Protected-change approval: harden repeatability headline assertions

[tomatotomata]

Proposed protected change

Follow-up to #2289. The repeatability test currently uses substring assertions such as "STABLE over 2 runs" in headline; because STABLE over N runs is contained in UNSTABLE over N runs, a regression to the negative verdict can pass.

Trust impact

This is a tests-only correction to the benchmark-integrity checks. It makes the test reject a false stable result; it does not change benchmark scoring or production behavior.

Scope

• anchor stable and unstable headline assertions at the repeatability: ... prefix
• add one regression case proving an UNSTABLE headline cannot satisfy the stable assertion
• leave implementation and benchmark outputs unchanged

Validation

• focused repeatability tests: 39 passed
• Ruff on tests/test_repeatability.py: passed
• git diff --check: passed

Prepared with OpenAI Codex under human direction, with the change reviewed before submission. Please apply benchmark-change-approved if this protected tests-only change is acceptable, or let me know what scope should change.

[vanguarstew]

This is a focused, tests-only correction that strengthens the benchmark-integrity checks by anchoring stable/unstable assertions at the repeatability: prefix and adding a regression case. It fits the roadmap's emphasis on rigorous, anti-gaming benchmark scoring and does not touch production behavior. Please apply benchmark-change-approved and proceed with the change as scoped; no further clarification needed.