feat(python): expose escalation mode to Python

[rbehal]

What

Binds LlmClassifierConfig::Escalation to Python as libsy.escalation(judge_target, efficient_target, capable_target, *, prompt, confirmations, recent_turn_window, window_message_chars, max_output_tokens), with type stubs and binding tests. Bumps the dist to 0.2.0+gumloop.0.2.0 (wheels for this version are already in the Artifact Registry).

Why

The backend's Chew mid-run step layer (escalation variant) runs libsy's escalation algorithm per agent step: each efficient answer is judged, and a confirmed streak of escalate verdicts latches the run's session onto the capable tier. The capability existed in the Rust core but had no Python door.

Notes

• Sessions ride request metadata (session-id header); a latched session serves the capable target directly with no further judge calls — covered in tests/test_libsy_escalation.py (streak below confirmations, latch, session isolation, judge fail-open, validation).
• Same conventions as the custom_classifier binding; no changes to existing bindings.
• Replaces the direct push to gumloop-main that was force-reset; the gumloop-v0.2.0+gumloop.0.2.0 tag was removed and should be re-pushed on the merge commit.

Made with Cursor