Give tied objects the same rank

[appleweiping]

Fixes #320.

Ranker assigned ranks by position in the sorted frame:

df[self.rank_attribute_name] = [idx + 1 for idx, _ in enumerate(df.index)]

Objects with identical scores therefore received different ranks, and which one came out ahead depended on the order pandas happened to produce. For integer-valued metrics like engagement counts or vote totals, where ties are common, the output reported a precision the data does not support.

Both transform and transform_objs now derive the rank from the score column with rank(ascending=False, method="min") — standard competition ranking. Tied objects share the best rank they collectively occupy, and the next distinct score resumes after the positions the tie consumed:

scores	before	after
10, 10, 10, 5	1, 2, 3, 4	1, 1, 1, 4
3, 2, 1	1, 2, 3	1, 2, 3

Ranks remain integers, and objects with distinct scores are unaffected.

Testing

Adds convokit/tests/ranker/test_ranker.py covering a tie through both transform and transform_objs, plus a distinct-score case pinning the unchanged behaviour. Verified against the released 4.1.2: the two tie tests fail (assert np.int64(1) == np.int64(2)) and pass with this change; the distinct-score test passes either way. convokit/tests/run_all_tests.py discovers all three.

The new test directory includes an __init__.py, matching the other test packages.

🤖 Generated with Claude Code

[laerdon]

Hi @appleweiping ! Thanks so much for looking into this! I'll review your PR within this coming week.

[appleweiping]

Focused validation is complete on the current head d5f816d.

• The ranker regression suite passed all 3 tests.
• The repository-wide runner reached a database mode corpus test that requires a live MongoDB. I stopped it after confirming that the block was environmental, not a ranker assertion.

The tied-rank behavior and its regression coverage are ready for review. The existing reviewer note from Sep 4 said a review would follow this week. Please take another look when available.