Global-ready: content language follows the reader + English code comments

[howieyoung]

Makes the project global-facing. Stacked on #1 (base branch feat/local-archive) — merge #1 first, then this retargets to main.

1 · Content language follows the reader (with an override)

New src/locale.ts resolves the language for LLM-generated content (topic labels, article summaries, social context, the weekly reading digest, the cover concept):

1. explicit override — USER_CONFIG.contentLanguage, when set and not "auto";
2. dominant pages.lang, when enough pages carry a real language tag;
3. a zero-dependency script heuristic (kana → Japanese, hangul → Korean, other CJK → Chinese, Latin → English) over recent reading — because Chrome/extension language tags are frequently empty (in practice ~80% here);
4. English as the ultimate fallback.

The generator prompts (enrich, digest, cover) are now English templates that emit output in the resolved language. Image-generation prompts stay English regardless (image models expect English). userConfig gains contentLanguage: "auto" as the per-user override (a BCP-47 code or a plain name like "English" / "日本語").

No behavior change for a reader whose content is already one language — the resolver returns that language; new content just follows what they actually read.

2 · Code comments → English

All Chinese code comments across the repo are translated to concise English so outside contributors can read the codebase. Comments only — every string literal is byte-identical: user-facing UI labels in the rendered issue/email/showcase, console/operator messages, thrown errors, and prompt strings.

Out of scope (deferred): the fixed UI chrome (rendered section labels, nameplate, html lang) stays Chinese for now — a separate follow-up if full UI localization is wanted.

Verification

• typecheck, the 18 security-invariant tests, and the security grep-gates all pass.
• String-survival check confirms the Chinese UI/console strings are untouched by the comment sweep.
• Resolver verified end-to-end (returns the reader's language; digests regenerated in-language at the tuned length).

🤖 Generated with Claude Code