fix: UTF-8 decoding corrupts multi-byte characters in streaming

[TheBlueHouse75]

Summary

Fixes garbled output for any non-ASCII text streamed from the LLM (French accents, Chinese, Japanese, emoji, etc.).

The bug

In stream_llm(), the SSE reader was doing:

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

Multi-byte UTF-8 characters span 2–4 bytes (é = 0xC3 0xA9, 中 = 3 bytes, emoji = 4 bytes). Decoding each byte individually fails for every byte of a multi-byte sequence, so each byte is replaced by U+FFFD. The result is unreadable output for any non-English response.

Example before fix (French)

Je peux t'aider avec une grande vari��t�� de t��ches ! Voici quelques-unes
de mes principales comp��tences :
  **R��ponses et informations** : Je peux r��pondre �� tes questions
  **R��daction et cr��ation** : Je peux r��diger des textes, emails, po��mes

The fix

Use codecs.getincrementaldecoder("utf-8") and read in larger chunks. The incremental decoder keeps state across decode() calls, so multi-byte sequences that straddle chunk boundaries are handled correctly.

decoder = codecs.getincrementaldecoder("utf-8")(errors="replace")
while True:
    ch = resp.read(4096)
    if not ch:
        buf += decoder.decode(b"", final=True)
        break
    buf += decoder.decode(ch)

Switching from 1-byte to 4 KiB reads is also noticeably cheaper on CPU.

Test plan

• Run python3 agent.py against a French prompt — accents render correctly
• Verified streaming still works token-by-token in the UI
• No behavior change for pure ASCII output