Derive semantic Markdown changes

[hweihwang]

Part 1 of the version-comparison work for nextcloud/collectives#599: the semantic model that turns two Markdown revisions into a typed list of changes.

• Two documents in, typed changes out, built on ProseMirror ChangeSet
• Changes classified as text, formatting, attribute, structure or unknown, with conservative move detection
• An explicit complexity limit: you get a bounded result, never a silent partial list

First of six. Followed by #9048, #9049 and #9050 in this repo, then nextcloud/collectives#2699 and #2700.

Full story, screenshots and the test matrix live in the #599 implementation update: nextcloud/collectives#599 (comment)

[benjaminfrueh]

@hweihwang Tested: 3-column table (headers A/B/C, data x/x/x), delete column C. Header change is reported correctly as C, but the data row shows column A as changed instead of column C. I think it could be because all three "x" cells are identical.

[hweihwang]

Thanks for the concrete reproduction. This is exactly the case T01 now pins.

At the commit you reviewed, the three identical body cells (x/x/x) gave the column alignment no evidence to tell which column was deleted, so the body rows could be attributed to the wrong column even though the header resolved correctly.

The fix landed as regression test T01 (attributes the duplicate-body column deletion) at the current tip: deleting column C from a 3-column header with two x/x/x body rows must emit a single table-column delete whose descriptors point at column C for the header and both data rows.

Reproduced against the current tip (c599581): one table-column delete, descriptors at before paths [0,0,2], [0,1,2], [0,2,2]. Header and both rows agree on column C. T01/T02 also cover insertion and first/middle/last variants.