Agent produces garbled multilingual output (Chinese/Cyrillic/Korean) when session exceeds context limit

Bug: When a CoPaw agent session grows too large, the LLM begins producing
garbled output — mixed Chinese, Cyrillic, Korean, and random code fragments.
This has occurred 3 times consistently on the same agent.

Steps to reproduce:
1. Run a CoPaw agent through multiple long sessions
2. Allow session/memory files to accumulate without clearing
3. Agent begins producing garbled multilingual garbage output instead of normal responses
4. Starting a "new chat" does not fully resolve the issue as memory files persist and reload

Expected: Agent should gracefully handle context limits — truncate old context,
summarise, or warn user rather than producing garbage output

Actual: Agent outputs garbled multilingual text with no warning. "New chat" button
does not clear memory files, so the problem persists across new sessions.

Root cause (suspected):
- Session files accumulate across chats (observed: 9.3MB of session files) --Is new chat, really a new chat?
- Memory files load on every new chat regardless of relevance
- No context window management or summarisation kicks in before degradation
- Model loses coherence when context is overloaded

Environment:
- CoPaw version: latest
- OS: Linux (Docker container)
- Sessions directory: /workspaces/cto/sessions/ and /workspaces/default/sessions/

Workaround: Manually delete all session JSON files from the sessions directories

Suggested fix:
1. Implement context window limit — auto-summarise or truncate when approaching limit
2. "New chat" should clear or archive session files, not just start a new one
3. Memory file loading should use semantic retrieval (only relevant context)
rather than loading everything on startup