Pre-existing ordering flake on main: test_fig_two_by_two_quadrant_labels fails under the Step 9c co-selection prefix (empty mpl title) while passing single-file

kind: infra

Goal

Root-cause and fix the main-side test-ordering interaction that makes tests/test_issue2162_figures.py::test_fig_two_by_two_quadrant_labels fail when executed after the Step 9c gate's co-selection prefix, while passing single-file and in isolation. Until fixed, any issue whose gate set includes this file can red on it and burn a compare/pristine cycle.

Measured evidence (#2168 Step 9c, 2026-08-18 — mechanical, not prose)

• Gate #2 (474-file set, 1h47m): the test FAILED with assert '1 cells without probe rows omitted' in '' — the spy on F._save captured ax.get_title() == '' on fig.axes[0], i.e. the title was never set (or axes[0] was not the scatter axes) under full-suite state.
• step9c_baseline.py compare --run-pristine --max-paired-files 500: the node PASSES the single-file pristine oracle on a scratch main tree (568440568f7e) but REPRODUCES under the gate's own 406-file co-selection prefix run ON PRISTINE MAIN — compare's ORDERING WARN classified it ordering_suspect: a pre-existing cross-module interaction on main, not branch-linked (the #2168 diff is except-tuple widenings; the test's imports are untouched).
• Compare JSON: /tmp/step9c-compare-issue-2168.json (ephemeral); durable record in #2168 events.jsonl (epm:test-verdict marker, 2026-08-18).

Suspected class

matplotlib global-state pollution (rcParams / style / figure-registry) by some predecessor in the prefix — the same broad class as the known tick_triage module-level-NOW flake (#2369) but a distinct mechanism and file. The contaminating predecessor is somewhere in the 406-file prefix; #2021 measured a distance-49 contaminator for a prior interaction of this shape, so nearest-N windowing is not a valid search shortcut — bisect the prefix instead.

Acceptance criteria

• The contaminating predecessor (or the figures module's vulnerable global-state read) is identified and fixed so the test passes under the full co-selection prefix.
• A regression guard at whichever grain fits (test isolation fixture, module-state reset, or the figure function setting its title unconditionally).
• No weakening of the test's assertions.

Provenance

Mandated by the compare ORDERING WARN follow-through ("route a workflow-fix candidate at the interaction itself — never a silent pass"), #2168 Step 9c verdict, 2026-08-18.