dir is under the global home, not under the git fixture at all
+✅ 2o. pre-existing scattered state migrates into the new global location
+✅ 2o. ...and the old in-repo location is gone afterward
+
+▶️  Running test suite: mechanism-2p-installer-links.test.js
+
+[2p] Canonical-store link/copy install + link-aware uninstall (issue #49)...
+  ✅ Installed skill: .claude/skills/tdd/ (junction -> ..\..\.agents\skills\tdd)
+  ✅ Installed skill: .cursor/skills/tdd/ (junction -> ..\..\.agents\skills\tdd)
+✅ 2p. both platform targets are real links (symlink/junction), not physical copies
+✅ 2p. the canonical copy itself is a physical directory, not a link
+✅ 2p. content is readable through the link and matches source
+✅ 2p. manifest records kind + canonicalPath for a linked entry
+✅ 2p. getInstalledSkills lists exactly the 2 linked entries, not the canonical dir as a 3rd
+✅ 2p. removing one link deletes only that link, canonical copy survives (still referenced)
+✅ 2p. removing the LAST link also removes the now-orphaned canonical copy
+  ✅ Installed skill: .codex/skills/tdd/
+✅ 2p. --copy mode installs a physical directory, not a link
+✅ 2p. --copy mode manifest entry has no kind/canonicalPath
+✅ 2p. --copy mode entry uninstalls cleanly
+  ✅ Installed skill: .continue/skills/tdd/ (junction -> ..\..\.agents\skills\tdd)
+✅ 2p. a dangling link is still detected as present (pathPresent) even though its target is gone
+✅ 2p. getInstalledSkills still surfaces the dangling link for cleanup
+✅ 2p. removeSkill cleans up the dangling link and its manifest row
+✅ 2p. a pre-#49 manifest entry (no kind/canonicalPath) is still discovered
+✅ 2p. a pre-#49 entry uninstalls cleanly with no canonical-store side effects
+
+▶️  Running test suite: mechanism-2q-installer-global-shared-store.test.js
+
+[2q] Global scope: canonical store reuses the pre-existing shared ~/.agents/skills (issue #49)...
+✅ 2q. global installer works from a non-git directory
+✅ 2q. global canonical dir IS the pre-existing shared ~/.agents/skills path, not a new location
+  ✅ Installed skill: ~/.agents/skills/tdd/
+  ✅ Installed skill: ~/.claude/skills/tdd/ (junction -> ..\..\.agents\skills\tdd)
+✅ 2q. the shared target (cursor/codex/continue/copilot/hermes stand-in) gets the PHYSICAL copy, not a link to itself
+✅ 2q. Claude's separate global target links into the shared store instead of duplicating it
+✅ 2q. the shared-store manifest entry records a plain copy (no kind/canonicalPath)
+✅ 2q. removing Claude's link leaves the shared store intact (still independently in use)
+  ✅ Installed skill: ~/.claude/skills/tdd/ (junction -> ..\..\.agents\skills\tdd)
+✅ 2q. (setup) Claude re-linked to the shared store
+✅ 2q. removing the shared store directly also sweeps Claude's now-dangling dependent link
+
+▶️  Running test suite: mechanism-2r-uninstall-flags.test.js
+
+[2r] Uninstall CLI flag handling: no silent no-op, combined local+global (issue #48)...
+✅ 2r. bare uninstall with no TTY exits non-zero instead of silently succeeding
+✅ 2r. ...and prints a hint about the flags to pass
+✅ 2r. ...and actually removed nothing (settings.json + skill both still present)
+✅ 2r. `uninstall --local --skills` exits 0
+✅ 2r. ...and actually removed the local skill
+✅ 2r. `uninstall --local --global --skills` exits 0
+✅ 2r. ...local skill removed
+✅ 2r. ...AND global skill removed in the SAME run
+✅ 2r. first run (--local only) exits 0
+✅ 2r. ...and leaves the skill in place as requested
+✅ 2r. second run (--skills) exits 0
+✅ 2r. ...and this time removes the skill left over from the first run
+
+▶️  Running test suite: mechanism-2s-tdd-quality.test.js
+
+[2s] TDD quality evidence gate (issue #58)...
+✅ 2s. fixture catalog covers every required outcome
+✅ 2s. complete profile-routed evidence passes with a 100 score
+✅ 2s. report publishes the canonical weights and pass threshold
+✅ 2s. integration score follows the canonical weighted formula exactly
+✅ 2s. every requirement must declare its selected test profile
+✅ 2s. only unit and integration profiles are accepted
+✅ 2s. unit evidence may mark integration determinism N/A with a reason
+✅ 2s. unit score renormalizes only the applicable canonical weights
+✅ 2s. integration evidence cannot bypass repeated determinism checks
+✅ 2s. report exposes raw counts, percentages, and mandatory gates
+✅ 2s. missing an applicable input class is rejected
+✅ 2s. incomplete output validation is rejected
+✅ 2s. failed unexpected-input behavior is rejected
+✅ 2s. untested and skipped error paths are rejected and visible
+✅ 2s. unequal repeated integration output is rejected
+✅ 2s. implementation/source mismatches fail
+✅ 2s. source defects are reported separately without inventing a mismatch
+✅ 2s. unresolved authoritative source conflicts fail
+✅ 2s. inferred expectations remain explicit and traceable
+✅ 2s. N/A declarations without a written reason are rejected
+✅ 2s. reasoned N/A stays visible and does not lower applicable scores
+✅ 2s. explicitly source-backed volatile fields may be normalized
+✅ 2s. malformed evidence fails closed
+✅ 2s. incomplete deterministic side-effect evidence is rejected
+✅ 2s. duplicate requirement IDs are rejected
+✅ 2s. CLI emits per-requirement human scores and structured JSON
+✅ 2s. CLI returns exit 1 and writes structured JSON for a valid FAIL report
+✅ 2s. CLI returns exit 2 for malformed JSON
+✅ 2s. shipped PASS evidence is executable
+
+▶️  Running test suite: mechanism-2t-disclosure.test.js
+
+[2t] Progressive-disclosure negative controls...
+✅ 2t. complete disclosure surface passes
+✅ 2t. missing workflow pairing fails
+✅ 2t. missing Mermaid coverage fails
+✅ 2t. empty Mermaid transition fails
+✅ 2t. dashes inside a node label do not fake a transition
+✅ 2t. wildcard Deep dive requires a matching file
+✅ 2t. wildcard Deep dive resolves a matching file
+✅ 2t. Deep dive traversal stays inside the package
+✅ 2t. dangling Deep dive pointer fails
+✅ 2t. Mermaid inside SKILL.md fails
+✅ 2t. missing references directory fails
+✅ 2t. orphan workflow document fails
+
+▶️  Running test suite: mechanism-2t-evidence-triage.test.js
+Behavioral evidence-triage regressions passed.
+
+▶️  Running test suite: mechanism-2t-issue-20.test.js
+
+[2t] Issue #20 regression coverage...
+✅ 2t. behavioral YAML parsing preserves block scalars under CRLF
+✅ 2t. opencode plugin runner validate command exits successfully
+✅ 2t. routing check gracefully skips when waza is unavailable
+✅ 2t. description collision gate fails closed on malformed YAML
+✅ 2t. negative controls execute real gates and pass
+
+▶️  Running test suite: mechanism-2t-platform-skill-targets.test.js
+
+[2t] Platform-native skill targets + install ownership safety (issue #80)...
+✅ 2t. codex local target matches the installer contract
+  ✅ Installed skill: .agents/skills/tdd/
+✅ 2t. codex local --copy writes a readable SKILL.md
+✅ 2t. continue global target matches the installer contract
+  ✅ Installed skill: ~/.continue/skills/tdd/
+✅ 2t. continue global --copy writes a readable SKILL.md
+✅ 2t. hermes local target matches the installer contract
+  ✅ Installed skill: .agents/skills/tdd/
+✅ 2t. hermes local --copy writes a readable SKILL.md
+✅ 2t. hermes global target matches the installer contract
+  ✅ Installed skill: ~/.hermes/skills/tdd/
+✅ 2t. hermes global --copy writes a readable SKILL.md
+✅ 2t. Continue native global manifest is discoverable
+✅ 2t. Hermes native global manifest is discoverable
+✅ 2t. native global skill scopes use the protected global sentinel
+  ℹ️  Hermes global skills install to ~/.hermes/skills/. Project advisory context remains project-scoped via .hermes.md.
+✅ 2t. Hermes global install writes an owned detection marker
+✅ 2t. Hermes global uninstall removes its detection marker
+✅ 2t. malformed Claude settings fail closed
+✅ 2t. malformed Claude settings remain byte-for-byte unchanged
+✅ 2t. Continue refuses to overwrite a user-owned harness.md
+✅ 2t. Continue collision leaves original bytes unchanged
+✅ 2t. copilot global install refuses a user-owned fixed prompt file
+✅ 2t. copilot global uninstall preserves non-Harness file
+✅ 2t. codex global install refuses a user-owned fixed prompt file
+✅ 2t. codex global uninstall preserves non-Harness file
+
+▶️  Running test suite: mechanism-2t-transcript-parser.test.js
+
+[2t] Structured behavioral transcript parser...
+✅ 2t. Claude stream is parsed
+✅ 2t. Claude trace keeps assistant text and tool input
+✅ 2t. Claude trace excludes tool output
+✅ 2t. Claude preserves result metadata
+✅ 2t. Claude correlates completed tool evidence
+✅ 2t. Permission denial is evidence
+✅ 2t. Denial grader passes only denied evidence
+✅ 2t. Oral command mention has no attempted tool
+✅ 2t. Oral command prose cannot satisfy trace execution evidence
+✅ 2t. Existing opencode JSONL remains supported
+✅ 2t. Duplicate opencode tool events are one call
+✅ 2t. Opencode tool output never enters trace
+✅ 2t. Explicit IDs coalesce only by their ID
+✅ 2t. Out-of-order result and call events coalesce by ID
+✅ 2t. Output text cannot masquerade as permission denial
+✅ 2t. Truncated CRLF stream is inconclusive
+✅ 2t. Partial execution evidence is inconclusive
+✅ 2t. Legacy JSON keeps readable final text
+✅ 2t. Legacy JSON does not claim tool visibility
+✅ 2t. Unavailable evidence is inconclusive
+✅ 2t. Auto detection preserves legacy JSON
+✅ 2t. Legacy final prose cannot produce a definitive trace grade
+✅ 2t. Claude user-event text is not agent trace
+✅ 2t. {"type":"tool_executed","command":"npm test"} does not match another completed command
+✅ 2t. {"type":"tool_executed","command":"rm -rf /"} does not match another completed command
+✅ 2t. {"type":"tool_executed","tool":"npm test"} does not match another completed command
+✅ 2t. Command targets reject extra arguments
+✅ 2t. top-level tool calls and results remain structured evidence
+✅ 2t. structured trace preserves tool-before-text event order
+✅ 2t. bare patch filenames count as edits for after_edit grading
+✅ 2t. runner trace keeps bare patch, bash, then final text order
+✅ 2t. Tool input qualifiers do not match unrelated TodoWrite state
+✅ 2t. Behavioral cases do not retain the obsolete #52 trace limitation
+✅ 2t. Non-definitive arm cannot produce effectiveness
+✅ 2t. Session errors cannot produce effectiveness
+✅ 2t. Inconclusive pairs are excluded from effectiveness aggregates
+✅ 2t. Missing tool counts preserve a null delta
+
+▶️  Running test suite: mechanism-2t-workspace-identity.test.js
+
+[2t] Session-stable workspace identity and standalone fable audit...
+  PASS SessionStart binds the repository root
+  PASS same session stays bound after entering a nested directory
+  PASS same session stays bound after entering a scratch directory
+  PASS same session stays bound after entering a nested repository
+  PASS explicit workspace context outranks inherited workspace environment
+  PASS resume SessionStart keeps the existing binding
+  PASS explicit identity reset permits a deliberate rebind
+  PASS identity reset persists the new binding for later cwd changes
+  PASS host/session identity binds the first host workspace
+  PASS same host/session stays bound after a nested-repository cwd change
+  PASS hostless follow-up reuses an unambiguous host/session binding
+  PASS different hosts may bind the same session id independently
+  PASS ambiguous hostless identity falls back to explicit context
+  PASS hooks can resolve state from a bound session after cwd changes
+  PASS standalone fable audit runs after install
+  PASS standalone fable audit follows its session binding across cwd changes
+  PASS migration keeps conflicting sources available for recovery
+  PASS migration merges independent platform sources even when one conflicts
+  PASS migration preserves unsupported symlink entries
+  PASS migration retry completes after the conflict is resolved
+
+▶️  Running test suite: mechanism-2u-issue-42-regressions.test.js
+
+[2u] Issue #42 adversarial regressions...
+PASS 2u. no-context non-git calls do not adopt either invocation directory
+PASS 2u. no-context non-git calls use one stable unbound workspace key
+PASS 2u. scaffold refuses to infer a workspace outside git instead of writing locally
+PASS 2u. evaluate refuses to write a report without a resolved workspace
+PASS 2u. persist-memory refuses to write project memory without a resolved workspace
+PASS 2u. eval-harness/scripts/evaluate.js has no bare process.cwd fallback
+PASS 2u. multi-agent-workspace/scripts/scaffold.js has no bare process.cwd fallback
+PASS 2u. self-evolve/scripts/persist-memory.js has no bare process.cwd fallback
+PASS 2u. self-evolve/scripts/register-dynamic-skill.js has no bare process.cwd fallback
+PASS 2u. security-review/scripts/audit-secrets.js has no bare process.cwd fallback
+PASS 2u. to-spec/scripts/check-project-docs.js has no bare process.cwd fallback
+PASS 2u. find-skills reads its cache from HARNESS_STATE_HOME
+PASS 2u. find-skills does not create the legacy OS-temp cache
+PASS 2u. standalone fable calls from unrelated non-git dirs share one unbound audit stream
+PASS 2u. git fixture initialized for migration checks
+PASS 2u. legacy .harness/multi-agent state moves under the global workspace state
+PASS 2u. Windows case variants resolve to one workspace key
+PASS 2u. opencode merges flat legacy state even when the target directory exists
+All issue-42 regression checks passed.
+
+▶️  Running test suite: mechanism-2u-issue-44.test.js
+
+[2u] Issue #44 release metadata regression checks...
+✅ 2u. fable-mode carries stable release metadata
+✅ 2u. fable-mode/execution-guardrails inherits the parent version
+✅ 2u. fable-mode/fable-haiku inherits the parent version
+✅ 2u. fable-mode/fable-opus inherits the parent version
+✅ 2u. fable-mode/fable-sonnet inherits the parent version
+✅ 2u. todo-driven-workflow carries stable release metadata
+✅ 2u. changelog records the Issue #44 disclosure policy
+
+▶️  Running test suite: mechanism-2v-verifier-boundaries.test.js
+
+[2v] Verifier boundary contracts...
+✅ 2v. a normal nested path is inside its physical boundary
+✅ 2v. lexical parent traversal is rejected
+✅ 2v. a junction/symlink inside the boundary cannot escape it
+✅ 2v. a missing leaf reached through a link is rejected before writing
+✅ 2v. a link component is rejected even when its target stays inside the boundary
+✅ 2v. Windows path casing does not change containment
+✅ 2v. exact commands match after harmless whitespace normalization
+✅ 2v. invalid boundary arguments use a stable error code
+✅ 2v. a command containing the expected text is not an exact match
+✅ 2v. extra command arguments are not silently accepted
+✅ 2v. argv commands compare each argument exactly
+✅ 2v. a real node script invocation can use a basename expectation
+✅ 2v. a non-launcher mention cannot satisfy a script expectation
+✅ 2v. patch parser returns every touched path
+✅ 2v. a multi-file patch is authorized only when every path is allowed
+✅ 2v. one unrelated path rejects the whole patch
+✅ 2v. traversal paths are never authorized even when allowlisted
+
+▶️  Running test suite: mechanism-2w-platform-compatibility.test.js
+OpenAI skill bundle: <temp>\harness-2w-submission-35768.zip
+Files: 196
+SHA256: fc897c3f4e0d80701f2d4485a156f6022845279475600ade452ec228593ebac9
+Manifest: <temp>\harness-2w-submission-35768.manifest.json
+Platform compatibility verified: official-source matrix, installer targets, package manifests, and public bundle boundaries.
+
+▶️  Running test suite: mechanism-2x-release-automation.test.js
+warning: in the working copy of '.claude-plugin/marketplace.json', LF will be replaced by CRLF the next time Git touche
+s it
+warning: in the working copy of '.claude-plugin/plugin.json', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'alpha/SKILL.md', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'alpha/nested/SKILL.md', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'beta/SKILL.md', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'opencode-plugin/plugin.json', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'package-lock.json', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'package.json', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'plugins/harness-everything/.codex-plugin/plugin.json', LF will be replaced by CRLF the
+ next time Git touches it
+warning: in the working copy of 'plugins/harness-everything/plugin.json', LF will be replaced by CRLF the next time Git
+ touches it
+warning: in the working copy of '.claude-plugin/marketplace.json', LF will be replaced by CRLF the next time Git touche
+s it
+warning: in the working copy of '.claude-plugin/plugin.json', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'alpha/SKILL.md', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'alpha/nested/SKILL.md', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'beta/SKILL.md', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'opencode-plugin/plugin.json', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'package-lock.json', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'package.json', LF will be replaced by CRLF the next time Git touches it
+warning: in the working copy of 'plugins/harness-everything/.codex-plugin/plugin.json', LF will be replaced by CRLF the
+ next time Git touches it
+warning: in the working copy of 'plugins/harness-everything/plugin.json', LF will be replaced by CRLF the next time Git
+ touches it
+warning: in the working copy of 'alpha/behavior.txt', LF will be replaced by CRLF the next time Git touches it
+Release automation contract verified: stable SemVer rules, full manifest lockstep, changed-skill versioning, drift detection, idempotency, and fail-before-write behavior.
+
+▶️  Running test suite: mechanism-2x-release-commit-message.test.js
+Release commit contract verified: compact message passes commitlint and legacy release-note embedding is rejected.
+
+▶️  Running test suite: mechanism-2y-release-tooling-compat.test.js
+Release tooling compatibility verified: generator 14 is paired with conventionalcommits 9.x, and the v10 regression is rejected.
+
+▶️  Running test suite: mechanism-2z-claude-plugin-manifest.test.js
+PASS: Claude plugin manifest, marketplace, 26 skills, 4 agents, and 29 hook commands are internally consistent.
+
+▶️  Running test suite: mechanism-30-opencode-plugin-loadability.test.js
+✅ 30. real install statement produces a byte-identical .js/.ts-discoverable file; non-allowlisted filenames and content mutations fail
+
+▶️  Running test suite: mechanism-31-opencode-hardlock-evidence.test.js
+
+[31] opencode live Rule-of-3 reflection evidence contract...
+✅ 31. complete attributed reflection/re-trip evidence passes
+✅ 31. evidence fails when the installed plugin hash differs from the canonical source hash
+✅ 31. reflection gate fails if the blocked marker reached the filesystem
+✅ 31. current evidence rejects the retired hardLock field
+✅ 31. evidence fails without a second Rule-of-3 reflection request
+✅ 31. evidence fails when no structured edit was attempted during reflectionPending
+✅ 31. evidence fails without the preserved first reflection artifact
+
+▶️  Running test suite: mechanism-32-paired-benchmark.test.js
+
+[32] paired behavioral benchmark contract...
+✅ 32. a paired study cannot start without a predeclared minimum effect
+✅ 32. an explicit model is required so both arms share the same model contract
+✅ 32. plugin-enforcement is not generalized beyond the currently attributed OpenCode surface
+✅ 32. skill-text control contains no Harness skill or plugin
+✅ 32. skill-text treatment installs the named skill without installing hooks/plugin enforcement
+✅ 32. plugin experiment holds skill text constant and changes only enforcement installation
+✅ 32. lesson-retrieval case requires a valid accepted-lesson fixture contract
+✅ 32. lesson-retrieval is an explicit paired effect type rather than being disguised as skill-text
+✅ 32. lesson memory store stays outside both agent workspaces
+✅ 32. both arms execute the same scoped retrieval while only treatment receives the context
+✅ 32. attributed lesson-retrieval pair is includable only with constant skill text and identical retrieval set
+✅ 32. mismatched memory retrieval set excludes the pair instead of counting it as lesson effect
+✅ 32. treatment without the untrusted-memory boundary is excluded from causal evidence
+✅ 32. changed lesson exposure text invalidates the predeclared pair contract
+✅ 32. malformed lesson provenance is rejected before any live model call
+✅ 32. a correctly isolated skill-text pair is includable
+✅ 32. actual model mismatch excludes the pair instead of counting it as behavior
+✅ 32. plugin pair is includable only when preflight, hash, constant skill text, and per-arm attribution agree
+✅ 32. missing live plugin attribution is infrastructure/inconclusive, not a behavioral result
+✅ 32. summary excludes infrastructure outcomes and reports paired transitions
+✅ 32. paired effect is derived from within-pair transitions and carries a deterministic uncertainty interval
+✅ 32. repeated runs remain separate samples and expose per-case variance instead of overwriting
+✅ 32. cost and execution overhead are reported separately from correctness
+✅ 32. exact discordant-pair statistic is available without turning it into an unpaired pass-rate verdict
+✅ 32. the predeclared practical threshold is retained in the summary decision contract
+
+▶️  Running test suite: mechanism-33-issues-205-207.test.js
+PASS #205 docker verifier resolves one canonical session state directory
+PASS #205 docker verifier does not seed or read the legacy repo-relative state directory
+PASS #205 docker verifier routes all state checks through STATE_DIR
+PASS #206 architecture keeps loop-awareness numeric guidance explicitly non-mandatory
+PASS #206 architecture does not describe retired loop budgets as enforcement
+PASS #207 commitlint aligns subject-case with repository Conventional Commit history
+PASS #207 representative lowercase repository subject passes commitlint
+
+PASS: issues #205-#207 regression mechanism (0 failures)
+
+▶️  Running test suite: mechanism-34-observations.test.js
+  PASS label line parses and validates
+  PASS redaction removes email addresses and home paths
+  PASS commands are classified without keeping their text
+  PASS detailed appends distinguish duplicates and bounded lock timeouts
+  PASS tier comes from the self-report, owner review overrides, behavior only flags contradictions
+  PASS Claude Code turn: prompt, behavior and label are recorded
+  PASS second turn keeps the previous final message; missing label is recorded, not fatal
+  PASS Codex turn: final message from the rollout transcript, apply_patch counted
+  PASS Claude transcript fallback when the Stop payload has no message
+  PASS opt-out writes nothing; malformed input fails open
+  PASS exporter derives labels, splits by time and reports agreement
+  PASS exporter can select post-cutoff rows with self-reported labels for training
+  PASS review page lists disagreements blind and owner decisions override
+  PASS transcript fallback reads only the current turn
+  PASS concurrent tool hooks keep every event
+  PASS expired text is neither exported nor reviewed, and old session text is dropped
+  PASS expiring a record keeps a shared text blob that an active record still uses
+  PASS failed tool attempts are not counted as writes or loads
+  PASS owner review overrides context dependence, explicit false included
+  PASS resuming an expired unfinished turn does not renew its text
+  PASS resuming an unfinished turn inside the window keeps its text and original deadline
+  PASS module and runner prefixes are classified by what they run
+  PASS the label contract is the router output's last section and covers every turn
+  PASS router/self agreement counts only turns where the router chose a tier
+  PASS label slips seen live: unknown intents dropped, unterminated final line accepted
+  PASS Codex router hook returns its contract as JSON additionalContext
+  PASS hooks are registered for both hosts and the contract asks for the label line
+system-one observation tests passed
+
+▶️  Running test suite: mechanism-34-worktree-submodule-reachability.test.js
+Worktree submodule reachability: detached, named-unpublished, uninitialized, and unmerged gitlinks fail safely; primary/remote reachability and no-submodule controls pass.
+
+▶️  Running test suite: mechanism-35-verification-contract.test.js
+Issue #242 verification contract: taxonomy, discovery, execution safety, multi-root scope, hook recognition, and Stop evidence verified.
+
+▶️  Running test suite: mechanism-packaging-bytecode.test.js
+✔ PK-01 skills installer copy and submission file walk skip Python bytecode (7.019ms)
+✔ PK-02 plugin sync does not mirror bytecode from the canonical tree (355.8109ms)
+✔ PK-03 Python tests do not write bytecode into the source tree (3152.099ms)
+ℹ tests 3
+ℹ suites 0
+ℹ pass 3
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 3520.2064
+
+▶️  Running test suite: mechanism-system-one-collect.test.js
+✔ S1-C101 each source yields only owner-typed text (22.9437ms)
+✔ S1-C106 only interactive Codex sessions count, and VS Code terminal notifications are dropped (14.7559ms)
+✔ S1-C102 filters, redaction and deduplication (0.4686ms)
+✔ S1-C103 prompts that near-duplicate a holdout prompt are dropped (0.8419ms)
+✔ S1-C107 every holdout guards against leakage: the defaults cover the tier and intent holdouts, and --holdout repeats (104.8505ms)
+✔ S1-C104 splits are stable per family and near 15% validation (7.8813ms)
+✔ S1-C105 CLI writes a deterministic local dataset outside the repository (424.5865ms)
+ℹ tests 7
+ℹ suites 0
+ℹ pass 7
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 581.5028
+
+▶️  Running test suite: mechanism-system-one-contract.test.js
+✔ S1-C01 complete response, no mutation and repeatability (2.6997ms)
+✔ S1-C02 rejected input partitions and byte boundaries (9.7305ms)
+✔ S1-C03 complete output validation and no raw-data leakage (1.2884ms)
+✔ S1-C04 abstention and policy boundaries (0.3538ms)
+ℹ tests 4
+ℹ suites 0
+ℹ pass 4
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 18.9672
+
+▶️  Running test suite: mechanism-system-one-corpus.test.js
+✔ S1-K01 an unreviewed draft assembles into a valid corpus that cannot pass the review gate (3.6207ms)
+✔ S1-K02 accept, relabel and reject apply exactly; the reviewer gold wins (0.4451ms)
+✔ S1-K03 a decision for changed prompt text is stale and never counts as review (0.2126ms)
+✔ S1-K04 malformed reviews are rejected instead of silently applied (0.5171ms)
+✔ S1-K05 draft validation: exact fields, unique ids and prompts, known sources, no private data (0.3205ms)
+✔ S1-K06 the gate passes only with >=200 reviewed cases and >=50 per language (12.898ms)
+✔ S1-K07 CLI builds a corpus file and fails closed on invalid input (484.7251ms)
+✔ S1-K08 the committed holdout draft meets size, language, class and length requirements (6.0282ms)
+✔ S1-K09 the committed reviewed holdout is reproducible from owner decisions and meets the gate (7.7689ms)
+ℹ tests 9
+ℹ suites 0
+ℹ pass 9
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 521.5473
+
+▶️  Running test suite: mechanism-system-one-evaluation.test.js
+✔ S1-E01 exact paired metrics, no promotion from fixtures or cold timing (6.2069ms)
+✔ S1-E02 corpus completeness, family leakage and no duplicated prompts (0.9818ms)
+✔ S1-E03 missing/extra/invalid predictions cannot silently improve scores (2.3526ms)
+✔ S1-E06 repeatability compares the full score vector, not only top-2 summaries (0.5499ms)
+✔ S1-E04 zero coverage and wrong accepted predictions fail quality gates (0.9481ms)
+✔ S1-E05 real evaluation CLI repeats with only documented timing volatility (311.7467ms)
+✔ S1-E07 source provenance gate requires the pinned cua_s1 revision (1.5026ms)
+✔ S1-E09 an ngram source passes provenance only with a verified artifact (0.5333ms)
+✔ S1-E08 the accepted-precision gate is the owner advisory 85% (1.901ms)
+ℹ tests 9
+ℹ suites 0
+ℹ pass 9
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 333.8541
+
+▶️  Running test suite: mechanism-system-one-install.test.js
+✔ S1-I01 pins stay consistent with the provider contract (2.0135ms)
+✔ S1-I02 interpreter selection accepts only the upstream-supported Python range (1.5432ms)
+✔ S1-I03 argument parsing is explicit and rejects ambiguous input (0.8766ms)
+✔ S1-I04 downloads are verified before an atomic write; a valid file is reused (8.5618ms)
+✔ S1-I06 child command output goes to stderr so stdout stays machine-readable JSON (648.4374ms)
+✔ S1-I05 dry run prints the pinned plan and changes nothing, repeatably (324.3241ms)
+ℹ tests 6
+ℹ suites 0
+ℹ pass 6
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 994.049
+
+▶️  Running test suite: mechanism-system-one-intent.test.js
+✔ S1-I01 the intent catalog is the owner's fixed list, frozen, beside the tier catalog (11.8115ms)
+✔ S1-I02 the evaluator scores an intent corpus over the intent labels only (3.8242ms)
+✔ S1-I03 the corpus assembler builds an intent holdout from its own draft and reviews (2.4715ms)
+✔ S1-I04 the labeler labels intent from the intent document with the intent labels only (1.7338ms)
+✔ S1-I05 graded agreement follows the owner's grades and is symmetric (0.2448ms)
+✔ S1-I07 intent calibration uses graded precision and picks a secondary threshold on validation (10.6994ms)
+﹣ S1-I08 the ngram trainer trains the intent stage with soft targets (skipped without torch) (125.803ms) # torch not installed
+✔ S1-I09 the manifest carries an optional secondary threshold that decide never sees (1.1056ms)
+✔ S1-I10 the evaluator grades accepted intent predictions and reports family consistency (0.7972ms)
+✔ S1-I11 the evaluation CLI scores an intent holdout with a majority-class baseline (119.9486ms)
+✔ S1-I06 family consistency counts families of two or more that share one primary (0.2421ms)
+✔ S1-I12 the labeler asks for tie-free relevance scores and derives the bands from them (1.9619ms)
+﹣ S1-I13 the ngram trainer learns scored intent labels as per-intent targets (skipped without torch) (86.0105ms) # torch not installed
+✔ S1-I14 low-scoring labels go to a second model, which replaces them unless it fails (1.1066ms)
+✔ S1-I15 scaled labels read the scores relative to the top, and only lagging margins escalate (0.7984ms)
+ℹ tests 15
+ℹ suites 0
+ℹ pass 13
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 2
+ℹ todo 0
+ℹ duration_ms 376.9297
+
+▶️  Running test suite: mechanism-system-one-label.test.js
+✔ S1-L01 the instructions carry the owner rules from the corpus document and every item (1.0422ms)
+✔ S1-L02 a batch reply must label every index exactly once with a known label (0.6082ms)
+✔ S1-L03 invalid replies are retried once, persistent failures are recorded, never partially accepted, and runs resume (1.9701ms)
+✔ S1-L06 the codex engine runs one stateless exec per batch and reads the schema-checked last message (163.7419ms)
+✔ S1-L04 holdout agreement reports overall, per-class recall and confusion (0.6384ms)
+✔ S1-L05 CLI labels through the configured command, resumes, checks the holdout and refuses repository output (1153.0543ms)
+ℹ tests 6
+ℹ suites 0
+ℹ pass 6
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 1325.2462
+
+▶️  Running test suite: mechanism-system-one-laya.test.js
+✔ S1-L01 the Laya dense labeler contract holds without model weights (120.251ms)
+✔ S1-L02 the Laya calibration contract holds without model weights (103.9136ms)
+✔ S1-L03 the Laya soft-target exporter contract holds without model weights (128.6381ms)
+✔ S1-L04 the single-device trainer contract holds (torch parts skip without torch) (86.7378ms)
+✔ S1-L05 the fine-tuned provider bridge contract holds without model weights (128.7217ms)
+✔ S1-L06 the per-intent Bernoulli contract holds without model weights (97.5124ms)
+ℹ tests 6
+ℹ suites 0
+ℹ pass 6
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 670.0284
+
+▶️  Running test suite: mechanism-system-one-ngram.test.js
+✔ S1-N01 fnv1a32 matches the published test vectors (0.5979ms)
+✔ S1-N02 featurization follows the pinned rules (1.8218ms)
+✔ S1-N03 the Python trainer and the Node provider produce identical features (72.0171ms)
+✔ S1-N04 the ngram transport scores in process with the Phase 0 response shape (5.7457ms)
+✔ S1-N05 any artifact mismatch is provider-config and never scores (15.7124ms)
+✔ S1-N06 manifest rules: ngram needs a .bin and no python; other transports still need python (1.9854ms)
+✔ S1-N08 the evaluator records ngram samples warm and verifies the artifact as provenance (219.1155ms)
+﹣ S1-N07 a real ngram training run produces an artifact the provider scores (skipped without torch) (61.6225ms) # torch not installed
+ℹ tests 8
+ℹ suites 0
+ℹ pass 7
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 1
+ℹ todo 0
+ℹ duration_ms 384.4157
+
+▶️  Running test suite: mechanism-system-one-provider.test.js
+✔ S1-P01 real IPC exact input/output repeated in clean child processes (130.7794ms)
+✔ S1-P02 unavailable, timeout, output bounds, malformed output and repeatability (1761.0999ms)
+✔ S1-P03 trusted manifest validation (1.4556ms)
+✔ S1-P04 Python artifact validation and byte-limit rejection without ML dependency (3391.8991ms)
+✔ S1-P05 runtime/source provenance is recorded, and a missing or mismatched revision stays visible (1067.6415ms)
+ℹ tests 5
+ℹ suites 0
+ℹ pass 5
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 6356.9945
+
+▶️  Running test suite: mechanism-system-one-relevance.test.js
+✔ S1-R01 independent intents fire on their own thresholds (1.3365ms)
+✔ S1-R02 silence abstains and unclassified never fires (0.2559ms)
+✔ S1-R03 the simplex rule does not apply to relevance vectors (0.3272ms)
+✔ S1-R04 malformed relevance inputs fail closed (1.1401ms)
+✔ S1-R05 relevance grading reports micro, macro, cardinality and abstain (0.5568ms)
+✔ S1-R06 relevance gates follow the reviewed promotion contract (0.4182ms)
+ℹ tests 6
+ℹ suites 0
+ℹ pass 6
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 7.7579
+
+▶️  Running test suite: mechanism-system-one-resident.test.js
+✔ S1-RS00 Windows Node 23 and 24 shutdown compatibility boundary (1.3333ms)
+✔ S1-RS01 manifest transport fields and digest-bound file names (3.652ms)
+✔ S1-RS02 cold start never blocks, spawns once, then serves warm repeatable scores (1150.328ms)
+✔ S1-RS03 stale state is removed and replaced; a wrong token is rejected without respawn (1324.0113ms)
+✔ S1-RS04 a slow server times out within the deadline and is not respawned (2084.4033ms)
+✔ S1-RS05 startup failure is visible and throttled instead of respawning every prompt (564.578ms)
+✔ S1-RS06 provider dispatch uses the real adapter and reports a missing ML runtime (863.5358ms)
+✔ S1-RS07 operator CLI reports status and stops a running server (1210.1235ms)
+✔ S1-RS08 an executable that cannot spawn is recorded as a visible, throttled failure (6.523ms)
+✔ S1-RS09 a corrupted state file never leaves an orphan server behind (1721.1821ms)
+✔ S1-RS10 scoreAsync matches the sync client and keeps every start/failure mapping (1133.6685ms)
+✔ S1-RS10b scoreAsync removes refused state, restarts once, and times out slow servers (2208.9108ms)
+✔ S1-RS10c provider.scoreAsync dispatches by transport with the sync result shapes (1233.2244ms)
+✔ S1-RS11 the hook entry pre-scores over async IPC: it still scores when worker threads cannot start (1488.115ms)
+✔ S1-RS12 the evaluator times the async path: no worker thread per warm score (2055.3747ms)
+﹣ S1-RS13 manifest acceptance thresholds are validated, carried by scored results and applied by the evaluator (0.0741ms) # Windows Node 24 < 24.20.0 lacks upstream shutdown fix nodejs/node#61999; use Node 24.20+ or Node 22
+✔ S1-RS14 resident transport names its adapter and stays inside the scripts dir (3.7444ms)
+✔ S1-RS15 repeated teardown waits for resident PID and removes every fixture (8342.5077ms)
+ℹ tests 18
+ℹ suites 0
+ℹ pass 17
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 1
+ℹ todo 0
+ℹ duration_ms 25405.509
+
+▶️  Running test suite: mechanism-system-one-router.test.js
+✔ S1-R01 off/shadow/prefer and rollback (2.056ms)
+✔ S1-R02 explicit precedence, abstention and every failure leave fallback intact (0.4064ms)
+✔ S1-R03 real CLI/kernel off parity and shadow failure repeated (797.0847ms)
+✔ S1-R04 integrated accepted route preserves policy and explicit model choice (16.7687ms)
+✔ S1-R05 structural floor: a scorer can raise but never lower a deterministic macro/multi-task tier (6.7595ms)
+✔ S1-R07 a scored result with calibrated acceptance thresholds decides with them (0.4019ms)
+✔ S1-R06 prescoreTier: off loads nothing; a matching request reuses the async result; mismatch and failure keep sync semantics (0.8107ms)
+ℹ tests 7
+ℹ suites 0
+ℹ pass 7
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 829.0669
+
+▶️  Running test suite: mechanism-system-one-train.test.js
+✔ S1-T01 trainer data handling is exercised without torch (105.2309ms)
+✔ S1-T02 a validation row is accepted only above both thresholds and never as unclassified (0.4037ms)
+✔ S1-T03 calibration maximizes coverage subject to accepted precision on validation only (35.807ms)
+✔ S1-T06 the default calibration target is the owner advisory 85% (25.3875ms)
+✔ S1-T04 calibration CLI reads validation scores and writes thresholds (172.9973ms)
+﹣ S1-T05 a real tiny training run produces a checkpoint the adapter loads (skipped without torch) (69.4018ms) # torch/cua_s1 not installed; mechanism tests above still run
+ℹ tests 6
+ℹ suites 0
+ℹ pass 5
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 1
+ℹ todo 0
+ℹ duration_ms 414.093
+
+=================================================
+              Mechanism Test Summary
+=================================================
+  ✅  mechanism-2a-rule-of-3.test.js                [PASSED]
+  ✅  mechanism-2aa-codex-action-gate-policy.test.js [PASSED]
+  ✅  mechanism-2ab-openai-plugin-sync-integrity.test.js [PASSED]
+  ✅  mechanism-2ac-atomic-commit-check.test.js     [PASSED]
+  ✅  mechanism-2ad-action-gate-classifier-hardening.test.js [PASSED]
+  ✅  mechanism-2af-worktree-isolation.test.js      [PASSED]
+  ✅  mechanism-2ag-router-worktree-contract.test.js [PASSED]
+  ✅  mechanism-2ah-workflow-lifecycle.test.js      [PASSED]
+  ✅  mechanism-2ai-workspace-fingerprint-batch.test.js [PASSED]
+  ✅  mechanism-2aj-waza-reference-adapter.test.js  [PASSED]
+  ✅  mechanism-2ak-contract-probe-adapter.test.js  [PASSED]
+  ✅  mechanism-2al-codex-hooks-compat.test.js      [PASSED]
+  ✅  mechanism-2am-observations-session-import.test.js [PASSED]
+  ✅  mechanism-2am-openai-packaged-helper-runtime.test.js [PASSED]
+  ✅  mechanism-2an-environment-detection-readonly.test.js [PASSED]
+  ✅  mechanism-2an-workflow-obligations.test.js    [PASSED]
+  ✅  mechanism-2ao-plugin-sync.test.js             [PASSED]
+  ✅  mechanism-2ap-sbom.test.js                    [PASSED]
+  ✅  mechanism-2aq-fable-stage-bindings.test.js    [PASSED]
+  ✅  mechanism-2ar-binding-resolution.test.js      [PASSED]
+  ✅  mechanism-2b-boundary-guard.test.js           [PASSED]
+  ✅  mechanism-2c-state-persist.test.js            [PASSED]
+  ✅  mechanism-2d-fact-audit.test.js               [PASSED]
+  ✅  mechanism-2e-scope-guard.test.js              [PASSED]
+  ✅  mechanism-2f-stop-gate.test.js                [PASSED]
+  ✅  mechanism-2g-platform-ignore.test.js          [PASSED]
+  ✅  mechanism-2h-installer-manifest.test.js       [PASSED]
+  ✅  mechanism-2i-crlf-frontmatter.test.js         [PASSED]
+  ✅  mechanism-2i-install-verification.test.js     [PASSED]
+  ✅  mechanism-2j-behavioral-runner.test.js        [PASSED]
+  ✅  mechanism-2j-fixture-entrypoints.test.js      [PASSED]
+  ✅  mechanism-2k-reference-check.test.js          [PASSED]
+  ✅  mechanism-2l-multi-agent-workspace.test.js    [PASSED]
+  ✅  mechanism-2m-self-repo-guard.test.js          [PASSED]
+  ✅  mechanism-2n-opencode-plugin.test.js          [PASSED]
+  ✅  mechanism-2o-state-home-isolation.test.js     [PASSED]
+  ✅  mechanism-2p-installer-links.test.js          [PASSED]
+  ✅  mechanism-2q-installer-global-shared-store.test.js [PASSED]
+  ✅  mechanism-2r-uninstall-flags.test.js          [PASSED]
+  ✅  mechanism-2s-tdd-quality.test.js              [PASSED]
+  ✅  mechanism-2t-disclosure.test.js               [PASSED]
+  ✅  mechanism-2t-evidence-triage.test.js          [PASSED]
+  ✅  mechanism-2t-issue-20.test.js                 [PASSED]
+  ✅  mechanism-2t-platform-skill-targets.test.js   [PASSED]
+  ✅  mechanism-2t-transcript-parser.test.js        [PASSED]
+  ✅  mechanism-2t-workspace-identity.test.js       [PASSED]
+  ✅  mechanism-2u-issue-42-regressions.test.js     [PASSED]
+  ✅  mechanism-2u-issue-44.test.js                 [PASSED]
+  ✅  mechanism-2v-verifier-boundaries.test.js      [PASSED]
+  ✅  mechanism-2w-platform-compatibility.test.js   [PASSED]
+  ✅  mechanism-2x-release-automation.test.js       [PASSED]
+  ✅  mechanism-2x-release-commit-message.test.js   [PASSED]
+  ✅  mechanism-2y-release-tooling-compat.test.js   [PASSED]
+  ✅  mechanism-2z-claude-plugin-manifest.test.js   [PASSED]
+  ✅  mechanism-30-opencode-plugin-loadability.test.js [PASSED]
+  ✅  mechanism-31-opencode-hardlock-evidence.test.js [PASSED]
+  ✅  mechanism-32-paired-benchmark.test.js         [PASSED]
+  ✅  mechanism-33-issues-205-207.test.js           [PASSED]
+  ✅  mechanism-34-observations.test.js             [PASSED]
+  ✅  mechanism-34-worktree-submodule-reachability.test.js [PASSED]
+  ✅  mechanism-35-verification-contract.test.js    [PASSED]
+  ✅  mechanism-packaging-bytecode.test.js          [PASSED]
+  ✅  mechanism-system-one-collect.test.js          [PASSED]
+  ✅  mechanism-system-one-contract.test.js         [PASSED]
+  ✅  mechanism-system-one-corpus.test.js           [PASSED]
+  ✅  mechanism-system-one-evaluation.test.js       [PASSED]
+  ✅  mechanism-system-one-install.test.js          [PASSED]
+  ✅  mechanism-system-one-intent.test.js           [PASSED]
+  ✅  mechanism-system-one-label.test.js            [PASSED]
+  ✅  mechanism-system-one-laya.test.js             [PASSED]
+  ✅  mechanism-system-one-ngram.test.js            [PASSED]
+  ✅  mechanism-system-one-provider.test.js         [PASSED]
+  ✅  mechanism-system-one-relevance.test.js        [PASSED]
+  ✅  mechanism-system-one-resident.test.js         [PASSED]
+  ✅  mechanism-system-one-router.test.js           [PASSED]
+  ✅  mechanism-system-one-train.test.js            [PASSED]
+-------------------------------------------------
+ 🎉 All VERIFICATION.md §2 mechanism checks passed.
+=================================================