feat: language-orthography plugin — enforce diacritical marks for non-English languages

[alissonlinneker]

Summary

Adds a new language-orthography plugin that enforces full orthographic correctness (accents, cedillas, umlauts, etc.) when the language setting targets a non-ASCII language.

• SessionStart hook reads the user's language from settings and injects an orthographic enforcement instruction
• No-op for English or when no language is configured
• Follows the same pattern as explanatory-output-style and learning-output-style plugins

Closes #32886

Context

The built-in language instruction template says "Always respond in pt-BR" but never mentions diacritical marks. The model interprets this loosely and frequently produces accent-less text — e.g., informacao instead of informação, voce instead of você. This affects every language with diacritics: Portuguese, French, Vietnamese, Czech, Turkish, Spanish, German, etc.

The problem gets worse after context compaction because:

1. The compaction/summarization step doesn't receive language rules, so the summary can lose proper diacritics
2. CLAUDE.md instructions that reinforce accents are wrapped in a "may or may not be relevant" disclaimer, which the model uses to deprioritize them

This plugin works around the issue at the prompt level. The long-term fix would involve strengthening the core r1z() language template and passing language rules to the compaction step — details in #32886.

What the plugin does

The SessionStart hook:

1. Reads language from ~/.claude/settings.json or settings.local.json
2. Skips if no language is set or if the language is English
3. Injects an instruction that frames diacritic omission as an orthographic error (equivalent to a typo in English), not a style preference

Test plan

• Tested with language: "pt-BR" — outputs correct enforcement JSON with accented examples
• Tested with language: "fr" — outputs enforcement for French
• Tested with language: "en-US" — silent no-op (exit 0, no output)
• Tested with no settings file — silent no-op
• JSON output validated — proper structure, accented characters preserved