Los casos diagnosticaban y no exigían. Cada uno cerraba con un veredicto —«un contrato que no
se aplica no es un contrato: es un comentario»— y esa lección no llegaba a ninguna parte: los
nueve apuntaban `related_modes` al modo que ESCRIBE expedientes. Un bucle circular. La serie
denunciaba el contrato desconectado y ella misma lo era.
El esquema ahora lo impide:
lesson_codified_in pares what/where con gramática de punteros verificados
(adr:NNN/id · qa:id · mode:id/paso · gate:just X · test:ruta · contract:ruta#campo)
lesson_debt lo que nada exige todavía, con la dirección de quien lo saldaría
LessonAccounted contrato que RECHAZA el caso que no llena ninguno de los dos.
Visto rechazar antes de rellenar nada. Confesar cuesta (adr-071).
replay la sesión repetida con protocolo: el prompt literal recibido (con su
procedencia declarada — una reconstrucción sin fuente es ficción), el
prompt corregido, las microtareas con su oráculo NOMBRADO, la puerta
antes de delegar y el disparador de adr?.
Contabilizados los diez casos × 2 idiomas. Cuatro llevan replay. El reparto es el hallazgo:
las lecciones que dolieron en MECANISMO acabaron en constraints de ADR; las que dolieron en
CONDUCTA DEL OPERADOR quedaron en deuda. Esa mitad era justo la que faltaba.
Nuevo: /docs/expedientes/protocolo.html y /docs/case-files/protocol.html — el protocolo de
sesión, PROYECTADO de los replays (gen-protocolo.nu, con --check en la cadena). No es una
tarjeta del grid: el grid indexa CASOS, y una proyección archivada entre las instancias de las
que se proyecta es el mismo error de categoría que un esquema archivado entre ellas. Lleva las
cuatro caras de una sola regla —el agente nunca es su propio oráculo— y la deuda viva de la
serie, que se recalcula sola: cuando una ⊘ se salda en su caso, desaparece de la página.
Nuevo expediente 8/0 — la cura que no volvió a casa (ES+EN, 'clinical). Las ocho líneas que
cerraron el caso fundacional existen, probadas y sirviendo, y no habían llegado a ninguna
parte: ni al commit de su instancia (donde alguien siguió añadiendo kinds A MANO hasta siete),
ni a las dos plantillas de las que nace cada sitio. 23 constraints en 8 ADRs, 0 ejecutadas.
Y su punto de abandono: el arquitecto recomendó borrar el atajo con «prueba empírica» construida
sobre dos árboles distintos — habría roto 6 de 8 rutas respondiendo 200. Lo paró el testigo que
el propio encargo exigía. La puerta cazó al arquitecto.
Deuda corregida en 404-PAP y 404-PAP-bis: decían que faltaban tres cosas cuando faltaba una.
El backport está hecho y la constraint escrita; sólo falta que algo la ejecute. Un howto no
expira solo — éste caducó en dos horas.
Y el mensaje de la puerta de ilustración daba rutas que no resuelven desde donde la puerta
corre. Corregido: las rutas son de la raíz de la constelación, y el conversor NO reescala.
Verificado: 22 casos reproducen · completos en en, es · servidos · informe reproduce ·
parity 11×2 · protocolo-check reproduce · oráculo de cobertura ALL PASS.
The published deck was a hand-written carousel of exported PNGs: the words of every
slide were an IMAGE of words. Nothing indexed them, no glossary tooltip could reach
them, no screen reader read them, and the page drifted from the source the moment a
slide changed.
The deck is NOT re-parsed. Slidev's markdown is not standard markdown, and its meaning
does not live in the markdown anyway — it lives in a compiled bundle of UnoCSS
utilities, theme rules and <style scoped> blocks. Re-implementing that is
re-implementing a renderer, in any language. So Slidev renders (its own /print route,
the one `slidev export` drives) and we capture what it produced: the DOM, and the CSS
the browser actually applied.
Measured on the served page: 0 → 2152 indexable words, 16/16 slides reachable, 0
third-party origins added, 0 broken references.
Twelve traps, none of which raised an error:
· The CSS must be asked of the BROWSER (17 stylesheets; index.html links 2).
· Collapsing :root/html/body imports the SPA's viewport cage — it clipped the deck
at 900px and swallowed 14 of 16 slides, silently.
· UnoCSS's preflight, scoped, kills the glossary underline it was supposed to enable.
· :not() inherits its argument's specificity — the boundary guard grew with every
exemption until it reverted the slide scaling. :not(:where(…)) contributes zero.
· scale(calc(100cqw / 980)) is invalid CSS and is dropped in silence; on a phone the
box was 328px while the content stayed 980px and two thirds of every slide was cut.
· The site's markdown renderer destroys an inline <script>: curly quotes, injected
<p>, &&. CSS and JS go to external files — correctness, not optimisation.
· The 980x552 frame CUT 12 of 16 slides (Slidev cuts them too; nobody had seen it,
because the slides were PNGs of themselves).
· transform: translateY(0) — an IDENTITY transform — makes <main> the containing
block for position:fixed, and the zoom modal landed at top = -221px.
Two views over one DOM (slider default, stacked reads with Ctrl+F), an index derived
from each slide's own H1, a sticky rail, zoom, and Slidev's own copy buttons revived —
they shipped in the captured DOM at 0x0, unclickable by any human.
Twelve gates (A–L), and IN THE CHAIN: `just decks` in sync-full, `decks-check` in
`check` (on disk), `decks-served` in `check-live` (the page). A generator outside the
chain is a generator that does not run.
The article was a projection; the informe was not. Same cases, two hands, and
they drifted: deriva's ledger carried 6 rows in the .md and 8 in the viewer,
recaida had no typed source at all, and espejo never made it into the informe —
so its own article linked to an anchor that did not exist. The English tree was
two hand-written .md nobody regenerated and a third case that was simply absent.
Three fixes, one shape:
- The <pre> was killing the render. `<section class="exp">` opens a CommonMark
HTML block of type 6, closed by the first blank line — and weapon.code has
blank lines. Joined with "\n" the whole body was ONE block that died inside
the snippet, so every `#` comment re-parsed as an ATX heading: 23 spurious
<h1> in espejo, `</pre>` swallowed by the last one. Chunks are now separated
by blank lines, so <pre> opens a type-1 block that only closes on `</pre>`.
The editorial prose stops being buried in raw HTML and renders as Markdown.
- The skin becomes a dimension of the contract. The informe carries prose per
case AND per skin; the old Expediente held one flattened skin, so the viewer
could never be projected from it. `skins.<skin>` now holds what the genre owns
(loglines, alibis, headings, stamps); the numbers, the code, the technical
names of the suspects and the closing stay neutral. A datum that changed with
the skin would be fiction, not genre.
- The locale becomes a dimension of the build. Both trees project from .ncl,
the renderer's vocabulary lives in scripts/ (not among the instances it types
— `contract-in-the-instance-tree`), and expedientes-parity fails when a case
exists in one locale and not the other.
Gates: expedientes-check · informe-check · expedientes-parity, all in the chain.
A projection nobody re-derives is not a projection, it is a copy you trust.
The howto called the CASE object the single source of truth. It was not. The rendered .md
was hand-written — 356 of its 500 lines were the shared glossary widget pasted in verbatim,
and it carried editorial prose (the balance, the closing) that the contract never modelled.
A generator run against the old contract would have DELETED that prose: the exact
work-in-the-generated-tree shape adr-070 names.
So the contract was extended first (balance / balance_outro / closing, additive — every
existing case validates unchanged), the prose was rescued into it, and only then was
gen-expediente.nu written. The glossary widget is now injected from its single source
instead of retyped. `just expedientes-check` asserts every .md body reproduces from its .ncl:
edit the source, never the render.
The generator dropped the hero and the "show full expediente" CTA on its first run — half an
hour after the ADR that forbids exactly that. It was caught only because the output was
diffed against the previous render instead of trusted because it printed a green tick. Both
are emitted from data now, and the check is why we know.
New case, clinical skin: 69/58 — "the mirror that reverted its own work". The tool reported
"index.json with 69 posts" while the disk held 58; eleven ADRs (059..069) were written and
never published; zero content reached production for four weeks; 214 lines of template work
sat one command from deletion with no witness anywhere. Every number in its ledger traces to
a log or an artifact, per the mode's own rule — none is recalled.
Its verdict is the one the series exists to ask and rarely gets: "with ontoref, would it not
have gone like this?" — and here the answer is NO. It happened INSIDE ontoref, with ADR-066
("the check decides, never the reporter") accepted a week earlier and proven by an eight-path
falsación run. The protocol did not fail. Difusión was the blind spot: the one surface the
protocol did not govern. Every mechanism that would have caught every one of these failures
already existed and was already accepted in this very repo. None of them was pointed here.
Also: adr-070's pages (the ADR projection now runs in the build chain).
Two failures, one shape: something living in a tree whose consumers assume it holds
only something else.
adr-pages: gen-adr-pages.nu existed but no recipe called it, so the ADR projection
silently stopped tracking the spine — ADR-059..069 were written months ago and never
reached the site. A generator outside the dependency chain is a generator that does
not run. Wired into sync-full ahead of `content` (it emits the markdown that content
indexes), and it is idempotent: this commit carries no ADR diff because regenerating
reproduces the tree byte for byte.
graph: gen-content-graph nickel-exports every .ncl under site/content, and
expedientes/_schema/expediente.ncl was a CONTRACT, not an instance — export demanded
`id` and the recipe aborted. `just graph` had been dead since expedientes landed. The
generator was right and the file was in the wrong tree: site/content/ holds instances,
so a schema there is a category error. Moved to site/schemas/expediente.ncl; the one
sidecar that imports it still type-checks against the contract. graph now builds
(mini:154 related:222 ontology:15).
sync-full ordering: graph / adr-map / about-json ran AFTER content, but the leg that
copies the four nu artifacts into the serve tree lived INSIDE content — so a full sync
left site/r rendering the previous run's About and ADR map. Added sync-artifacts as the
final link. Both trees now agree on all four.
Gates
-----
adr-check spine vs site: fails if the spine holds an ADR the site never published.
Verified by deleting adr-069 — the gate caught it (69 vs 68, exit 1) and
sync-full rebuilt it unattended.
This is the third gate (with templates-check and posts-check) and, like the others, it
runs before the damage, not after. None of them run on deploy yet: `just authoring
publish` still ships without asking. That is what keeps these gates a habit rather
than a law — and is the subject of the governed-delivery ADR still to be written.
Refs: ADR-066 (the check decides, never the reporter).
Versión corta, si la prefieres:
site: wire the ADR projection into the build chain; move a contract out of the data tree
gen-adr-pages existed but no recipe called it — eleven ADRs never reached the site. Now in
sync-full, with an adr-check gate that fails when the spine holds an ADR the site doesn't.
`just graph` had been dead since expedientes landed: a schema sat inside site/content/, which
gen-content-graph nickel-exports as if everything there were an instance. Moved to
site/schemas/. Also fixed sync-full ordering so the serve tree stops rendering the previous
run's About and ADR map.
First real commit of site/. Until now outreach/.git tracked exactly one file
(README.md) and every content pack stamped git_sha=5619a40 regardless of what it
shipped — a signature with no provenance behind it. The site had no witness, which
is why a deploy could erase two sessions of work with nothing to diff against and
nothing to restore from.
Tracked here (authored): site/content, site/config, site/i18n, site/public/images,
justfile, scripts/, .just/, run*.sh, rustelo.manifest.toml.
Ignored (reproducible, or secret):
.k, .env real keys — outreach mirrors to a PUBLIC remote
site/r, site/public/r content indexes, rebuilt by `just content`
provisioning/.content-packs 228 MB of deploy tarballs
cache, data, logs_*.out runtime droppings
Three trees are deliberately NOT ignored despite looking generated, because their
declared source cannot currently reproduce them — ignoring them would delete the
only copy:
site/public/images 196 of 209 images exist nowhere else (assets/site is empty)
site/content/{adr,catalog} projections whose generators are unwired (gen-adr-pages)
site/content/domains or failing (`just graph`)
They become ignorable when an audit proves regeneration is a no-op, not before.
Fixes carried in this import
---------------------------
content: the mirror ran one way and reverted its own work. content_processor writes
site/r, but the recipe ended with `rsync -a site/public/r/ -> site/r/`, so each fresh
index was clobbered by the stale deploy-tree copy: the tool reported "index.json with
69 posts" while the disk kept 58. Meanwhile public/r is the tree the deploy pack ships,
so no new content could ever reach production. The two trees own different files, so
the mirror now runs both ways: content indexes r -> public/r (with --delete, or a
renamed slug survives as a live URL), and the four nu artifacts (about, adr-map,
taglines, content_graph) back public/r -> r, excluded from the outbound leg so stale
copies cannot overwrite the fresh originals.
templates: the assembler has two layers (framework defaults, source-project templates)
but the level chain has three (rustelo -> website-htmx-rustelo -> outreach/site). The
site level had no overlay slot, so its template work had nowhere legitimate to live and
ended up in htmx-templates/ — the one tree `just templates` rm -rf's. That is how the
nav submenus were lost. site/templates-overlay/ is the missing third slot; 214 lines
(nav submenus, subscribe CTA, post engagement block) now live in a declared source and
survive regeneration byte-identical.
adr: projected ADR-059..069 into the content tree. They were written months ago and
never reached the site because gen-adr-pages.nu is wired into no recipe — a generator
outside the dependency chain is a generator that does not run.
Gates added
-----------
templates-check reassembles into a temp tree and asserts an empty diff against the
live one: any drift means htmx-templates/ holds work that exists
live one: any drift means htmx-templates/ holds work that exists
nowhere else and the next run deletes it.
Both gates run before the destructive operation, not after. A check you can only run
afterwards is not a gate, it is an autopsy.
Refs: ADR-062 (projection-not-own-repo: site is a Projection of outreach, not its own
repo — different deploy target is not a different visibility boundary), ADR-048
(materialization), ADR-066 (the check decides, never the reporter — which the content
pipeline, the surface that publishes ontoref, was the last place not to honour).
