PDF Rotation Text Verdict

Tells you whether the text a PDF extractor returned for a page is character-reversed, and repairs it, from facts you already have -- the page rotation and its character list. No I/O: it never opens a PDF.

THE SILENT FAILURE. On a page carrying /Rotate 180 or /Rotate 270, pdfplumber returns every text cluster with its characters in EXACTLY REVERSED order, while the row and column structure around it is completely correct. Measured on 0.11.10:

  /Rotate   0 -> 'Gizmo' and '$3.50'      correct
  /Rotate  90 -> correct characters
  /Rotate 180 -> 'omziG' and '05.3$'      reversed
  /Rotate 270 -> reversed

THE CASE THAT MAKES THIS DANGEROUS. A reversed NUMBER is still a valid number. '2024' becomes '4202'. '40071' becomes '17004'. '1250' becomes '0521', which int() reads as 521. A quantity, a year, a product code or a total silently changes value and every parser downstream accepts it, because nothing about it is malformed. There is no exception and no warning.

THE STRUCTURE IS NOT THE PROBLEM, AND THIS SAYS SO. Cell placement is correct at every rotation including 90, confirmed by hand-deriving pdfminer.six's own CTM formulas and by pixel-un-rotating pdfplumber's own render. Only the per-character join direction is wrong, and only at 180 and 270. A product that claimed a geometry bug would be selling a fix for something that works.

AND THERE IS NO DOCUMENTED OPTION FOR IT. extract_table has no docstring at all. The one parameter that looks relevant, text_use_text_flow, makes it worse: every character space-separated at 180, newline-separated at 270, still silently wrong.

A SAFE PAGE ROTATION DOES NOT MEAN SAFE TEXT. A sideways stamp, a margin note or a rotated column header -- drawn with a local rotation while the page's own /Rotate stays 0 -- is reversed too. An adversarial reviewer found this package returning a confident NOT_AFFECTED for exactly that page. The character facts carry the signal even though the rotation does not, so a page whose characters are not uniformly upright now returns NOT_DETERMINABLE naming the mixed evidence instead of a pass.

THE REPAIR IS OFFERED AS A PURE FUNCTION and it fails closed: it refuses right-to-left script by default, naming the exact codepoint, because visual and logical order already differ there and blind reversal would corrupt correct text.

VERIFIED: 96 tests, measured by running the suite. Every mutation observed FAILING before restore, with one genuine survivor found and closed.

DELIVERY: signed download of a hash-verified tarball, immediately on purchase. Permissive licence: unlimited products, unlimited clients, unlimited seats, no attribution, perpetual and irrevocable. One restriction, do not republish the source as source.