fix(user-input-fold): preserve supplementary characters at prose cutoff

[ooiuuii]

Problem

Fixes #585.

The first-line prose cutoff in folded user messages can split an emoji or supplementary CJK character, producing a replacement character when terminal output is encoded as UTF-8.

Value

Keep the preview faithful to complete input characters while retaining the existing preview size and folding behavior.

Approach

Move the existing cutoff back one code unit only when it falls between a matching high/low surrogate pair. No new helper, configuration or rendering framework; all other folding paths remain unchanged.

Validation

• Actual finalized-user Markdown transformer: supplementary characters straddling the 1,200-unit boundary fail the UTF-8 round-trip assertion before the fix. Focused suite: 26 pass / 1 fail before, 27/27 pass after.
• Controls cover a complete pair fitting the budget, BMP text, existing fold marker and unchanged streaming input.
• node --experimental-strip-types --test tests/extensions/user-input-fold/index.test.ts, bun run check, git diff --check — pass.
• Scoped structured review (autoreview --mode local --engine codex --model gpt-6-astra --thinking xhigh --no-web-search): helper exit0, no actionable findings.
• Synthetic source-checkout fixtures, no private messages or provider calls. No installed-app E2E claim. Full local suite has separately reproduced host baseline timeout/Windows process-cleanup limitations; fresh CI is the broader gate.

Impact

• User-visible: no broken supplementary character at this folded prose boundary.
• Model-visible context and stored Session content: unchanged; original input remains intact.
• Runtime/lifecycle, config and compatibility: no changes. Existing code-unit budget, thresholds, streaming behavior and fenced-block handling preserved.