Complete FB50 native LoRA/reference/calibration companion evidence.

Unpack all parts into one directory. With NumPy, run:
python verify_reference_inventory.py
python verify_native_and_calibration.py --root .

The first verifier reconstructs ten reference operators from the retained query
Gram, document products and norms, and checks all method top10/paired contrasts.
The second checks all twelve full200-query native matrices, four80-query Doc-Z
and centered controls, reduced-candidate calibration results and paired contrasts.
Both distinguish query-matched decoy top1 from any-distractor top1. The latter
retains the original score precision; official metrics cast tofloat32. Differences
are reported, not silently reconciled. Actual target validation is retained in
target-accuracy.json, but these CPU verifiers do not repeat model inference.

This is native LoRA-gradient attribution, not all-trained-parameter attribution:
the FB target also trained embeddings/head outside the tracked LoRA inventory.
The shared50-function target differs from the separate100-function transfer
target. All facts here were exposed development data. The original joint CE
reference criterion FAILED on RR, regardless of this FB companion outcome.
Doc-Z is credited to learned_aggregator_v1, paper3v1; it is not a new operator.
https://rstation.ai/problems/rule_routing_attribution/papers/3/v1

All original per-task raw matrices remain in the worker workspace. This package
adds all native and calibration raw scores to the complete reference evidence;
the reference README explains which other AUCs are retained measurements.
Execution records retain completed chunks and the scheduling-only storage stop.
Publication is a separate scientific review step. Code MIT.
