⚡ Phase 5 — Lightning training loop

WP-039 — a package whose finding is that no code was needed

Deterministic checkpoint and resume was proved rather than implemented: a four-epoch reference run against a run stopped after two epochs and resumed produces the same per-step loss trajectory over the post-resume epochs (rel 1e-4), with global_step, current_epoch and every MuSGD momentum buffer restored. No source changes were needed — Lightning's checkpointing plus existing module and optimizer state already round-trips.

Two scoping choices made that provable: all runs share max_epochs, so the progressive-loss alpha ramp (a pure function of current_epoch / max_epochs, recomputed each epoch) is identical either side of the resume; and the data stream is a fixed unshuffled in-memory dataset, keeping sampler-state restoration out of scope. Without the first, a correct resume would still have produced divergent losses.

WP-040 — overfit-100 on an accelerator

Train-set recall 0.9747 against the 0.95 floor, MPS, 4:47 wall clock, decoded NMS-free through the deployed path.

The structural decision that outlived it: goldens/gpu/ is excluded from default harness discovery and recomputed only under --include-gpu, which is what keeps the offline gate accelerator-free while still holding accelerator numbers under version control. The cost of that split is that a gpu/ golden can go stale unnoticed — shapes_regression_det.json currently has an outstanding re-freeze.

Superseded by WP-167. The outstanding re-freeze is closed: all five files under goldens/gpu/, shapes_regression_det.json among them, were regenerated on an L4. The sentence above is left as written because the staleness it predicted is exactly what happened — golden-gpu returned 46/48 on the pre-WP-169 values — and a log that edits its own correct predictions into the past tense stops being evidence that anything was predicted.