Route augmented noul through choice head for training (Phase C)

[Oaklight]

Summary

Phase C of #104: training integration for augmented noul questions.

• Teacher selection (data/pipeline.py): _select_teacher_probs() selects the best teacher from the nested multi-teacher cache (jev_augmented > luna_augmented for augmented noul, jev > luna > qwen for standard). Fixes existing dead code where synthetic teacher_probs were never consumed by the training loop.
• Training routing (model/training/supervised.py): Augmented noul items with 5-option teacher_probs are routed through the choice head (KL-divergence against teacher distribution) instead of the binary noul head. Standard noul items unchanged.
• Eval routing: Augmented noul predictions mapped back to binary accuracy (pred_key == "true"/"false").
• Uncertainty weighting: Augmented noul uses "choice" log_var since it goes through the choice head.
• 9 new tests for teacher selection, 557 total pass.

Test plan

• 557 tests pass (including 9 new teacher selection tests)
• Verify augmented noul items go through choice head during training
• Confirm standard noul items still use binary path (backward compat)
• A/B training comparison: with vs without augmented noul routing

[milo-oaklight]

PR #113 (krino) — APPROVED ✅

Clean routing of augmented noul through the choice head. Teacher selection logic is well-structured with clear priority chains.

Path	Teacher Priority	Head	Loss
Augmented noul	jev_augmented > luna_augmented > standard	Choice (5-way softmax)	KL divergence
Standard noul	jev > luna > qwen	Binary noul	BCE

Key changes verified:

1. _select_teacher_probs() — Correct priority fallback; empty dicts properly skipped
2. Routing gate — augmented_options and len(teacher_probs) > 2 cleanly separates paths
3. Uncertainty weighting — Augmented noul correctly uses "choice" log_var
4. Eval mapping — pred_key == target_key correctly maps 5-way predictions back to binary accuracy

Minor observation: _compute_augmented_noul_batch processes items sequentially (not true batch), matching the unified head path for standard noul. Fine for now; batch optimization can come later if needed.

Tests: 9 new tests cover teacher selection edge cases comprehensively.

CI green. Ready to merge.