bug(sdd): disabled failed verification cannot enter unmanaged remediation

[dnlrsls]

Pre-flight Checklist

• I have searched existing issues and this is not a duplicate
• I understand that PRs will be rejected if the linked issue does not have status:approved

📝 Bug Description

When receipt-driven review is disabled and recorded as disabled/unmanaged, an admitted failed SDD verification cannot enter bounded remediation because sdd-status still requires a bounded review transaction that policy intentionally prevented from being created.

This is distinct from #2128, which covered archive-gate treatment of discovered invalidated receipts. Here the verification failure is substantive and correctable, but remediation remains unavailable: remediationState.required is false and the dispatcher reports bounded review transaction is missing.

🔄 Steps to Reproduce

1. Start and complete an OpenSpec-backed SDD change with receipt-driven review disabled.
2. Record delivery as disabled/unmanaged and complete every implementation task.
3. Persist a validator-admitted gentle-ai.verify-result/v1 report with verdict fail, correct non-zero requirement/scenario totals, and at least one substantive blocker.
4. Settle the failed verification and explicitly authorize a bounded remediation work unit.
5. Run gentle-ai sdd-status <change> --cwd <repo> --json --instructions.
6. Observe that apply, verify, and archive remain blocked because no bounded review transaction exists.

✅ Expected Behavior

An admitted failed verification under legitimate disabled/unmanaged delivery should preserve the concrete failure and permit one explicitly authorized bounded correction followed by fresh independent verification. It must not require review authority that the kill switch prevented from existing, and it must never fabricate approval.

❌ Actual Behavior

The dispatcher rejects remediation and reports a blocker equivalent to:

verify evidence cannot enter remediation: <substantive verification reason>; bounded review transaction is missing

remediationState.required remains false, so no supported correction route is available.

Gentle AI Version

v2.2.4 and v2.2.5 pre-release

Operating System

Windows

AI Agent / Client

OpenCode

📋 Affected Area

CLI (commands, flags)

💡 Logs / Error Output

nextRecommended: resolve-blockers
remediationState.required: false
blockedReasons:
  - verify evidence cannot enter remediation: blockers must be zero for archive readiness; bounded review transaction is missing

Additional Context

Split from the follow-up FAIL occurrences on #2137. The original #2137 scope remains PASS evidence misclassified after requirement-count mismatch.

Related but distinct:

• fix(sdd): archive gate blocks on a discovered invalidated receipt while review mode is disabled #2128 fixed disabled archive-gate provenance, not failed-verification remediation.
• fix(sdd): accept approved successor when failed verify lacks native evidence revision #1649 concerns failed evidence without a native evidence revision and an approved successor.
• bug(sdd): passing verify stays blocked at resolve-review — verify total 11 does not match actual requirement count 0; bounded review transaction is missing (v2.2.4) #2137 concerns requirement counting and PASS-to-remediation misrouting.

Acceptance criteria:

• Correctly counted, admitted FAIL evidence under disabled/unmanaged can expose one maintainer-authorized bounded correction route.
• Enabled review mode continues to require native review authority.
• Explicit invalid review artifacts and unreadable mode state remain fail-closed.
• No path reports or implies review approval while review is disabled.

[MarsSall]

Confirmed on the current upstream main HEAD after rebuilding and installing gentle-ai 2.2.5-0.20260801165051-b6ecb3ef6896.

Environment

• Gentle AI: 2.2.5-0.20260801165051-b6ecb3ef6896
• OS/architecture: Windows x64
• Client: OpenCode 1.18.11
• Artifact mode: repository-local OpenSpec
• Effective receipt-driven review mode: globally off, recorded as disabled/unmanaged

Bounded retest and outcome

1. Updated the installed binary to the exact current main commit and verified toolchain health passed 8/8 with zero warnings.
2. Queried structured SDD status once against the preserved failed-verification state.
3. No acquire, reset, remediation, review, repair, or source mutation was performed.
4. All 30 tasks remain complete, while apply, verify, and archive remain blocked with nextRecommended: resolve-blockers.
5. The blocker is unchanged: verify result total 15 does not match actual requirement count 0; bounded review transaction is missing.
6. remediationState.required remains false, with an empty failed-evidence revision, lineage, and binding revision in dispatcher status.

Expected

The admitted FAIL evidence under legitimate disabled/unmanaged delivery should preserve the 15-requirement failure and expose one explicitly authorized bounded correction route, without fabricating review approval.

Actual

Current main still counts 0 requirements and requires review authority that the kill switch prevented from existing, leaving no supported remediation route.

Safe opaque evidence

• Runtime ledger revision: sha256:6b980454c3838106e3b04998b32a1c88b2285a182ecfb3153d831f29a5c91756
• Failed verification evidence revision: sha256:1587cab5fddfa5335af8b532624ff547b336dd9e2a1927b38f5fd08fb93e4434
• Verify report hash: sha256:155c735375a465987dc010073dfcd4176414259848fc9c464842d50429c69d94

Preserved state

The retest was read-only. It changed no source, tests, SDD artifacts, review authority, Git lifecycle state, remotes, deployments, or external data. No root cause is asserted.

[Alan-TheGentleman]

I reproduced this on current main; it remains a current bug and must stay separate from #2185 and #2137. I need a bounded authority design before approval: explicit maintainer runtime authorization bound to the admitted failed-evidence revision, one bounded unmanaged correction route, and fresh verification afterward. That design must issue no receipt or approval under disabled/unmanaged, leave enabled-mode review unchanged, and fail closed for invalid artifacts or unreadable mode state. I am moving this to status:needs-design pending that design.

[sinapsecl-dev]

New occurrence on stable gentle-ai 2.2.4 with receipt-driven review disabled globally. This reproduces the same dead end after a substantive failed verification and a maintainer-authorized bounded correction.

Observed evidence

• Environment: macOS arm64, OpenCode 1.18.11.
• Independent verification correctly failed with 3 critical findings (29/35 scenarios compliant) and persisted failed evidence revision sha256:8eec148d1c494cbcf832678f4f6fe0cbc876f6ce2c12f54af720b89adea4eab4.
• The maintainer explicitly authorized one bounded correction. Native SDD objective generation 13 completed in one attempt with a 122-line correction.
• Focused correction tests passed: 28 focused, 172 registry, and 134 cumulative. Full manifest loading, syntax checks, YAML parsing, and diff checks passed.
• Delivery and release remained disabled. No review transaction was created, as required by the global kill switch.
• After correction, sdd-status still reports apply complete but verify/archive blocked, routes to resolve-review, and says the failed verification cannot enter remediation because a bounded review transaction is missing.

Expected

The completed maintainer-authorized correction should permit a fresh independent verification while delivery remains disabled/unmanaged. No review receipt or approval should be required or fabricated.

Actual

There is no supported transition to fresh verification. Starting review would violate the disabled-mode contract, while the dispatcher requires review authority that cannot exist.

Minimal reproduction

1. Disable receipt-driven review globally.
2. Complete an SDD apply and persist an admitted failed verify report with substantive blockers.
3. Explicitly authorize and complete one bounded correction through the native attempt ledger.
4. Keep delivery/release disabled and create no review transaction.
5. Query structured SDD status.
6. Observe resolve-review and the missing bounded-review-transaction blocker instead of fresh verification readiness.

Consumer state is preserved: no commit, push, merge request, archive, external delivery, or review mutation occurred after the correction.

[dnlrsls]

Accepted design: unmanaged failed-verification remediation

The fix will extend the existing native SDD runtime ledger rather than create authority inside the disabled review system.

Command and authority

Add a typed reset disposition to the existing mutation surface:

gentle-ai sdd-attempt reset \
  --disposition failed-evidence-remediation \
  --remediates-evidence-revision <failed-evidence-revision> \
  --work-unit <bounded-remediation-label> \
  --evidence-goal <bounded-remediation-goal> \
  --max-changed-lines <positive-limit> \
  --maintainer-authorization <exact-binding>

The authorization must be persisted in the Git-common-dir runtime ledger and bound to the current runtime revision, change, predecessor objective and generation, admitted failed-evidence revision, failed candidate identity/tree, disabled/unmanaged delivery, fixed remediation scope, actor, and reason. It creates exactly one successor objective with max_attempts: 1; previous attempts, ordinals, cumulative charges, and lifetime charges remain immutable and monotonic.

State transitions

1. A current, substantive verification FAIL with a matching terminal failed runtime attempt exposes authorize-remediation.
2. An explicit maintainer authorization appends one immutable authorization/reset record under the existing runtime lock and CAS.
3. The next sdd-attempt begin must match the authorized work unit, evidence goal, target, and line ceiling exactly.
4. The correction must change the candidate and finish with non-zero line accounting, distinct correction evidence, diagnosis, harness disposition, cleanup evidence, and process evidence.
5. A successful correction routes only to fresh independent verification. Archive remains blocked until a new admitted PASS evaluates the corrected candidate.
6. Fresh FAIL, interruption, over-budget work, or failed cleanup consumes the only unmanaged correction and cannot open another attempt.

Safety invariants

• This path must never create or imply a review transaction, receipt, approval, binding, or GateAllow.
• Enabled review mode continues using the existing bounded-review remediation path unchanged.
• Unreadable or unknown mode, explicit invalid/foreign/ambiguous review state, stale or mismatched evidence, target drift, malformed ledger state, and conflicting replay fail closed.
• Exact request replay is idempotent; changed input under the same request ID is rejected.
• Re-enabling review mode before authorization, begin, or finish invalidates unmanaged completion.
• A consumed authorization cannot be reset or renamed into another correction budget.

Compatibility and recovery

The runtime ledger remains the only authority and CAS domain. Existing reset/begin/finish behavior and v1 records remain compatible. The new operation uses the same locking, immutable publication, atomic HEAD update, Git common-dir sharing, bounded reads, symlink rejection, and pre-/post-HEAD crash-recovery behavior. OpenSpec and Engram consume the same native status and do not persist independent counters or authorization.

Required proof

Implementation must include behavior-first coverage for OpenSpec and pure Engram routing, Windows and Linux, exact replay and conflicting replay, concurrent authorization, evidence/target/scope mismatch, enabled and unreadable mode, explicit invalid review state, linked worktrees, one-attempt exhaustion, cleanup/process evidence, fresh verification, and pre-/post-HEAD publication failures.

Add one benchmark E2E journey that starts from an admitted disabled/unmanaged FAIL, authorizes and consumes one bounded correction, proves no review authority was created, records a fresh independent PASS, reaches archive through delivery=disabled/unmanaged, and proves replay cannot create a second correction.

Scope boundaries

This does not change #2137/#2185 stale non-FAIL routing, #2128 archive-gate provenance, #1649 missing failed-evidence revisions, or enabled-mode review remediation. No second authorization store or generic unchanged-candidate reset is permitted.

With this contract accepted, the issue is implementation-ready.

[desatatufuria]

@dnlrsls, this occurrence appears to exercise the accepted design you documented above and adds a stale-evidence boundary worth covering explicitly.

Additional occurrence with stale verification totals

This is an additional sanitized occurrence on released Gentle AI 2.2.4, Linux x86_64, using OpenCode with repository-local OpenSpec artifacts. Receipt-driven review was globally disabled and remained disabled/unmanaged.

Bounded history and mutation outcome

Before final verification, 12 bounded objectives were recorded: 11 passed and one broad release objective was interrupted before mutation so it could be safely re-sliced. Previous evidence, budgets, and authority history remained preserved. No Git common-dir authority files were manually edited or deleted.

Final verification correctly returned FAIL for a substantive missing public CLI composition requirement. The admitted report declared 13 requirements and 27 scenarios, while subsequent native status independently counted 28 scenarios. The failed evidence revision was sha256:8b59eb032fc678b12eef7609a40006f9f10ada6d8bd59d8b14876ee83a0984ed.

After the failed evidence was settled, structured status routed to resolve-review, reported both the scenario-total mismatch and a missing bounded review transaction, and exposed no authorized evidence-refresh, apply, or unmanaged-remediation transition. No remediation mutation occurred, no review authority was fabricated, and consumer state remained preserved.

Minimal reproduction

1. Initialize an OpenSpec-backed SDD change.
2. Disable receipt-driven review globally.
3. Complete implementation through bounded runtime objectives.
4. Admit a substantive failed verification report whose scenario total is stale.
5. Settle the failed verification evidence.
6. Query structured SDD status.
7. Observe resolve-review, the totals mismatch, the missing review transaction, and no authorized evidence-refresh or unmanaged-remediation route.

Expected

The stale verification total should permit one bounded evidence correction or fresh verification. Once consistent failed evidence is admitted, disabled/unmanaged mode should expose the accepted evidence-bound unmanaged remediation flow without requiring or fabricating review authority.

Actual

The consumer remains blocked even though all authority and evidence are preserved. Changing branches, merging, or rewriting the local report cannot produce a valid native transition because the authority lives in the Git common-dir.

This appears to exercise the same disabled failed-verification dead end already covered by the accepted design, while adding a stale-evidence boundary that the implementation should handle explicitly.