Correct data, in the most hostile encoding a real system ever sees

extreme=true on every generator that already took edge. Untrimmed and non-breaking whitespace, zero-width characters with a leading BOM, homoglyphs (Cyrillic and Greek lookalikes that read as Latin), RTL overrides and stacked combining marks — one class per record, rotating across a batch.

The question it exists to answer: does your pipeline trim, normalise and compare safely, or does a zero-width space break an equality check and a homoglyph slip past a filter?

The constraint that made it usable: identifiers stay ASCII-clean. Your PESEL still checksums after normalisation. Only the human-facing string gets mangled, and it stays recoverable.