fix: preserve split UTF-8 in JSON-RPC reader

[zhuwenzhuang]

Summary

• decode Codex app-server stdout with a stateful UTF-8 decoder
• preserve compatibility when the readable already emits strings
• add a regression test that delivers one Chinese character as three separate byte chunks

Root cause

Calling Buffer.toString() independently for every stdout chunk replaces an incomplete multi-byte UTF-8 sequence with U+FFFD. Since stream chunk boundaries are arbitrary, valid JSON text could therefore be corrupted before JSON parsing.

Validation

• npm test: 401 passed, 28 skipped
• npm run typecheck
• npm run build