benchmark: add a small regression fixture for over-selection noise

[vishal2612200]

Goal

Add one small fixture or benchmark case that catches a noisy selection pattern without tuning to one repository name.

Suggested files

• tests/test_ranking_evals.py
• tests/fixtures/ or benchmark fixture helpers

Acceptance criteria

• Test fails before the intended ranking improvement.
• The scenario is general, not named after one benchmark repo.
• The test documents why the noisy file should not outrank the target file.