feat(chat): context overflow protection — auto-summarize conversation when approaching model context limit

Summary

Long conversations silently hit the model's context window limit, causing the model to lose early context, hallucinate, or fail entirely. Users have no warning and no automatic protection. This is especially acute for LLC agents running multi-hour tasks.

What Users Want

Detect when conversation is approaching the context limit (>80% full — builds on feat(chat): context window usage indicator — show token consumption during conversation #8990)
Auto-summarize the earliest portion of the conversation history and replace it with a compact summary
'Rolling window' mode: always keep the last N tokens verbatim; summarize everything older
User preference: warn only / auto-summarize / hard stop
For LLC agents: checkpoint before summarization so the original history is preserved
Summary injected as a system message so the model knows context was compressed

AutoBot Fit

autobot-backend/chat_history/ manages session history. New ContextCompressor service: monitors token count, triggers summarization via the LLM gateway when threshold is crossed. Summarization prompt: 'Compress the following conversation segment preserving all decisions, facts, and action items.' LLC heartbeat runs: compress between task steps to stay within context.

Acceptance Criteria

Token count tracked per message (uses data from feat(chat): context window usage indicator — show token consumption during conversation #8990)
Warning shown at 80% context fill
Auto-summarize triggered at 90% (when mode = auto)
Summary prepended as system message marker
Original history preserved in DB (only context window is compressed, not the stored record)
LLC agent: auto-compress between tool call batches during long runs