fix(csa): preserve diverse seed eligibility

[isty2e]

Fixes #106.

pick_diverse_low_value_seed() assumes at least one observed distance score is greater than or equal to their arithmetic mean. Float accumulation can violate that assumption by one ULP, leaving the eligible set empty even when every distance is finite and non-negative.

The normal selection path is unchanged. If rounding empties the eligible set, selection falls back to all candidates at the maximum observed distance and applies the existing objective-value ordering. The regression test uses six repeated 1 / 58 distances, which reproduce the failure on main.

Verification:

• uv run --extra test pytest -q tests/csa/test_csa_selection.py
• Ruff check and format
• basedpyright on the changed source and test
• pre-commit on both files