…99 P0)

Two P0 defects surfaced by Stage D 2026-05-20: 25h of GB10 training
produced final_loss=3.58 (real convergence) but a 200-byte empty
model.safetensors. Root cause is a defect-pair:

  Defect 1: pipeline.export() serializes the `student_weights` HashMap,
            which PMAT-698f's APR short-circuit returns empty (the read
            side correctly handles APR; the write side then ships
            nothing). The trained GPU weights in CudaStudentProvider
            were never pulled back to disk.

  Defect 2: zero intermediate checkpoints across the 25h run. A crash
            at step 49999 would have produced identical loss to what
            we observed.

Fix:

1. New trait method `StudentLogitsProvider::save_checkpoint(&mut self,
   path: &Path) -> Result<()>` with a no-op default (preserves
   FixtureStudent behavior).

2. `CudaStudentProvider::save_checkpoint(path)` delegates to
   `trainer.save_apr(path, "albor-distilled-v2", "Qwen2ForCausalLM")`.
   GPU weights → APR v2 file with full metadata.

3. `pipeline.export()` calls `self.student.save_checkpoint(<output_dir>/model.apr)`
   after the metadata sidecar write. CUDA path produces a real APR;
   fixture path still writes only the metadata-only safetensors.

4. Periodic checkpointing in `pipeline.train()`:
   - Every APR_DISTILL_CHECKPOINT_EVERY steps (default 5000)
   - Files: `<output_dir>/ckpt-step-{N:06}.apr`
   - Write failures are logged but don't fail training (loss progress
     is more valuable than pristine intermediate snapshots)
   - Set to 0 to disable (smoke tests)

Test plan:
- [x] cargo check (both with/without cuda feature) clean
- [x] 61 distill lib tests pass (FALSIFY-APR-DISTILL-TRAIN-001/002
      unchanged; fixture path semantics preserved via no-op default)
- [ ] Live gx10 re-dispatch (Phase 4 short re-train, per user direction):
      verify model.apr is written with non-trivial size AND
      ckpt-step-NNNNNN.apr appears every 5000 steps

Unblocks Phase 5 (HumanEval) and Phase 6 (publish), which both require
a model with real weights.

Stage D 2026-05-20 produced loss-verdict evidence (-66% over 25h) but
no usable model. Per user direction (2026-05-21): land this P0 fix,
then dispatch a 5-10K-step re-train (~5h) to produce a usable Phase 5/6
checkpoint. Future runs benefit from periodic checkpoints automatically.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>