fix(openai-live): chunk Chinese and Japanese context appends at sentence endings - #6413

Fixes #6412

Summary

_chunk_text() (used by OpenAILiveLLMService._send_context_append() to split long delegation output into session.commentary.append / session.thinking.append events) only treated ASCII .!? followed by whitespace as a sentence boundary. Chinese and Japanese end sentences with 。！？ and no space, so a CJK result became one long "sentence" that _split_at_token_limit() cut wherever the budget ran out, i.e. mid-sentence:

• before: _chunk_text(cjk_text, 100) chunks ended with 测试句, 一个测, 这是一, か？这, …
• before: _chunk_text("First sentence. 第二句话。Third sentence.", 6) → ['First sentence.', '第二句话。Thir', 'd sentence.']
• after: every chunk ends at 。/！/？, and the mixed example gives ['First sentence.', '第二句话。', 'Third sentence.']

Changes:

• _SENTENCE_BOUNDARY also splits after full-width 。！？, with no whitespace required.
• When rejoining pieces into a chunk, no space is inserted after a CJK sentence ending (Latin sentences are rejoined with a single space exactly as before), so "".join(chunks) reproduces CJK text unchanged.

Latin-script behaviour is unchanged; the existing chunking tests pass as-is.

Tests

• Added test_chunk_text_breaks_cjk_at_full_width_sentence_endings and test_chunk_text_keeps_spaces_between_latin_sentences_next_to_cjk to tests/test_openai_live_service.py; both fail on main and pass with the fix.
• Changelog: changelog/6413.fixed.md.
• uv run pytest tests/test_openai_live_service.py → 50 passed (also verified with the change applied on current main); ruff check / ruff format --check clean.