fix(eval): sort json2md GT when extra has no usable relation list - #4

Bug and impact

bigos.eval.omnidocbench._merge_truncated_json2md silently skipped the final sort-by-order step when a manifest row had extra set to a truthy dict but extra["relation"] was missing, None, or not a list. When triggered, prepare_annos_json2md returned blocks in manifest order rather than reading order, causing the json2md GT pipeline to emit misordered GT markdown.

Concrete blast radius for OmniDocBench evaluation runs:

• gt_markdown_json2md writes GT pieces in the wrong order for affected rows.
• _gt_markdown_for_eval (json2md strategy, the default) feeds that misordered GT into character-level metrics — cer, ned, cer_normalized, ned_normalized, len_ratio — biasing per-sample and aggregate numbers persisted in eval/results/baseline-*.json and the markdown report.
• _compute_per_category_metrics walks prepare_annos_json2md(row) and zips against pred blocks in reading order; misordered annos cause wrong-cell attribution of text_ned, table_teds, and formula_ned to the wrong category buckets.

Trigger: any OmniDocBench manifest row whose extra dict carries keys other than a list-typed relation. The reference tools/json2md.py always finishes with sorted(merged_annos, key=lambda x: x['order']) (documented in docs/eval/json2md-reference.md), so the previous behavior diverged from the documented contract whenever this branch was reached.

Root cause

if not extra:
    return sorted(annos, key=lambda x: int(x.get("order") or 0))
relations = extra.get("relation")
if not isinstance(relations, list):
    return annos  # <-- skips the final sort

The if not extra branch and the post-merge return both sort by order. The middle branch — taken when extra exists but doesn't contain a usable relation list — was the only exit that returned the filtered annos in their original manifest order.

Fix and validation

Sort by order in the no-relation branch as well, mirroring the reference tools/json2md.py end-of-pipeline sort.

Validation:

• Added tests/test_eval_omnidocbench.py::test_extra_without_relation_still_sorts_by_order covering three concrete triggers: extra with only unrelated keys, extra["relation"] is None, and extra["relation"] set to a dict. Confirmed the test fails on the previous code (assert 8 < 0 from 'second\n\nfirst') and passes on the fix.
• Full fast suite (pytest -m "not slow") green: 54 passed.
• Full eval suite including the slow Docling smoke test green: 13 passed.

No broad refactor; the change is one branch and one regression test.