[translator] Epic: translation style controls — glossary, tone & brand voice

## Summary

Add first-class, structured control over **how** content is translated — a glossary, tone/voice, and protected (do-not-translate) terms — instead of relying only on a free-form system prompt.

## Motivation

Comparable tools (e.g. Sanity AI Translations) expose brand-voice controls as structured, maintainable settings rather than prompt text:

- a **glossary** of preferred/forbidden term translations per locale, with a do-not-translate flag;
- a **style guide** for formality and tone/voice;
- **protected phrases** the model must never translate (brand names, product names, set terms).

Today our plugin can only influence this through the system prompt, which is hard to maintain, not per-term, and not enforceable. This epic closes the most significant capability gap versus Sanity.

## Scope (high-level — to be refined)

- **Glossary**: preferred and forbidden translations per term, per target locale; a do-not-translate flag.
- **Tone / style**: formality and voice guidance applied consistently to translation requests.
- **Protected terms/phrases**: a hard constraint so listed terms are passed through untranslated.
- Make these **editable and maintainable** (config and/or editor-facing) and feed them into the translation prompt automatically.
- Consider hard (rule-based) vs soft (prompt-level) enforcement, and how quality is validated.

## Notes

- Data model, enforcement strategy and UX to be defined in follow-up sub-issues.
- Relates to the async system-prompt builder (#109) and the provider abstraction.

Priority: **High**. Details are intentionally high-level for now; sub-issues will follow.

Start by reading the async system-prompt builder in #109 and the provider abstraction referenced in the issue. The scope is intentionally high-level, so this epic should first be refined into sub-issues that define the data model, enforcement strategy, editor or configuration surface, and validation criteria before implementation begins.

Written by the indexing model from the issue text.