fix: preserve conversation language in GenUI model feedback

[wang-kaopu]

Description

Refs #165.

GenUI model-facing feedback contained Chinese natural-language diagnostics and copy-ready Chinese example values. These inputs can influence the language selected for the model's next response, especially when validation or fence repair runs immediately before generation.

This change:

• preserves the complete GenUI system-prompt teaching content and component guidance
• replaces the prompt's user-visible title example with <user-language text>
• strengthens the rule that reply language comes from the conversation
• uses language-neutral metadata for the bundled skill catalog and SKILL.md frontmatter
• defines <user-language ...> as a required meta-placeholder that must never be emitted literally
• returns validator and render feedback through structured protocol fields
• sends fence repair guidance through [genui-fence-repair]
• replaces high-induction user-visible literals in SKILL.md examples with <user-language ...> placeholders
• keeps built-in templates and English/Chinese i18n resources unchanged

Validation

• pnpm run check — 776 passed, 99 skipped; typecheck and build passed
• node scripts/verify-pack.mjs — 2.05 MB tarball, 7.22 MB unpacked, 164 files

[wang-kaopu]

理论上 fewShot 减少了，需要验证 GenUI 生成质量是否会因此下降，如果会就要再单独优化 examples