spec: seed the 4.3 cycle with one line per block

[dhh1128]

Finding #90: the ratified text wraps at eighty columns, so a line number names a sentence fragment rather than anything citable, and the wrap puts line breaks inside digests. Measured 2026-08-26 across spec/: 23 of 27 files carry a sha256 broken across a line boundary, 49 in all, each one a seed-header pin naming the ratified edition that seed repairs. Reading rule 2 says a digest whose preimage is not stated "is not a pin; it is decoration" (L993-1010), and one nobody can select or paste without repairing it by hand is close to that. The ratified document itself is clean at 11 whole digests, so the damage sits in the drafting corpus.

The seed proposes one line per block — paragraph, list item, table row, fenced block, heading. Roughly 370 blocks against 3,940 lines.

This is the piece of the carriage work that is provable rather than argued, which is the whole reason it stands alone. A reflow is correct when normalizing whitespace on both sides and diffing gives nothing, so a check can state that an editorial pass altered no text. tools/check-block-lines.py enforces the resulting shape and is version-gated at 4.3, so nothing older has to migrate.

It also survives re-rooting. The clean-root question #77 lists last constrains what the successor says; this constrains only how its lines are broken.

Split out of the original locator seed deliberately. That seed bundled this with locator tokens, with dropping markdown bullets, and with <dfn>, which are judgment where this is checkable.

Relates to #90, #75, #85, and #77.

[Nicholas-Keystate]

Substance-gate review at the branch bytes, with one live exhibit the seed did not have when filed.

The finding is now self-demonstrating on main. Running check-block-lines.py at today's tree: the covenant-seal carriage seed merged this afternoon (#81, spec/custos-4.3-seed-covenant-seal-carriage.md) fails it — 80-column wrapped, 163+ continuation lines, and its header pin splits the ratified sha256 across a line boundary at L5-6. Finding #90's defect, reproduced in a file that entered the tree today, hours before this review. No stronger evidence for version-gating the check is available.

Verified: the checker's semantics (block = paragraph / list item / table row / fenced block / heading; 4.3-gated so nothing legacy migrates); the reflow-provability claim (whitespace-normalized diff = empty is a mechanical test); the separation discipline (locator carriage held, <dfn> withdrawn pending a term test — correctly refusing to bundle judgment with the one provable change).

Merge carries one obligation: with the CI hook active, the #81 seed must reflow to one-line-per-block in the same motion or main goes red. The reflow is exactly the checkable operation this seed defines — whitespace-normalized diff empty, no text altered — so it will be applied as a follow-up commit with the check's own proof cited. The 606-citation cost the seed prices lands only on forward pointers into the successor, as stated; citations into ratified 4.2 are undisturbed.

One discipline note for the drafting corpus: future seeds should arrive one-line-per-block from birth; the checker makes that a gate rather than a request.