Regenerate groundtruth for docling-parse#349 (curly double quotes fold to "): 147 pages on 65fd81c7
Groundtruth for docling-project/docling-parse#349 (fold curly double quotes to ", not '), regenerated against dataset main 65fd81c7 so it carries the word/line bbox precision from 3826b7b7 and the AcroForm groundtruth from 65fd81c7. Replaces discussion #5, which was generated against 571f34b1 and would have reverted those bboxes.
Regenerated with pytest tests/test_regression_parse.py::test_reference_documents_from_filenames --update-groundtruth on docling-parse fix/curly-double-quotes rebased onto main at 7.22.0 (95d24e3). 147 pages in 76 documents change; every changed cell differs only by ' becoming " at the same position (checked character by character after decompressing the .json.gz files, not just by hash). 588 files are touched; every other file in groundtruth/parser/ is byte-identical to main.
Compared with discussion #5 this drops one page: 6480366468566741514.pdf page 1. On 571f34b1 that page's „Heute: KOHLHAAS“ came through as 'Heute: KOHLHAAS' and #349 turned it into "Heute: KOHLHAAS"; since #352 (null-cmap fallback) main decodes the same two glyphs as ãHeute: KOHLHAASÒ, so there is no quote left to fold. That looks like the fallback picking a Latin-1 table where the font wants Mac Roman (0xE3 and 0xD2 are „ and “ in Mac Roman); I will raise it on the docling-parse side separately.
Side note: groundtruth/parser/ on main also contains 753 *.cells.delta.txt / *.word.delta.txt files. These are the diagnostic files test_regression_parse.py writes on a mismatch and deletes on success, so they look like an accidental commit in 3826b7b7. This PR does not touch them.
Closing as superseded. The curly-quote groundtruth for docling-parse#349 was regenerated and landed directly when that PR merged on 2026-09-28 (dataset main is now 6d3458e2, and docling-parse main pins that revision), so this PR's artifacts are redundant. Nothing here needs review — closing to keep the queue accurate.