Case curation and oracle challenge

Dated 2026-09-19, before subject dispatch. Work order GHI #1045.

The main-session agent curated H from the original handoff complaint and the historical producer, both renderers, session-start budget handling and resume skill. The issue's claim that traversal was absent was already false at the selected snapshot: resume_handoff populated chain, while the renderers did not consume it. The task retains the observed symptom rather than prescribing the issue's diagnosis. Expected criteria therefore reward finding existing machinery and its omitted consumers, not building a second traversal.

A quality-reviewer independently derived C from issue #989 and historical acceptance input identity, review admission, readiness, provenance checks, tests and governance prose, before comparing the repair diff. The existing test and prose explicitly prescribed environmental invalidation. The repair's changed-file list omitted that prose, demonstrating why historical patch recall is not the oracle. Interpreter/platform policy is not a mandatory scored remedy.

A second quality-reviewer independently derived S from original issues #771/#708 and historical sync, trailer validation, tests and rules. The initial suggested “fingerprint reuse” case was a mistaken recollection: this was auto-add staging and semantic attribution. Local history resolved the repair as 45b4a51a1 and its parent as da4470b8d. The reviewer also withdrew a proposed fail-closed criterion after reading the historical function and tests' explicit fail-open inspection-error policy. Those corrections preceded the frozen criteria and all subject runs.

The C curator then challenged the H oracle and common protocol against the extracted snapshot. Accepted with one clarification, incorporated before freeze: H3 can preserve bounded traversal while explicitly declining completeness; it must not demand new truncation metadata or traversal/API redesign. H1–H6 and the translation of C's oracle were otherwise accepted. S is explicitly a weaker hidden-discovery case and a useful state/attribution control.

All six runs use fresh contexts with inherited parent model settings and no model override. The mapping lives in inputs.json; the scorer receives response labels and case identities, but not that mapping. Content may reveal treatment.

Snapshots use git archive of the recorded parent commit. Python's safe tar filter rejected three absolute/outside harness symlinks in H and C; each omitted path and reason is listed in cases.json. All ordinary source, tests, docs and configuration remain. No Git checkout, worktree or production workflow was created. Historical snapshots are temporary and are not installed in main.