EqualifyEverything / equalify-iris

#374's last open item was a decision rather than a patch: amend `SplitWord.split`'s contract so a break inside one cell is in scope, or accept a floor of about 4 per 100 pages and write it down beside the other limits. The decision was to document it, so this is a comment and two tests — no runtime path moves, and nothing about the delivered HTML changes. The shape: `Compos-<br>ite` inside one `<th>` puts the hyphen and the break in a single cell. `splitWordContradictions` reads `WORD` over `textOf`, where the tag has become a space; the join's rewrite runs `WORD` over each text run between tags, and `Compos-` ends its run. No letter follows the hyphen either way, so the word is a candidate for neither pass — the same blindness the file already documents for a wrapped source, arriving as markup instead of whitespace. Census on the shipped arm: 24 breaks carried into column heads, `joinBrokenWords` closes 17, 3 of the 7 survivors land on limits already named, the last 4 are this shape on one page. Why the repair costs more than the defect: a wider pattern needs a distinction between an element that ends a CELL and one that ends a LINE, and `textOf` renders both as one space, so `<td>Total-</td><td>farm</td>` cannot be told from `Compos-<br>ite`. It also decides what `split` REPORTS, and `split` is contractually a string the document contains — nothing on the page reads `Compos-ite`. Passes downstream read that contract. `agents/page.md` now tells the page agent a word broken at the foot of a column is one word (#450), which lowers the rate of the input for free. Pinned rather than left as prose, in three tests that fail for different reasons: widening the pattern alone reddens the wrapped-source case, and widening it AND taking whitespace out of `splitWordAudit`'s lookup key — the whole repair — reddens the `-<br>` limit and the cell boundary as well, all three in one run. Separate tests on purpose: `assert` stops at the first failure, so while the cell-boundary pin shared a test it reported nothing and the widening's cost would have arrived one edit cycle after its gain. Two review rounds, both APPROVED. Round 1 carried two non-blocking notes, both correct and both fixed: the masked cell-boundary pin above, and "both read `WORD` over `textOf`", which was wrong about the join's width. Round 2 re-ran the mutation itself, got the three failures and the exact reported values the comment claims, and returned nothing new. Verified: `tsc --noEmit` clean, `npm test` 1700 / 1700. Closes #374 Co-Authored-By: bbertucc <bbertucc@users.noreply.github.com> Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>