fix(eval): use unique image_rel as sample_id to stop silent dump_dir overwrites

[cursor]

Bug and impact

Running bigos eval --benchmark=omnidocbench --dump-dir=... silently dropped most per-sample diagnostic JSON dumps. With the tables subset (~hundreds of pages spread across many PDFs), users would see only one or two *.json files in dump_dir/ instead of one per processed sample. The aggregate report numbers were correct, but the per-sample diagnostic data — used to inspect failing pages, compare pred_preview vs gt_preview, etc. — was lost.

Root cause

In bigos.eval.omnidocbench.evaluate, sample_id was built from:

        sample_id = str(
            page_info.get("page_no") or row.get("page_id") or image_rel or f"row_{idx}",
        )

page_info.page_no is the page number within a source PDF (1, 2, 3, …) and is not unique across manifest rows: every multi-page document starts at page 1, so picking page_no first meant every "first page" sample collapsed to sample_id == "1". The dump path is computed as dump_dir / f"{_safe_dump_stem(sample_id)}.json", so all those samples wrote to dump_dir/1.json and silently overwrote each other.

I reproduced this with 5 manifest rows sharing page_no=1: the eval reported 5 samples, but only 1 dump file was written (the last one).

Fix and validation

• Reorder the sample_id priority so image_rel (e.g. paper_2401.05459_page_001.png) — which is unique per row and carries both document and page info — is preferred over page_no. page_id (which would also be unique if present) keeps top priority.
• Add a regression test (test_dump_dir_writes_one_file_per_sample) that processes 5 rows sharing page_no=1 and asserts dump_dir contains 5 distinct files and 5 distinct sample_ids.

Validation:

• New regression test passes; previously it would have written 1 file.
• Full non-slow test suite passes (43 passed).
• ruff check clean.
• Manual rerun confirms: 5 samples → 5 dump files (doc0_page1.png.json, …).

No public API change; affects only the value used for result.sample_id and the per-sample dump filename.