fix(segment): a table cell is not a sentence — split it, like a paragraph

[groeimetai]

What changes

CORE.LEN.01 measures a sentence. The segmenter did not honour that inside tables: every
table cell became one unit, however much prose it held, while a paragraph has always been
collected first and only then split into sentences. A cell is now treated like a paragraph.

No rule file changed. This is a segmentation fix in tools/klaar/src/segment.js.

Why

The old behaviour did not just over-count — it made the rule unfixable.

A field-reference cell carrying five sentences was reported as one sentence of 85 words. The
author follows the rule's own fix text ("split the sentence at the conjunction"), splits neatly
at the full stop, and the number does not move, because the cell is still one unit. The only
remaining options are cutting past the point of meaning, or raising the baseline.

Both were taken in anger. In the Studio Groei platform the KLAAR baseline was raised six times
in six days (304 → 312 → 362 → 366 → 379 → 387 warns), and the waarom on the last bump says
so itself: "TWEEDE ophoging in dezelfde sessie... dit is een gewoonte die je niet moet laten
sluipen." Meanwhile 7 of the last 8 red CI runs on that repo were this ratchet.

A gate whose fix text does not work is worse than no gate: it teaches people to edit the
baseline instead of the text.

Evidence

Measured on a real corpus — the 39 spec documents of the Studio Groei platform, not an invented
example.

	errors	warns	sentences	density (warns/1000)
cell = 1 unit	23	387	9978	38.79
cell split	23	368	10301	35.72

The net figure is the least interesting part. The composition matters:

• 31 warnings disappeared — cells that were only "too long" because they held several
  sentences.
• 12 warnings appeared — real sentences over 30 words that were previously invisible,
  because their cell had already been counted once as a single unit.

So the measurement gets sharper in both directions, not merely gentler. Errors stayed at 23:
no document moves from conforming to failing.

Real example of a newly visible one, from docs/architecture/data-layer.md:

"De koppeling beeld ⇔ analyse is tweezijdig geklemd, want een designAnalyse z…"

And a representative disappearing one, from the table that triggered this work — a cell holding
five sentences of telemetry documentation, reported as a single 85-word sentence.

What deliberately does not change

• Unit kind stays line. A cell is not a free-standing sentence in a paragraph, and the
  shape rules lean on that distinction.
• All sentences of one cell keep the same block id, so a rule declaring unit: 'block' still
  sees the whole cell — groupByLine() groups on block id, not on line.

Only the boundaries move. maskInline is length-preserving, so an offset into the masked cell is
the same offset into the raw cell, and therefore into the source column.

Version: MINOR (0.2.0 → 0.3.0)

Part 5 asks: can a document that conformed yesterday fail today? No — conformance hangs on
errors, and those do not move. But PATCH does not cover it either: PATCH is for "tool changes
that do not change the verdict", and the verdict does change. A gate that ratchets on warning
counts can shift once because of this. Hence MINOR rather than something quieter.

Checks

• node tools/klaar/bin/klaar.js check passes
• node tools/klaar/bin/klaar.js conformance passes — 248 passed, 0 failed, over 30 rules
• cd tools/klaar && node --test test/ passes — 41/41
• npm run lint:self exit 0, 0 errors

Mutation-tested, not assumed. With only the production file reverted and the new tests kept,
the suite drops from 41/41 to 37/41 — all four new tests fail. They cover the split itself,
the shared block id, the per-sentence column offset, and a cell containing inline code (which
fails the moment maskInline stops preserving length).

Clean-room declaration

• I have not copied rule text, vocabulary entries, examples, numbering or section structure
  from ASD-STE100 or from any other licence-restricted specification.
• Every lexicon term I added carries a source from the allowlist in CLEANROOM.md.
  (No lexicon terms were added.)
• Every example I added is my own, written about software delivery.
• I have not compared a KLAAR lexicon against another standard's vocabulary.