content/page/acerca.de.md

@@ -1,8 +0,0 @@
----
-title: über
-lang: de
-manual_translation: false
-translated_from: es
----
-
-**Viena Latina** ist die lateinamerikanische Gemeinschaft in Wien: Tourismusführer, kulturelle Agenda, Gastronomie, Gemeinschaftsleben und lateinamerikanische Geschäfte in der Stadt.

content/page/acerca.pt-br.md

@@ -1,8 +0,0 @@
----
-title: Sobre
-lang: pt-br
-manual_translation: false
-translated_from: es
----
-
-**Viena Latina** é a comunidade latino-americana em Viena: guias de turismo, agenda cultural, gastronomia, vida comunitária e lojas latinas na cidade.

content/page/contacto.de.md

@@ -1,8 +0,0 @@
----
-title: Kontakt
-lang: de
-manual_translation: false
-translated_from: es
----
-
-Schreiben Sie uns an **hola@vienalatina.com**.

content/page/contacto.pt-br.md

@@ -1,8 +0,0 @@
----
-title: Contacto
-lang: pt-br
-manual_translation: false
-translated_from: es
----
-
-Escreva-nos a **hola@vienalatina.com**.

content/post/mi-test.de.md

@@ -1,11 +0,0 @@
----
-title: Testartikel
-date: 2026-09-16
-lang: de
-manual_translation: false
-categories:
-- Comunidad
-translated_from: es
----
-
-Dies ist ein Test des automatischen Übersetzungsflusses im Grätzl. Viena Latina veröffentlicht in Spanisch, Deutsch und Portugiesisch.

content/post/mi-test.pt-br.md

@@ -1,11 +0,0 @@
----
-title: Artigo de prova
-date: 2026-09-16
-lang: pt-br
-manual_translation: false
-categories:
-- Comunidad
-translated_from: es
----
-
-Este é um teste do fluxo de tradução automática no Grätzl. Viena Latina publica em espanhol, alemão e português.

scripts/translation/markdown.py

@@ -5,11 +5,9 @@ model has no equivalent, so structure is protected here instead: non-prose
 blocks pass through untouched, and inline constructs are masked with opaque
 placeholders whose survival is verified after the round trip.
 
-Placeholders are word-shaped (``Zq0Xv``) rather than punctuation. Measured on
-M2M100 418M: the OpenNMT ``｟0｠`` convention was dropped on every single
-occurrence, because SentencePiece fragments punctuation runs and the model then
-fails to copy them. A token that looks like an unknown proper noun gets carried
-through, since the model has nothing to translate it to.
+Placeholders use OpenNMT's protected-sequence convention (U+FF5F/U+FF60).
+SentencePiece keeps these atomic; ``{{x}}``, ``<x>`` and ``%s`` get fragmented
+by BPE and dropped by the model.
 """
 
 from __future__ import annotations
@@ -18,8 +16,7 @@ import re
 
 from .provider import SITE_TO_MODEL, Provider
 
-_MASK = "Zq{}Xv"
-_MASK_RE = re.compile(r"Zq\s*(\d+)\s*Xv", re.IGNORECASE)
+OPEN, CLOSE = "｟", "｠"
 
 # Community vocabulary that must reach readers unchanged. Not inherited from
 # the WordPress plugin, which had no glossary at all — edit freely.
@@ -46,6 +43,9 @@ _INLINE = (
     re.compile(r"https?://\S+"),   # bare URLs
 )
 
+_PLACEHOLDER = re.compile(re.escape(OPEN) + r"\s*(\d+)\s*" + re.escape(CLOSE))
+
+
 class PlaceholderError(RuntimeError):
     """A masked span did not survive translation intact."""
 
@@ -57,7 +57,7 @@ class _Masker:
 
     def _take(self, match: re.Match) -> str:
         self.spans.append(match.group(0))
-        return _MASK.format(len(self.spans) - 1)
+        return f"{OPEN}{len(self.spans) - 1}{CLOSE}"
 
     def mask(self, text: str) -> str:
         for pattern in _INLINE:
@@ -68,10 +68,10 @@ class _Masker:
         return text
 
     def restore(self, text: str) -> str:
-        # Models pad, re-case and reorder placeholders; normalise before matching.
-        text = _MASK_RE.sub(lambda m: _MASK.format(m.group(1)), text)
+        # Models pad and reorder placeholders; normalise spacing before matching.
+        text = _PLACEHOLDER.sub(lambda m: f"{OPEN}{m.group(1)}{CLOSE}", text)
         for index, span in enumerate(self.spans):
-            token = _MASK.format(index)
+            token = f"{OPEN}{index}{CLOSE}"
             seen = text.count(token)
             if seen != 1:
                 raise PlaceholderError(