parse_otsl_output (used by the Granite Vision table-structure stage and the MinerU helper) accepts both closed cells (<fcel>x</fcel>) and unclosed ones (<fcel>x<fcel>y<nl>). When an unclosed cell came before a closed cell of the same tag, the closed-tag branch matched first. Its non-greedy .*? then ran to the first </fcel> and swallowed the tokens in between:

parse_otsl_output("<otsl><fcel>a<fcel>b<nl><fcel>c</fcel><fcel>d</fcel><nl></otsl>")
# before: seq ['fcel', 'fcel', 'nl'], 1 row, cells ['a<fcel>b<nl><fcel>c', 'd']
# after:  seq ['fcel', 'fcel', 'nl', 'fcel', 'fcel', 'nl'], 2x2, cells ['a', 'b', 'c', 'd']

With the fix, closed-tag text cannot contain another OTSL token, so the unclosed-cell branch handles those cells instead. Output that is fully closed or fully unclosed parses the same as before.

Checklist:

• Documentation has been updated, if necessary.
• Examples have been added, if necessary.
• Tests have been added, if necessary.

Review evidence:

• tests/test_otsl_utils.py::test_closed_cell_does_not_swallow_preceding_open_cells fails on main and passes with the fix. The second test checks that fully closed output is unchanged.
• I ran the existing parse_otsl_output tests in tests/test_table_structure_granite_vision.py (-k parse, 13 tests) against the fix with the model imports stubbed out, since I did not install the ML extras, and they all pass. I did not run the full suite.
• ruff format --check and ruff check pass on the changed files.

AI assistance: I used an AI coding assistant (Grok Bot) to help find and write this fix. I reviewed the change and ran the tests above myself.