test(docx-compare): characterize remaining LibreOffice note projection gaps

[stevenobiajulu]

Characterize remaining LibreOffice note projections after the reorder repair

Independent dynamic Claude Opus 5 review approved repair head a8c467c, but identified the following bounded follow-ups. Ref: #941, #973, #852.

Observed reader limitation (N1)

A synthetic document with an inline-edited footnote keeps correct in-process XML Accept/Reject projections, but LibreOffice Accept produces an empty note body. The reviewer reproduced the same result using a temporary emulation of base 31b1d8f (not an actual baseline checkout). An edited footnote combined with a moved endnote also loses note-body text in LibreOffice; the repaired head preserves paragraph projections while the emulated base does not.

The current fallback activates for copied footnotes, not merely copied endnotes. Do not claim that all mixed-kind note documents are repaired.

• Reproduce on an actual clean baseline and current main, with public synthetic inputs, both move modes and Accept/Reject.
• Assert anchor-to-note-body bindings separately from paragraph text.
• Characterize inline-only edits and edited footnotes alongside moved endnotes.
• Decide whether to extend the collision-safe separated-definition fallback across note kinds or separately repair inline-note reader handling. Preserve both projections and document history-granularity tradeoffs.

Lower-risk characterization (N3/N4 and coverage gap)

• Probe mixed edited/moved inputs with lexical note IDs such as 01 and +1. The low-level normalizer canonicalizes IDs, but the pipeline fallback compares raw attribute strings. This was a static concern, not a reproduced regression.
• Add a lone-stable-anchor mixed-note regression. That branch remains in original reconciliation, but the reviewer did not dynamically exercise it.
• Evaluate whether the ordinary-property subtree allowance should explicitly exclude sectPr and non-property children. The reviewer observed annotationRef under rPr and plain sectPr under pPr pass the copy guard; the former is already schema-invalid input. No new identity collision was demonstrated.

Word is UNVERIFIED. Open-cleanly evidence is not an Accept/Reject oracle. Use public/synthetic inputs only; no private documents or derivatives. Keep the terminal-empty-paragraph limitation #891/#973 separate. Any production fix needs regression tests, full pre-submit, required real corpus, schema/projection checks and independent dynamic review.