fix(session): prevent context overflow by adding safety margin to compaction check

[o1x3]

Summary

Fixes #6068 - Auto compaction not triggering reliably with Anthropic models causing "prompt is too long" errors.

Root cause: The overflow check happens before system prompts (~25K tokens) and tool definitions (~10K tokens) are added to the request, so 40-60K tokens of overhead aren't accounted for.

Changes:

• Add 50K overhead buffer to isOverflow() calculation
• Add mid-turn overflow detection in finish-step handler
• Add 30K truncation for MCP tool outputs and webfetch results

Test plan

• [x] Typecheck passes
• [x] Session tests pass
• [x] Tool tests pass