Mace-legal teaches Mike to read the red ink

Tracked changes, deletions, and reviewer comments now reach the assistant instead of being scrubbed at the door.

Most document ingestion pipelines quietly flatten a Word file or PDF into clean final text - which means everything a reviewer marked up, struck through, or commented on disappears before the AI ever sees it. JonasBoury's change to the Mace-legal fork keeps that editorial layer intact. DOCX insertions, deletions, moved passages, and comment bubbles are passed through as inline markers, and PDFs get an optional path (via PyMuPDF, a library for parsing PDF internals) that picks up colored redline annotations.

The prompts on the other end were updated to match, so the assistant can actually reason about who proposed what, what got cut, and what a reviewer flagged - rather than treating a marked-up draft as if it were a clean one. Documents with no markup go through the existing fast path, so nothing slows down for the common case.