fix(distill): non-degenerate smoke batch — per-row input matches label (PMAT-698m)

[noahgift]

Summary

Phase 3 GB10 smoke completed end-to-end after the 7-PR cascade but reported initial_loss=6.0760 → final_loss=8.3914 — loss INCREASED. The pipeline was working; this was a smoke-setup artifact.

Root cause

let dummy_batch = vec![vec![0u32]; batch_size];               // all token 0
let labels      = (0..batch_size).map(|i| i % num_classes);   // 0,1,2,...

Same input paired with N distinct labels. The student is asked to emit different targets for IDENTICAL features — impossible to learn, CE diverges. With teacher==student in the smoke, KD ~ 0 → CE dominates → final > initial.

Fix

Per-row input matches per-row label. Trivial identity mapping (input token → predict same token). CE decreases monotonically.

dummy_batch[i] = vec![(i % num_classes) as u32]
labels[i]      = i % num_classes

For real distillation with a dataset, callers override the pipeline's batch construction entirely.

Test plan

• 58 distill lib tests pass (FALSIFY-APR-DISTILL-TRAIN-001/002 + falsify_pipeline_001_end_to_end_loss_decreases all green)
• Live gx10 dispatch: final_loss < initial_loss (F-DISTILL-SMOKE-001 positive)

Relationship to cascade

The PMAT-698e..m cascade closed Phase 3 dispatch readiness:

#	PR	What
1-8	various	Pipeline mechanics (workspace, format, pre-warm, macro, key align)
9	THIS PMAT-698m	Smoke contract semantics — F-DISTILL-SMOKE-001 actually meaningful now

🤖 Generated with Claude Code