bug(review): assess rejects already_reviewed passive delta when exact candidate is unconsumed #1954

Before submitting

• I searched open and closed issues and did not find a report of this problem.
• I reviewed this report and removed credentials, tokens, private paths, hostnames, and other sensitive data.

These checks were performed by the reporting agent. Searches included already_reviewed, "passive delta", "acknowledged_predecessor", and "consumed" "assess". Related #1175 concerns optional reason fields, and #654 concerns acknowledgement losing its committed selector; neither matches this reproduction.

Problem

The Shell assessment decoder rejects a native assessment after an acknowledged committed review is followed only by a passive ODD task note.

Native Gentle AI correctly reports review_due: false, review_due_reason: "already_reviewed", and candidate.consumed: false. The exact new candidate has not been consumed; its executable changes are covered by the acknowledged committed predecessor. A fresh inspection of the new range also stops with acknowledged_predecessor_passive_delta.

The facade instead returns risk: "unassessable" with schema-incompatible, losing the native medium risk and review eligibility evidence and falling back to high-risk-equivalent verification. No data loss or unsafe approval was observed.

The conflicting check remains present in current main and in the tested source commit: validateReviewDueConsistency in lib/review-risk-assessment.ts requires already_reviewed if and only if candidate.consumed === true:

const alreadyReviewed = reviewDueReason === REVIEW_DUE_REASON.ALREADY_REVIEWED;
if (alreadyReviewed !== (candidate.consumed === true)) {
  throw new TypeError("review assessment review_due_reason already_reviewed must be reported exactly when candidate.consumed is true");
}

This equivalence excludes the native acknowledged-predecessor passive-delta case. The exact candidate must remain unconsumed; accepting the assessment must not manufacture closure or delivery authority.

Steps to reproduce

1. Use the beta versions below, with RDD enabled, in a disposable Git repository.
2. Commit a baseline containing increment(value) returning value + 1 and a passing test for increment(1) == 2. Retain this commit as <base-commit>.
3. Add double(value) returning value * 2, plus a test for double(3) == 6. Leave increment and its original test unchanged. Observe RED before adding double, then GREEN for both tests; commit this additive change.
4. Run an ordinary committed-only native review against <base-commit> with the reliability lens through the facade. Complete the provider-issued capture and exact acknowledgement. In this reproduction the review approved and acknowledgement returned authority: "burned".
5. Add and commit only odd/tasks/beta-note.md, containing a short passive note about the approved candidate. Do not change executable files or the original base selector.
6. Call gentle_review with operation: "assess", the fixture workspaceRoot, and input: {"baseRef":"<base-commit>","committedOnly":true}.
7. Compare with the read-only native command for the same committed range:

gentle-ai review assess --cwd=<fixture> --base-ref=<base-commit> --committed-only --json

8. A fresh facade inspection of the new range with the same base selector returns the terminal stop acknowledged_predecessor_passive_delta. This is inspection of the new candidate, not a bound STATUS replay of the consumed lineage.

The fixture was removed after collecting the results. This was an actual approved committed review, not a mocked receipt or an acknowledgement of pending workspace changes.

Expected and actual behavior

Expected: accept the supported native acknowledged-predecessor assessment, preserve candidate.consumed: false, expose reviewDue: false and already_reviewed, and derive verification from the native risk tier without inventing exact-candidate closure.

Actual: the facade rejects the native envelope, returns risk and nativeRisk as unassessable, candidate: null, and applies its conservative high-risk fallback. The native command exits 0 with a usable medium-risk assessment, and fresh inspection recognizes the passive delta.

Suggested regression: decode this native envelope with already_reviewed, review_due: false, and candidate.consumed: false, while retaining malformed-envelope rejection and exact-candidate consumed/closure semantics.

gentle-pi version

4.0.0; tested source commit 12a69a36524d4f239aefc3db7a25bce99c9e1bf5. Native Gentle AI beta override: 4.0.1-0.20261008053821-b15198ba3876. This report concerns the beta pairing, not a claim about the bundled stable binary.

Pi version

1.1.0

Operating system

Linux

Relevant logs or error output (optional)

Native assessment (exit 0; base commit replaced with a public-safe placeholder):
{
  "schema": "gentle-ai.review-assessment/v1",
  "risk": "medium",
  "reasons": [{"code": "executable_change", "path": "calc.py"}],
  "changed_paths": 3,
  "changed_lines": 10,
  "candidate": {
    "kind": "base-diff",
    "base_ref": "<base-commit>",
    "consumed": false
  },
  "review_due": false,
  "review_due_reason": "already_reviewed"
}

Facade assessment excerpt:
"risk": "unassessable"
"nativeRisk": "unassessable"
"reasons": [{
  "code": "schema-incompatible",
  "detail": "native review assess failed; no sanitized stderr diagnostic is available."
}]
"candidate": null
"plan": {
  "writerSelfVerification": true,
  "structuralReadbackOnly": false,
  "independentVerifier": true,
  ...
}

Fresh inspection of the new range:
"next_transition": {
  "kind": "stop",
  "reason_code": "acknowledged_predecessor_passive_delta"
}