feat(sensei): update agent to consume structured pipeline (DOJ-2439)

[dbejarano820]

Summary

• Removed old additionalContext string-based trigger sections from sensei agent prompt
• Rewrote delegation section with structured JSON parsing protocol for pending-lessons queue
• Added auto-coaching examples for Haiku model calibration across belt levels
• Created manual test checklist for end-to-end hook-to-sensei flow

Closes DOJ-2439 — [SRD-T0-1c] Update sensei agent to consume restructured pipeline

Test plan

• Verify micro-lesson trigger (first encounter) produces 2-3 sentence explanation
• Verify inline-insight trigger (repeat encounter) produces 1-2 sentence explanation
• Verify belt calibration reads from trigger JSON, not profile
• Verify existing /explain, /quiz, /why commands work unchanged
• Verify rate limiting and session cap still function
• Full checklist: docs/testing/doj-2439-test-checklist.md

🤖 Generated with Claude Code

[dbejarano820]

🤖 This PR has had no activity for 2+ days.

If it's still needed, please:

• Resolve any conflicts
• Address review feedback
• Push an update

If it's no longer needed, please close it.

— Dojito (automated PR gardener)

[bitfalt]

this was implemented properly. agent went crazy and pushed the commits with some fixes to main :')