What to conclude

FM-01 should remain the primary operational model.
BARTHAG alone carries a lot of signal, but its compressed nature means it overlaps materially with seed and the core efficiency terms.
The core block is the main source of operational forecasting lift.
Four Factors are useful for interpretation, but under the current sample they add more instability than score improvement.
FM-03 is the formal challenger because it is the closest validated alternative to FM-01 in the current logistic roster.
FM-02b is worth carrying as a context-diagnostic layer, but not as a primary bracket driver.
NL-02 does not capture enough useful nonlinearity to justify promotion; in this pass it mostly overfits.
Any future mixture-of-models layer is still deferred; it would need its own bounded task family and should not be conflated with the current direct hierarchy.