Fix unbounded query context growth in chat sessions

[kovtcharov]

Problem

`src/gaia/chat/ui/server.py:690-723` — chat history grows without checking total token count against the model's context window:

messages = db.get_messages(request.session_id, limit=20)
# Takes last 4 pairs = 8 messages
# If each message is 10KB, that's 80KB of context per query
# Over many queries, LLM context fills up, response quality degrades

No truncation if history + RAG context + system prompt approaches model limit.

Impact

• Long conversations degrade response quality as context fills up
• Very large messages (pasted documents, code blocks) overflow context
• No token counting before sending to LLM
• RAG chunks may be pushed out of context by long history

Proposed Fix

1. Estimate token count of history before sending to LLM
2. Implement sliding window with token budget (e.g., 80% of model context)
3. Truncate oldest history pairs when budget exceeded
4. Reserve fixed budget for RAG chunks and system prompt
5. Add `max_history_tokens` config parameter
6. Log when history truncation occurs

Files

• src/gaia/chat/ui/server.py (lines 690-723, _get_chat_response)
• src/gaia/chat/sdk.py (conversation history management)

Acceptance Criteria

• History truncated when approaching model context limit
• RAG context always has reserved token budget
• Config option for max history tokens
• Test: long conversation → verify truncation occurs cleanly

[kovtcharov]

Status: Partially Fixed

What works: Session cleanup exists with age-based TTL (30 days) and count limit (50 sessions) via SessionManager.cleanup_old_sessions().

What's still missing: Per-turn conversation history within a session grows unbounded. The conversation_history list in agent.py can grow indefinitely per turn. No max_history limit, no token-level context truncation, no sliding window.

Remaining work:

• Add max_history_pairs limit to conversation history (e.g., keep last N turns)
• Add token-level context truncation to prevent exceeding model context window
• Consider summarization of older turns rather than hard truncation