chore: update site

site/.gitignore

@@ -17,10 +17,11 @@ site/public/r/
 provisioning/.content-packs/
 
 # ── Generated: runtime + build outputs ───────────────────────────────────────
-# htmx-templates/ and htmx-assets/ are ASSEMBLED (`just templates` rm -rf's them).
-# They stay tracked until templates-overlay/ is the declared source for every
-# hand-authored template and `just templates-check` proves regeneration is a
-# no-op — ignoring a generated tree that still holds unique work loses it.
+# htmx-templates/ and htmx-assets/ are ASSEMBLED (`just templates` / `just htmx-assets`
+# rm -rf them). They stay tracked until site/_htmx/ is the declared source for every
+# hand-authored template and asset, and `just templates-check` + `just assets-check`
+# prove regeneration is a no-op — ignoring a generated tree that still holds unique
+# work loses it.
 cache/
 data/
 logs_*.out

site/justfile

@@ -152,12 +152,18 @@ deck-new source id title_es="" slug_es="" title_en="" slug_en="" source_en="":
 
 # Gate (on disk): the deck's CSS bundle cannot reach out of its wrapper, the site's CSS cannot
 # reach in — except the glossary, which is entitled to and which is the whole point.
+#
+# The ids come from lib/variants.mjs, the one authority on the deck→artifact key. This recipe used
+# to re-derive them with `jq -r '.decks[].id'`, from before a bilingual deck became one source (and
+# one asset dir) PER LANGUAGE: it asked for `ai-infra`, which holds thumbnails and no bundle, while
+# the pages load `ai-infra-es`. It crashed on the deck built after the split and passed on the one
+# whose pre-split bundle still sat on disk — gating a file no page loads. Ask the authority.
 [no-cd]
 decks-check:
     #!/usr/bin/env bash
     set -e
     cd "{{ project_root }}/../presentation"
-    for id in $(nickel export --format json ../scripts/decks/decks.ncl | jq -r '.decks[].id'); do
+    for id in $(node ../scripts/decks/lib/variants.mjs --ids); do
         node ../scripts/decks/gates/artifact.mjs "$id"
     done
 

site/site/public/docs/case-files/protocol.html

@@ -132,7 +132,7 @@
   <div class='kicker'>Session protocol · Case-file series</div>
   <h1>The session <span class='red'>protocol</span></h1>
   <p class='logline'>This protocol was not written: <b>it is projected</b> from the case files' replays — the session replayed with protocol that every case leaves written and typed. Change a case and this page changes with it. Editing it by hand is impossible by construction, because a hand-kept checklist is the next drift.</p>
-  <div class='meta'><span class='tag'>Replays&nbsp;<b>4</b></span><span class='tag'>Cases&nbsp;<b>11</b></span></div>
+  <div class='meta'><span class='tag'>Replays&nbsp;<b>4</b></span><span class='tag'>Cases&nbsp;<b>12</b></span></div>
 </header>
 
 <section>
@@ -262,6 +262,7 @@ the one you deduce by reading the file.</pre>
 <tr><td><a href='/case-files/the-relapse-hardcoded'>404-PAP-bis</a></td><td>⊘ The relapse had a cause, and it was not memory: <b>the first fix had never gone back to the template</b>, so it was not a relapse but reinfection from the source. It has gone back now — the backport is in both templates and the hand-written list is forbidden by a typed constraint. What remains is the usual: <b>nothing runs it</b>. Rustelo's justfile calls no validator, so the gate that would prevent the next reinfection exists and is silent. A <span class="mono">just</span> recipe in its chain would settle it. Case 8/0 tells why the fix took months to come home.</td></tr>
 <tr><td><a href='/case-files/the-generator-that-reported-success'>1/1852</a></td><td>⊘ The frontier stays open where data becomes text: whatever is written past it is published on someone's word. A contract over the EMITTED artifact — a schema for the generated resource — would settle it, not one over its source alone.</td></tr>
 <tr><td><a href='/case-files/the-cure-that-never-went-home'>8/0</a></td><td>⊘ Nothing runs the new constraint: the recipe that runs all 23 and reports per constraint is unwritten — and without it, this rots again exactly the same way. A <span class="mono">just</span> recipe in rustelo's chain would settle it, kept out of <span class="mono">check-strict</span> until the 11 real hits are triaged. And four remain ownerless: the third list in <span class="mono">htmx_pages.rs</span> (3 hits), the 11 real hits of <span class="mono">no-hardcoded-route-paths</span>, <b><span class="mono">templates/</span> unversioned</b> — the gravest, and no ADR says whether it is policy or oversight — and an instance ↔ template drift-check, which has neither mechanism nor owner.</td></tr>
+<tr><td><a href='/case-files/the-discipline-without-a-trigger'>8/4</a></td><td>⊘ «--artifacts» declared list&lt;string&gt; and unreachable from the CLI wrapper — the five historical runs carry an empty column because of it; settled by accepting a comma-separated string or a repeatable option. «verify» in zero of the mode's eleven steps; adding it to the three with an oracle is approved (assemble_case: the .ncl exports; render: the .md reproduces; publish: expedientes-check). The howto's step list is not checked against the DAG; settled by a gate comparing its ids against mode show --fmt json. And the underlying one, with no gate possible: a sentence is not an artifact — the only oracle of the human↔agent channel is the human.</td></tr>
   </tbody></table>
 </section>
 

site/site/public/docs/expedientes/protocolo.html

@@ -132,7 +132,7 @@
   <div class='kicker'>Protocolo de sesión · Serie Expedientes</div>
   <h1>El protocolo de <span class='red'>sesión</span></h1>
   <p class='logline'>Este protocolo no se redactó: <b>se proyecta</b> de los replays de los expedientes — la sesión repetida con protocolo que cada caso deja escrita y tipada. Cambia un caso y esta página cambia con él. Editarla a mano es imposible por construcción, porque una checklist mantenida a mano es la siguiente deriva.</p>
-  <div class='meta'><span class='tag'>Replays&nbsp;<b>4</b></span><span class='tag'>Casos&nbsp;<b>11</b></span></div>
+  <div class='meta'><span class='tag'>Replays&nbsp;<b>4</b></span><span class='tag'>Casos&nbsp;<b>12</b></span></div>
 </header>
 
 <section>
@@ -262,6 +262,7 @@ deduce leyendo el fichero.</pre>
 <tr><td><a href='/expedientes/la-recaida-hardcoded'>404-PAP-bis</a></td><td>⊘ La recaída tenía causa, y no era la memoria: <b>el primer arreglo nunca había vuelto a la plantilla</b>, así que no era recaída sino reinfección desde la fuente. Ya volvió — el backport está en las dos plantillas y la lista a mano está prohibida por una restricción tipada. Lo que queda es lo de siempre: <b>nada la ejecuta</b>. El justfile de rustelo no llama a ningún validador, así que la puerta que impediría la próxima reinfección existe y está callada. La saldaría una receta <span class="mono">just</span> en su cadena. El expediente 8/0 cuenta por qué el arreglo tardó meses en volver a casa.</td></tr>
 <tr><td><a href='/expedientes/el-generador-que-informo-exito'>1/1852</a></td><td>⊘ La frontera sigue abierta donde el dato se vuelve texto: lo que se escribe tras ella se publica bajo palabra. Lo saldaría un contrato sobre el artefacto EMITIDO —un esquema del recurso generado—, no sólo sobre su fuente.</td></tr>
 <tr><td><a href='/expedientes/la-cura-que-no-volvio-a-casa'>8/0</a></td><td>⊘ Nada ejecuta la restricción nueva: la receta que corra las 23 y reporte por restricción está sin escribir — y sin ella, esto se vuelve a pudrir exactamente igual. La saldaría una receta <span class="mono">just</span> en la cadena de rustelo, fuera de <span class="mono">check-strict</span> hasta que los 11 hits reales estén triados. Y quedan cuatro sin dueño: la tercera lista de <span class="mono">htmx_pages.rs</span> (3 hits), los 11 hits reales de <span class="mono">no-hardcoded-route-paths</span>, <b><span class="mono">templates/</span> sin versionar</b> —la más grave, y ningún ADR dice si es política o descuido—, y un drift-check instancia ↔ plantillas, que no tiene ni mecanismo ni dueño.</td></tr>
+<tr><td><a href='/expedientes/la-disciplina-sin-disparador'>8/4</a></td><td>⊘ «--artifacts» declarado list&lt;string&gt; e inalcanzable desde el envoltorio del CLI — los cinco runs históricos llevan la columna vacía por eso; lo salda aceptar cadena separada por comas o una opción repetible. «verify» en cero de los once pasos del modo; aprobado añadirlo a los tres con oráculo (assemble_case: el .ncl exporta; render: el .md reproduce; publish: expedientes-check). La lista de pasos del howto no se coteja contra el DAG; lo salda una puerta que compare sus ids contra mode show --fmt json. Y la de fondo, sin puerta posible: una frase no es un artefacto — el único oráculo del canal humano↔agente es el humano.</td></tr>
   </tbody></table>
 </section>
 

site/site/rbac.ncl

@@ -257,7 +257,18 @@ let E = {
       # Per-post engagement (views/ratings/messages) — public, anonymous by design.
       { resource = "endpoint", pattern = "/api/htmx/post/*",          outcome = "allow", methods = ["GET", "POST"], effects = [] },
       # Contact form — public submission, no auth required.
+      # Superseded by the declarative endpoint below; kept while the old handler still exists.
       { resource = "endpoint", pattern = "/api/htmx/contact",         outcome = "allow", methods = ["POST"], effects = [] },
+      # Declarative forms — one endpoint, one policy. A form is DECLARED in
+      # site/config/forms/<id>.ncl and served by POST /api/htmx/form/<id>: no Rust, no rebuild.
+      # The handler refuses an id that is not [a-z0-9-] and refuses a posted field the spec does
+      # not declare, so opening the pattern does not open a hole — the spec is the boundary.
+      { resource = "endpoint", pattern = "/api/htmx/form/*",          outcome = "allow", methods = ["POST"], effects = [] },
+      # Newsletter subscription — public, anonymous by design. The handler RELAYS to the provider in
+      # newsletter.ncl and answers with a fragment, so the reader stays on the site and learns
+      # whether it worked. The list is read from that config, never from the request: opening this
+      # pattern lets a visitor subscribe, not choose which list they land on.
+      { resource = "endpoint", pattern = "/api/htmx/newsletter",      outcome = "allow", methods = ["POST"], effects = [] },
       # Leptos server functions for OTP auth flow — hash suffix varies per build,
       # glob * matches the numeric hash without requiring a path separator.
       { resource = "endpoint", pattern = "/api/server_request_otp*",       outcome = "allow", methods = ["POST"],  effects = [] },