The first development full-pipeline run completed 1,200 predictions and printed:
typed F1: COPY .8956512924480486; prior136 .9763311247704958; new .9984887867980414.
Birth new-endpoint links: TP302 FP74 FN1, F1 .8895434462444771.
It then failed while serializing a NumPy int64 into the prediction JSONL.
No final evaluation or final data generation had occurred.

The serializer is corrected to convert NumPy numeric types without changing
numeric calculations. The weakness also motivates a separate learned growth
availability gate, trained on source and frozen-backbone predicted geometry.
No target geometry is supplied to that gate at runtime.

Pointer weights also underfit some legal extreme region/cell priority drills.
The next training version gives those drills equal batch mass and includes
their pairwise errors in development selection. All initial weights preserved.
