docs: one fact spelled once, and the marker the reader can see - #454

Closes #443, closes #444.

bun run typecheck clean, bun run lint clean, bun run selftest 564 PASS / 0 FAIL / 0 SKIP / 0 HOLE, exit 0 (562 on main). The +2 is exactly CUR09 and GT06.

🔒 No summary figure was edited, which is #452's promise being kept rather than asserted: TY10 reports the identical 548 literal runs, 48 clause openings, 46 value-filled, 40 n('…') calls before and after.

#443 — measured before choosing, and the measurement changed the answer

The card assumed the sed copies might be dead weight. They are not: all three run and all three reproduce their refusals, and "degrees": 12 occurs eight times in gallery/nod/motion.json, so the one-liner and the recipe are the same replace-all rather than two edits that happen to agree.

So the argument had to be settled on other grounds — and there, the two fences turn out not to be the same object, which the card's framing does not distinguish and the tree does:

• The portrait pitch fence is a pure duplicate. One sed, one build, and the recipe two lines below states exactly that edit and that verb. Deleted whole. The prose above pointed into the fence (`/tmp/pp.json` above), so the two paragraphs around it were merged into one that names no path.
• The nod fence is half a duplicate. It brackets the fold angle from both sides: the 34° build is the recipe spelled twice, the 33° build is the green side. ⭐ The green side is not expressible as a recipe by construction — GT05 faults a recipe whose edit leaves the build green — so the grammar structurally cannot carry it, and that line is the only copy of that fact. The 34° line is deleted, the 33° line kept, and the lead-in reworded from "Two builds:" to "The green side of the bracket:".

Rejected, on a closed question rather than on taste

Generating the sed line from the recipe. A recipe carries from, to, example, spec and verb; the sed line also carries a redirect target, an --out directory and a # green / # A39 fires comment, none of which the recipe has or could derive — a generator would have to mint them, and a minted convention is the next hand-kept thing. Decisively, #442 already refused ending a recipe in a runnable command because "that command as written does not run … it would look copy-pasteable and not be." Re-adding a copy-pasteable rendering reopens a question this repository closed.

A checker over the surviving sh fence — the forbidden third state, which gates the copy instead of removing it.

The four refusal blocks reproduce exactly as before, proved rather than asserted

GT02 4 blocks / 29 wrapped lines before and after; GT01 found 25 / byTag 14 / byHead 6 / byRecipe 4 / verified 18 / declared 7, identical; GT03 92 planted edits, 0 misses, identical; GT05 notDragged 14, 7 grammar refusals, 4 elision probes, identical. Only line numbers moved.

GT06, and the population that makes it a gate

GT06_THE_EDIT_A_REFUSAL_RECIPE_STATES_IS_SPELLED_ONCE reads every fenced block the transcript scan skips — the ones with an info string — against the recipes the gallery carries, and faults a line spelling both halves of one edit.

Its negative population is real rather than synthetic: 10 lines name a recipe's from and patch it to something else (the surviving 33° line, the amplitude and wavelength sweeps) and are all left alone. Its red-first probe is generated from the grammar, so it cannot drift away from it.

⚠️ gallery/README.md is deliberately outside that scan and the reason was measured, not assumed. The index shows one filled-in recipe inside a ```markdown fence, so it carries both halves of an edit on one line — and it is not a copy, because GT05 already requires that example to parse and to be a recipe some README really carries. Widening the scan would have faulted the one copy in the tree that is already derived. The reason is written into the case.

#444 — an HTML comment explaining an HTML comment

docs/SPEC_COVERAGE.md opened with eight lines of comment above the marker comment. Both are gone; one visible sentence stands there:

A dated record: what this page states was measured on the date it carries and is not kept current, so the selftest's currency gate leaves its figures alone.

• One string. The sentence assembles from one constant, the pattern is generated from its lead, and the header comment names the constant rather than quoting the syntax — which is what docs(gallery): the marker that tells the gate how to check a block tells the reader too #442 did to its own header.
• Exactly one form parses. CUR09 proves the retired comment declares nothing, and that the sentence wrapped back into a comment declares nothing either.
• A near miss faults by name. A header line opening with the lead and then not the sentence is reported, rather than read as "this file never asked to be exempt" — the outcome that would be worse than no marker at all.
• The date is deliberately not a field. The page carries 2026-08-22 in its own prose; interpolating it would be a second copy of exactly the kind this scan refuses. CUR01 still requires an ISO date in the header and the marker points at it.

The hand-written decoder #444 said to look for was there — the eight-line comment existed only because the marker was unreadable. Deleted.

🔒 CUR01's population did not move, figure for figure: 22 docs read (14 shipped, 8 one hop off README), 1 dated record, and every "sites read" limb unchanged. No floor was lowered — two clauses were added, and the second is a floor the marker never had: before this, a marker form that stopped matching would have taken the page silently back into the scan.

Mutants — planted in the real file, run, restored, checksummed

#444 red-first ①, the retired comment form back where the sentence is — 561 PASS / 3 FAIL / exit 1:

CUR01: not one document carries the dated-record marker and one is supposed to: either the
sentence moved in the tree and not here, or the escape hatch has stopped being taken and a
page whose figures are known to be stale is now being read as current
CUR02: 1 stale tally: docs/SPEC_COVERAGE.md:89 "31 assertions" states 31, and the assertion
registry holds 41
CUR09: no document in the tree carries the marker, so every clause above was checked against
synthetic text only

⭐ CUR02 is the proof the escape hatch was doing work: with the marker gone, the page's own stale figure is refused against the registry.

#444 red-first ②, a near miss — the message names both strings and says why a near miss is worse than a file that never asked:

CUR01: a header line opens with the dated-record marker's lead and is then not the marker:
docs/SPEC_COVERAGE.md reads "**A dated record:** measured a while ago and not since." … so a
near miss is an escape hatch an author meant to take and did not — which is worse than a file
that never asked for one

#443 red-first, the deleted sed line put back verbatim — 563 PASS / 1 FAIL / exit 1, and the count is the point: GT01–GT05 stay green, so the refusal block still reproduces with the copy beside it and only the new gate sees it.

GT06: gallery/nod/README.md:866 spells a recipe's whole edit a second time, in a `sh` fence
this scan skips — the recipe is "replace `"kind": "yaw", "radius"` with `"kind": "pitch",
"radius"` in `gallery/portrait/motion.json`, then run `build`."

Negative controls: a command naming only a recipe's from; one naming neither half; the ten real gallery lines; the marker sentence itself; a marked page with an ISO date. None caught.

🚨 A check that printed the pass sentence while failing

CUR09's first version went red with the marker reverted and its detail read "…and all 0 marked document(s) carry the sentence verbatim" — a FAIL whose message says everything is fine. The floor was a bare conjunct in the verdict rather than an entry in probes. Fixed here and re-run, and recorded because it is the doctrine's first bullet exactly: the message is the product.

Reported, not fixed

• The marked set has a floor but no cap. Nothing refuses a second document taking the escape hatch; it would appear in a PASS detail and nothing would go red. The deleted comment's "Do NOT copy it…" was the only thing that ever said so, and it was a comment in one file that nobody read. How many documents may be exempt is a decision, not a measurement, so a note stands beside CUR01 instead of an invented number.
• The nod 34° refusal is a two-line refusal quoted one line deep. The build prints A39 twice — same triangles and areas, two keys — and the README quotes only the first. GT02 permits this, so it is not a gate defect; it is a block showing half of what a reader would see. Left alone, because changing it is a content edit and two refusal recipes are also spelled as sed one-liners in sh fences, and that copy is ungated #443 forbids one.

The next drift this change makes

Asked before finishing, and three hand-kept things came back out: a comment stating "exactly one document needs it" (a count the run already prints) became "one document takes it today" with the cap gap named beside it; GT06's comment stopped naming 33° and states what is structural instead; CUR09's detail interpolates the header-line count rather than spelling it.

What could not be removed: the marker sits three paragraphs above the page's own 🔼 "this note is dated" blockquote, so a reader meets the idea twice. They are not the same claim, and making the blockquote itself the marker fails on mechanics — it is > -prefixed and wrapped over three source lines, and anything looser is a second accepted spelling.