generate: a Norwegian run on a local model, and what it took to read it

Every prompt and reply is now kept under <out>/llm/. An empty reply is
a retry, not a fatal error; a spec that never parses stops the run
before the structure phase; --visitor-llm lets a different model play
the visitor. The file-block parser finds markers anywhere in a line,
accepts a short END and lets a new FILE close an open one, after a
reply glued two files together with `<<<END>></think>`. The reference
gains a language section: copy in the intake's language, identifiers
in ASCII.

runs/rehab-entreprenor-no-2026-09-21: nemotron-3.5-lightning:30b-a3b
did not produce a loadable site. Written up with the reason and the
way forward: structure from needs.yaml deterministically, words from
the model.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH

generator/reference.md

@@ -158,6 +158,17 @@ title: <business name>
 lang: en
 ```
 
+## Language
+
+Everything a person reads is written in the intake's `language`: page
+names, headings, descriptions, field labels, button labels, the
+`who`/`wants` of each need, and the personas. Write it the way people
+in that trade and country actually speak, not as a translation.
+Everything a machine reads stays plain ASCII snake_case or kebab-case:
+need ids, persona ids, bucket names, state names, field `name`s, file
+names and paths. Letters such as æ, ø and å belong in the copy and
+never in an identifier. Set `lang` in site.yaml to the intake's code.
+
 ## Copy rules
 
 - Headings say what the visitor gets, in the visitor's words.