build-review-test-quality

Purpose

Judge the Test Quality concern for one engine-managed build_review rubric branch. This is a judgement-only contract: the engine owns scope selection, evidence assembly, result validation, finding identity, the stamped result envelope, and the outer gate verdict.

Input projection (v3)

Use only the supplied projection version v3. Its closed input contains:

• the lap ID, snapshot digest, and top-level contentDigest;
• the in-scope changed tests: only changed tests with a current-feature-owned, resolvable Covers: binding introduced or updated after the review base, pointing to an approved story criterion, an active feature requirement (FR-N), or a task Done when: check, represented as immutable content-region references;
• the changed diff by reference (changedFiles: per-file path, change kind, and hunk ranges), anchored by mergeBase and headSha;
• the current code-valid test_suite PASS; and
• typed reverted-production preflight evidence, including its source identities, classification, scoped-run result, executed selectors, and bounded failure excerpt when applicable.
• any concrete fallback candidates in testScope, each with its engine-established candidate ID, pinned source region, and allowed Covers obligation references. Unchanged legacy bare markers are excluded before projection; their ordinals never gain authority from the active plan.
• testScope.evidence records are identity-only references: source, region, startLine, endLine, and contentHash. They are not embedded source content or a replacement for the engine's authority.

The session runs inside the feature worktree. The diff content is not embedded: read referenced files and obtain any per-path diff with git diff <mergeBase>..HEAD -- <path> (or the merge-base form with git show <mergeBase>:<path>). Those reads are part of this closed input. Do not infer facts from a maker transcript, task-status narrative, prior review, or state outside this projection and its referenced content.

Judge only the reviewScope delta hunks. Treat changedFiles as context only. If that review identifies a finding outside the delta, set causeHunk to the delta hunk that causes it.

For every testScope.evidence region you inspect, re-read the file at the pinned ref for its source side (mergeBase for base regions, headSha for head regions). The region is the raw bytes from byteRegion.start (inclusive) to byteRegion.end (exclusive) of that output. byteRegion is in UTF-8 bytes; region, startLine, and endLine are character positions for identity and orientation only, and diverge from byte offsets once the file holds a non-ASCII character. The engine computed contentHash from that same immutable commit. Do not recompute it (a read-only review session has no hashing tool); copy it verbatim where the schema asks for it. A candidate's evidence record is the one whose id equals its candidateId. An unreadable region, one whose pinned read fails, is not judged: return its fallback candidate as indeterminate with a non-empty missingEvidenceReason. Your own read never becomes authoritative. The engine rejects a finding anchored to a contentHash that is absent from the projected evidence and candidates.

Judgement

For each in-scope changed test, judge whether its assertion actually distinguishes the behavior its Covers: binding names. Raise test-insensitive only when the test has a concrete, stub-passable assertion: it could pass while the changed behavior is absent or replaced with a stub. Keep independent tests or behaviors as separate findings.

The reverted-production preflight is evidence, never a finding by itself. Report its optional counterfactualSensitivity judgement using this closed vocabulary:

• supports means either an executed in-scope example fails on the reverted tree, or the reverted production causes the intended tests to fail during collection or load.
• indeterminate means an environment failure prevents the intended tests from bearing on behavior before that can be determined — for example, the #1915 database-auth or boot failures. It is neither sensitivity support nor a finding.
• not-applicable means the counterfactual evidence does not apply to a sensitivity judgement.

A stayed-green result is not automatically a concern: read the test and cite the concrete stub-passable assertion before finding it insensitive. An infrastructure failure is not a finding; do not invent evidence, downgrade it to a pass, or turn it into content criticism. Tests outside the supplied in-scope set are not this rubric's concern.