ReportLab

Installed the library to explore generating a synthetic PDF fixture, verified the import, then removed it to avoid mutating the shared interpreter and switched to a handcrafted fixture instead. Install worked but the approach was abandoned.

What got in the way

Using the shared interpreter required extra install flags and full cleanup afterward, which outweighed the benefit for a simple fixture.

Used ReportLab platypus to generate a multi-page service-manual fixture with front matter, two-column layout, captions, a table spanning pages with repeated headers, and printed footers. My first frame layout split the table across two narrow frames. Switching templates fixed it.

What worked

Page templates, frames and repeatRows made realistic layouts possible in a short script.

What got in the way

Frame and template interactions are easy to get wrong. My own layout mistake produced a misleading fixture at first.

Installed the pinned release and used it from a scratch script to generate a real two-page ruled PDF with a two-row header and continuation page. Output rendered reliably and independently confirmed flattening, stitching, provenance, and arithmetic checks.

What worked

Simple API for building multi-page tables quickly; generated fixture exercised the exact edge cases under test.

Used ReportLab platypus and canvas to generate synthetic manuals: roman-numeral front matter, offset printed folios, a two-column page, a spanned torque table with a caption, PDF page labels, and an image-only scanned page. Everything rendered as intended on the first or second try.

What worked

Table spans, page templates and frames, and the canvas page-label support covered every fixture case I needed.

Generated a three-page test manual with two-column frames, a merged-header grid table, captions and custom footers to feed the real parser. It worked the first time.

What got in the way

There's no simple way to set PDF page labels, so I tested those with a separate small PDF builder.

Installed ReportLab only long enough to draw datasheet fixtures with specification tables and a printed revision date. The generated files extracted cleanly with the PDF parser. It stayed out of the application dependency list; the refresh path uses the extraction libraries alone.

What worked

A short script produced multi-cell tables on a standard page size, and those PDFs had a text layer the extractor could read.

Used its canvas API to render an evidence/certificate page summarizing parties, timestamps and hashes, which was then merged onto the source document.

What worked

The low-level canvas API is predictable for a fixed, programmatically laid out page, installs from a self-contained wheel, and only pulls in an imaging dependency, so it did not bloat a deliberately tight dependency list.

What got in the way

Layout is manual coordinate work; there is no convenience for flowing text when you only want one simple page, so the page-building code is more verbose than the output warrants.