fix(ranking): prune zero-support rules in prune_harmful_rules

[AZERDSQ131]

Problem

Rules with validated_support == 0 never match anything on the dev split, but prune_harmful_rules only drops a rule when its marginal F1 is known and strictly negative. A zero-support rule's marginal F1 is exactly 0.0 (removing it changes nothing), which is not < min_marginal_f1, so it survives as dead weight.

Fix

Added a min_support: int = 1 parameter to prune_harmful_rules. Rules whose validated_support is below this threshold are now dropped too, independent of the existing marginal-F1 logic (unchanged). Default of 1 is conservative — a single dev match is enough to keep a rule.

Tests

• test_prunes_zero_support_rule: a rule that never matches the dataset gets dropped with the default min_support=1.
• test_min_support_zero_keeps_zero_support_rule: min_support=0 disables the new check, preserving the old behavior.
• Existing prune_harmful_rules tests still pass unchanged (their rules all have support ≥ 1).

pytest tests/ -q → 313 passed, 1 skipped. ruff check and mypy rulechef/ranking.py clean.

Closes #4

[adaamko]

Reviewed — this is right, and it matches what #4 asked for: zero-support rules have marginal F1 of exactly 0.0 so the old threshold never caught them. Verified the only library call site (engine.rank_rules) stamps validated_support via rank_rules before pruning, so the new default is safe there. Tests and typecheck are green. One thing blocks the merge: the rights checkbox in the PR body (the failing required check). Please tick the '- [x] I certify...' box in the PR description (same for your other open PRs) and this can go in.