workflow-fix: slow-tests roster misses the live-bundle no-flags lint test

Goal

Stop tests/test_workflow_lint_upload_or_true.py::test_no_flags_default_run_bundles_check from deterministically timeout-killing an implementer's local Step-9c-matched pytest run, by either registering the file in scripts/select_step9c_tests.py's slow-tests roster with a measured surcharge, or refactoring the test to the sub-second fixture-tree form its own sibling already uses.

The gap

That test executes a FULL live-tree no-flags workflow_lint.main([]) in-process. Measured: >533 s for that single test; the whole live bundle runs ~25-35 min. The file is ABSENT from select_step9c_tests.py's slow-tests roster, so the selector neither defers it nor sizes a surcharge for it — and an implementer running the gate-matched selection locally under the 600 s Bash cap gets a deterministic rc=124 timeout kill with no indication of which test stalled.

The roster exists precisely to prevent this class. The omission is the bug.

Driving incident — #2253 round 1 (2026-08-21)

The #2253 implementer hit three consecutive timeout-killed chunk attempts before isolating the offender via --collect-only index-splitting. Evidence recorded in that round's epm:results v1:

• /tmp/i2253-upload-b.out — rc=124 on a 10-test slice, stall at the 8th id.
• tests/test_workflow_lint_upload_or_true.py:465-473 — no _REPO_ROOT monkeypatch, so the test builds no fixture tree and runs against the real repo.

Note the failure mode is silent-by-construction: a timeout kill names no test, so each attempt costs a full 600 s before yielding any diagnostic. That is what turned one slow test into three wasted cycles.

Two candidate fixes (the implementing session picks; both are in-scope)

1. Roster registration — add tests/test_workflow_lint_upload_or_true.py to the slow-tests roster with a MEASURED surcharge (the >533 s figure above is a floor, not a basis; measure at the production entrypoint). Cheapest, preserves the test's current live-tree coverage semantics.
2. Refactor to the fixture-tree form — monkeypatch wl._REPO_ROOT onto a tmp fixture tree exactly as the sibling exemplar tests/test_workflow_lint_scripts_import_guard.py::test_check_bundled_in_no_flags already does. Sub-second, and it retains the same mutation-visibility property (deleting the or no_flags branch still fails the test). Strictly better if the live-tree read is not load-bearing — verify that before choosing it.

Prefer (2) if the live-tree execution is incidental; prefer (1) if it is deliberate coverage. Do not do both.

Acceptance

• An implementer's gate-matched local run over a diff selecting this file completes without a timeout kill, OR the selector defers the file with a measured surcharge and says so.
• If fix (2): the test still FAILS when the or no_flags bundling branch is deleted (mutation check), proving coverage was preserved rather than dropped.
• If fix (1): select_step9c_tests.py --json reports the file in slow_tests_selected with a non-default recommended_timeout_s, and a pin test asserts the roster membership so a future roster edit cannot silently drop it.
• No other selector behavior changes; the existing roster pin tests stay green.

Related

• scripts/select_step9c_tests.py (the slow-tests roster — the target surface)
• tests/test_workflow_lint_scripts_import_guard.py::test_check_bundled_in_no_flags (the sub-second exemplar to copy)
• #2253 (driving incident; its epm:results v1 carries the full evidence and the NOT-RUN single-test record)

Provenance

• workflow_fix_target: scripts/select_step9c_tests.py
• candidate-fingerprint: slow-roster-missing-upload-or-true-live-bundle-test

Surfaced as a formal <!-- workflow-fix-candidate v1 --> block by the implementer at #2253 implementation round 1, and routed by the #2253 orchestrator per the workflow-fix-on-bug protocol. The orchestrator verified the two claims that matter: the file is genuinely absent from the roster, and neither this test nor the file is present in #2253's own 134-test Step 9c selection (base_identical_excluded), so #2253's own gate is unaffected — this is filed for the fleet, not to unblock #2253.