Say what the scope rule is, and stop promising enforcement

[aaddrick]

Reddit user CaptainFilipe called this out. Asking about my setup, they asked how the style would "guarantee" it never touches code while still touching prose inside code — docstrings, log messages, HTML content, LaTeX body text. The question was correct, and the README produced it.

Two problems in one section

It promised enforcement. "What it never touches", "character for character", "and nothing else" describe a lock. An output style is system prompt text. Nothing outside the model holds that boundary. The honest claim is a strong default, and the section now says so, with a line telling readers to check output when it matters.

It dropped a target, and dropped the one they asked about. The canonical Scope section in output-styles/attention-control.md names four:

1. Prose the agent writes itself — apply the style
2. Code, commands, file paths, identifiers, error messages — verbatim
3. Quoted text — verbatim
4. Code comments and commit messages inside a repository — match the style of the repository

The README listed the first three. So a reader asking what happens to a docstring found no answer, when the style file has one: defer to the surrounding code. All four are in the README now.

Scope of the change

Six files. README.md plus the five translations, which carried the same omission and the same promise. The two paragraphs are translated in each.

tests/test_readmes.py never covered this: no fences, no code spans, no numbered rules in that section. That is why the omission survived from the first commit.

Not in this PR

They also identified a real gap: prose the agent writes into a document file — LaTeX body text, HTML copy, a log string it authors — is a fifth target the style does not name. That belongs in output-styles/attention-control.md, which changes the style hash and every generated copy, so it gets its own PR and its own eval consideration.

Gates

sync_style.py --check     All 5 generated files match their sources.
check_configs.py          All 8 shipped config files parse.
run_evals.py validate     Evaluation cases are valid.
unittest discover         Ran 75 tests ... OK