fix(core): preserve unicode in grep previews

[kitlangton]

What

Prevent oversized Ripgrep line previews from persisting half of a Unicode character. A malformed preview could display as � and poison subsequent provider requests.

Before / After

Before

A grep match longer than 2,000 JavaScript string units could place an emoji across the preview boundary. slice(0, 2_000) retained only the emoji's first UTF-16 code unit, which JSON serialized as an unpaired \ud83d. The TUI displayed a replacement character, and OpenAI rejected later requests with HTTP 400 because the request body was not valid Unicode JSON.

After

When the preview cutoff lands after the first half of a Unicode character, Ripgrep drops that incomplete code unit before appending the truncation marker. Persisted tool output remains valid Unicode and can be submitted to providers.

How

• packages/core/src/ripgrep.ts: remove a trailing unmatched high-surrogate code unit from truncated line previews.
• packages/core/test/ripgrep.test.ts: cover an emoji crossing the exact 2,000-unit preview boundary.

Scope

This fixes the producer that created the malformed durable tool output. It does not add a provider-request sanitizer or rewrite existing session history.

Testing

• bun run test test/ripgrep.test.ts from packages/core: 3 passed
• bun typecheck from packages/core: passed
• Push hook workspace typecheck: 30 packages passed
• End-to-end A/B replay against OpenAI:
  • original exported history containing one unpaired \ud83d: HTTP 400 Invalid body: failed to parse JSON value
  • identical history with only that unpaired code unit removed: normal assistant response