kev-finetune: two rubrics on one state are not a label conflict

[jaredpalmer]

split_data.py drops a state when one question has two labels. It decided the question from the id and the instructions alone. The same ticket scored on a 2-level scale and again on a 3-level scale was a conflict, and both records were deleted. Reordered choice keys were already the same question. Criteria that actually differ were not.

This continues #240. The conflict key is the question id plus question_form: type, instructions and criteria, dumped with sort_keys so key order does not matter. The label histogram and the missing-label warning still added every record with that id into one counter, so index 1 on the short scale and index 1 on the long scale were one bucket, and the warning read the criteria off whichever record came first. label_groups counts each form on its own. The summary name is chosen once from every record the run writes, the training file and the holdout together, and every partition reuses it. A file with one form per id keeps the question id, in the order those ids first appear, which is the key order summary.json stored before this change. A second form is named rating (2 levels). When two forms would share that name, or the name is already some other question id, the next one is answer (2 options) #2, and that other id keeps its own counts.

The assumption is that a normal fine-tune has one type, one instructions string and one criteria object per question id. That file's report does not change. Two forms are two questions that share an id, and their level indices are counted apart. data-format.md says this. A model that will be served one question should still be trained on one definition. The script no longer deletes the second rubric to enforce that.

Test plan

• uv run --extra serve python -m pytest tests/test_skill_scripts.py -q: 52 passed
• Two score lengths on one state stay in the file, share a partition, and appear as rating (2 levels) and rating (3 levels) in summary.json
• The same two scales on different states keep those names in a partition that contains only one of them. A holdout on the other scale does too
• Reordered choice keys, including é / naïve, with different labels are still a conflict. Omitted noul criteria and "criteria": null are the same form
• Different instructions, a reversed score scale, and noul criteria text are not conflicts
• label_table on a single-definition record still keys department, escalate and frustration. A file whose ids appear as zebra then apple keeps that order in summary.json
• Three choice forms with the same option count keep the same #2 and #3 names when the file is reversed. A question id answer (2 options) is not overwritten
• A score with no criteria, and a choice whose criteria is a list, are invalid lines and do not enter the histogram
• warnings_for reports missing level 2 only on the 3-level scale, and names a noul that has criteria answer (noul) even when that criteria object has two keys