Root cause of the Shira hallucination incident in production case 46 Friedman:
DocumentAnalyzer::extractFromDocx relied on PhpOffice\PhpWord, but on the
EspoCRM prod image the PHPWord files are present in vendor/ yet are NOT
registered in the composer PSR-4 autoload map — class_exists() silently
returned false, the method returned null, and Shira's read_document fell
back to OCR. OCR then only saw the signature image, the AI got "[signature]"
as document content and fabricated the entire CTS appeal as a knee injury.
This change rewrites extractFromDocx to use ZipArchive + a small regex
parser of word/document.xml. Independent of PHPWord, more robust on
tables/footnotes/hyperlinks (which PHPWord's element walker missed at depth
>1), and verified on prod against the same Friedman appeal: 80 paragraphs /
16633 clean chars extracted (vs 0 before).
The paired Python fix in shira-hermes (commit 7b517e1) makes the OCR
fallback also read document.xml, so even if this PHP fix regresses again,
the AI will not be fed "[signature]" as document content.
Refs Task Master #5
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
chaim
89b3f5844e
