Add strict cross-run holdout test

[rickyjreyes]

Adds a bidirectional 00382466↔00382467 holdout test for the request-48 B0→K*0 μμ sample. Discovery (KDE baseline, high-k scan, amplitudes and phases) is performed on the training run group only; the resulting base and alternative shapes are frozen before evaluation on the held-out group. The held-out statistic is a conditional multinomial log-likelihood ratio calibrated under the frozen training-null shape with 100,000 toys. No held-out frequency/Koide/phase/amplitude selection is allowed.

The workflow streams only the required ROOT branches from the public LHCb request-48 HTTPS files and uploads the JSON/CSV results as an Actions artifact.

This is intended as internal run-group predictive replication, not independent experimental confirmation or a WCT-specific test.