---

kind: infra workflow: v1

Step 9c paired oracle: add a SUFFIX-replay arm so later-sorting-file contamination is attributed mechanically instead of surfacing as an unattributed NEW failure

Goal

Close the documented residual blind class in the Step 9c paired-prefix oracle: test-order contamination originating in a file that sorts AFTER the candidate is present in the gate process but absent from every prefix replay, so the oracle cannot reproduce it and ordering_suspect comes back empty. Today that class surfaces as a fail-closed NEW classification against an innocent issue's gate, and the cost is paid by whoever happens to be merging.

Evidence

• The blind class is documented, not hypothetical: .claude/skills/issue/SKILL.md Step 9c 1d names it directly.
• #2059's gate hit it. A 239-file gate run failed tests/test_issue1090_fu3_dispatcher.py::test_conv_context_is_wildchat_family while the same file passed alone and passed in the single-file pristine-main oracle, so step9c_baseline.py compare classified the node NEW (fail-closed). The paired-prefix oracle over 34 prefix files did NOT reproduce it → ordering_suspect: []. The contaminator (tests/test_issue1481_analysis.py) sorts after the victim, so no prefix replay could ever contain it.
• Cost of the miss: a manual provenance-override on #2059, plus an entire diagnosis cycle on #2217 (which found the offender in one full-tree --collect-only sweep with a per-collector registry-delta plugin — exactly ONE offender out of 24,775 collected tests).
• #2217 shipped a mechanical recurrence guard for this class AT SOURCE (tests/conftest.py collection hooks + tests/test_no_import_time_registry_mutation.py), but its watch list is exactly CONTEXTS + NEGATIVE_PANELS. The guard's own docstring names the residual: a NEW module-level registry (fu4.ROUNDS / RUNS_BY_ROUND, trait registries, columns.CONTEXTS) must EXTEND the conftest hook or it is unwatched. The #703 env/logging contamination family is likewise outside a registry-key guard.
• Deliberately filed separately rather than ridden into #2217: the arm is an estimated ~150-300-line diff on scripts/step9c_baseline.py, which is 3,919 lines and fleet-critical (every /issue session's gate runs it). Landing it on a test-hygiene fix would have broken one-commit-per-leg revertibility and put the whole fleet's gate at risk behind an unrelated change.

Suggested approach

1. Add a SUFFIX-replay arm to the paired oracle: when a node classifies NEW and the prefix replay yields no ordering_suspect, replay the candidate together with the files that sort AFTER it (the complement of the prefix set), bisecting to attribute the offender.
2. Keep it bounded — the arm exists to attribute a failure the gate already saw, so it should trigger only on the NEW-with-empty-ordering_suspect path, never on every run.
3. Report the attributed offender in the compare output so the failing session gets the offender name instead of an unexplained NEW node.
4. Because the target is fleet-critical, the change wants its own smoke over a synthetic later-sorting contaminator (the #2217 incident reshaped is the natural fixture) plus a negative control proving a clean tree still classifies nothing.

Acceptance criteria

• A synthetic later-sorting contaminator, replayed through the oracle, is ATTRIBUTED by name rather than returning ordering_suspect: [].
• A clean tree produces no new suspects and no change in verdict (no false-positive regression).
• tests/test_step9c_baseline.py covers both arms.
• The existing prefix-replay behavior is unchanged on every path that already attributes an offender.

Provenance

Carried action item (1) recorded on #2217 before its Step 10d merge, filed after that merge landed. Recommended file-separately in #2217 plan v2 §4 leg 4 + §11.6 and endorsed independently by the Alternatives critic, whose note was that an unfiled Decision line is how this blind class lost its owner the first time.