W2 tail ULP v5

This deterministic, code-only Cascade generator retains the complete W2 prior and moves the terminal observation of each generated series one representable float64 value toward positive infinity.

The fixed Cascade training sequence uses causal predictions from patch p to target patch p + 1 and drops the final patch output. Consequently the final observation cannot influence an earlier prediction. Its one-ULP target change is also below float32 resolution after causal normalization. The temporal prior, family mixture, random draw stream, and every other observation remain intact.

Submit as generator.Generator.