fix: UTF-8 corruption in streaming — read chunks instead of byte-by-byte

[mitre88]

Problem

stream_llm() in agent.py:547 reads SSE responses byte-by-byte (resp.read(1)) and decodes each byte individually as UTF-8:

ch = resp.read(1)
buf += ch.decode("utf-8", errors="replace")

This corrupts all multi-byte UTF-8 characters:

• Emojis: 🍎 → `` (4 bytes decoded as 4 replacement chars)
• Accented: ñ, é, ü → `` (2 bytes each)
• CJK: 中文 → `` (3 bytes each)

Every streamed response with non-ASCII characters is garbled.

Fix

Read 4096-byte chunks instead of single bytes:

chunk = resp.read(4096)
buf += chunk.decode("utf-8", errors="replace")

Multi-byte characters are now correctly assembled before decoding.

Impact

• Before: Any emoji, accent, or non-ASCII char in streamed output is corrupted
• After: All UTF-8 characters stream correctly
• Performance: 4096-byte chunks are also faster (fewer syscalls)
• Zero breaking changes: Same SSE parsing logic, just reads bigger chunks

[mitre88]

@walter-grace One-line fix: stream_llm() reads byte-by-byte which corrupts all multi-byte UTF-8 (emojis, accents, CJK). This PR reads 4KB chunks instead. Same issue as PR #10.