fix(theming): lo que el adversarial del cierre midio - numeros de prosa y el exit de --debt

La verificacion adversarial del cierre (11 confirmados, 2 refutaciones de
NARRACION, 7 mutaciones con arbol restaurado byte a byte) dejo cuatro
correcciones, ninguna de mecanismo:

- El "114" de palabras eran 117 nombres --palabras-* consumidos (103 con
  fallback + 14 sin, los 117 bajo --palabras-scheme-*): el punto doctrinal
  sale REFORZADO - todo el canal es proyeccion del scheme. Corregido en el
  README y en la cabecera del ledger.
- Los "18 sin contrato" de recipe-contract SS4 son 23: el titulo dejaba
  fuera a los cinco estructurales que su propia enumeracion lista dentro.
  Renumerado con el desglose recontado.
- --debt no fijaba exit code: un modo de chequeo que no puede fallar es
  medio guard. Ahora process.exitCode=1 con deuda nueva o STALE (verificado
  por mutacion: HEAD exit 0, con literal inyectado exit 1, restaurado).
- El 5o test del suelo se llamaba "mutation probe" y no muta nada: renombrado
  a lo que es (canario de cordura del scanner), con la historia de donde SI
  se vio el rojo por mutacion.

Hallazgos del adversarial que quedan registrados sin ejecutar: las fichas de
docs/audit/theming estan OBSOLETAS respecto a la clase structural (--report
sin re-correr; tarea aparte con supervision por ficha) - el desglose "Knobs
de apariencia" de la ficha no suma (preexistente) - el suelo de reach tiene
~0,5pp de holgura por redondeo.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/canon/recipe-contract.md

@@ -210,13 +210,16 @@ key. The old coarse ceilings of `theming-reach-floor.test.ts` (`maxLiteral`
 564 / `maxGlobal` 1156) are RETIRED by it: a total only sees a net move, so
 five regressions under five fixes read green.
 
-**R-5.2 is the contract half**, and its 18 contract-less recipes were
-adjudicated one by one rather than waived: eight paint nothing; the
-typographic primitives (`heading`, `text`, `display`) and the calendar-layer
-consumers are all-`system`, and their theming surface IS the layer they
-consume (D-TH.2-b); five are `structural`; the WIP tracks sit outside the
-audited catalogue; and the real debt (`field-langs`, 31 raw globals with no
-contract) is covered BY the ledger — the entry IS its written registry. What
+**R-5.2 is the contract half**, and its **23** contract-less recipes were
+adjudicated one by one rather than waived (recounted by the adversarial pass
+of the same day — an earlier draft said 18 by leaving the structural five out
+of its own total): eight paint nothing; the typographic primitives
+(`heading`, `text`, `display`) are all-`system` and their theming surface IS
+the layer they consume (D-TH.2-b); five are `structural`; the three
+calendar-layer consumers keep bridge privates and their raw knobs in the
+ledger; `palabras` is a WIP track outside the audited catalogue; and the real
+debt (`field-langs`, 31 raw globals with no contract) is covered BY the
+ledger — the entry IS its written registry. What
 the rule does NOT adjudicate, and says so in its own PASS note, is the
 `var(--{c}-x, fallback)` reference that no `base.ts` entry declares (`mockup`
 ×11, `text-scramble` ×2): `recipe-css-contract.test.ts` sanctions that idiom

scripts/theming-census-debt.ts

@@ -67,8 +67,9 @@
  * tokenizes, not by a commit of this axis.
  *   ⚠ ONE CAVEAT, MEASURED AND WRITTEN DOWN (2026-08-25, adversarial review
  *   of the debt-B axis): for `palabras` the census's classification is IN
- *   QUESTION. Its two recipes consume 114 distinct `--palabras-*` names with
- *   `var(…, fallback)` and declare ZERO of them; those names are not a theme
+ *   QUESTION. Its two recipes consume 117 distinct `--palabras-*` names (103
+ *   with a `var(…, fallback)`, 14 without — all 117 under `--palabras-scheme-*`,
+ *   recounted by the adversarial pass) and declare ZERO of them; they are not a theme
  *   contract but the live projection of the document's `scheme-style`,
  *   written per instance as `--palabras-scheme-*` by `schemeCssVars`
  *   (`palabras-scheme.ts`). That is a VALUE channel, another axis — so the

scripts/theming-census.ts

@@ -647,6 +647,9 @@ async function writeDebtBaseline() {
 
 function reportDebt(only?: string, asJson = false) {
 	const { debt } = censusAudit(only);
+	// A check mode that cannot fail is half a guard: red findings set the exit
+	// code so `--debt` can gate a script or a hook, not only a human reader.
+	if (debt.newDebt.length > 0 || debt.stale.length > 0) process.exitCode = 1;
 	if (asJson) {
 		process.stdout.write(
 			JSON.stringify({ newDebt: debt.newDebt, stale: debt.stale }, null, '\t') + '\n'

src/uix/eidos/components/palabras/README.md

@@ -29,9 +29,10 @@ Lo que sí está medido es el tamaño de la deuda. El censo del eje
 knobs** de apariencia con un alcance nominal del **22,2 %**: 103 leen nombres
 `--palabras-*`, 272 primitivos globales y 87 son literales. Pero ese 22 % **no
 es superficie de tema**: las dos recetas no declaran ni una sola de esas
-propiedades (114 nombres distintos consumidos con `var(…, fallback)`, cero
-declaraciones), porque son la **proyección en vivo del `scheme-style` del
-documento** — `palabras-scheme.ts` las escribe por instancia como
+propiedades (117 nombres distintos consumidos — 103 con `var(…, fallback)` y
+14 sin él, los 117 bajo `--palabras-scheme-*`; recontado por la pasada
+adversarial —, cero declaraciones), porque son la **proyección en vivo del
+`scheme-style` del documento** — `palabras-scheme.ts` las escribe por instancia como
 `--palabras-scheme-*`. Ese canal pertenece al esquema del documento, no al tema.
 
 La deuda, por tanto, no se salda aquí: se salda declarando el bloque `palabras`

src/uix/eidos/theming-reach-floor.test.ts

@@ -122,13 +122,15 @@ describe('theming reach — the floor only moves up', () => {
 	});
 
 	/**
-	 * The mutation test the plan demands (memory
-	 * `a-guard-that-inspects-nothing-passes`): a guard is only trustworthy once
-	 * you have SEEN it go red. This one proves the classifier still reacts —
-	 * without it, a broken scanner would report a perfect catalogue and every
-	 * assertion above would pass over nothing.
+	 * A SANITY CANARY on the scanner, not a mutation (memory
+	 * `a-guard-that-inspects-nothing-passes`): it proves the classifier still
+	 * returns sane values for a known-clean subject, so a broken scanner cannot
+	 * report a perfect catalogue and let every assertion above pass over
+	 * nothing. The real red was SEEN by mutation on 2026-08-25 (a literal
+	 * injected into `mark`, then into `avatar` by the adversarial pass) — a
+	 * test file cannot mutate the recipes it guards.
 	 */
-	it('the classifier still reacts to a literal (mutation probe)', () => {
+	it('the classifier still reacts to a literal (scanner sanity canary)', () => {
 		const scan = scanComponent('mark');
 		expect(scan, 'mark should scan').not.toBeNull();
 		const before = scan!.row.literal;