Missing values

NaN is a first-class counterfactual value: "close the delinquency record" can be a legitimate, priceable recommendation.

from treecf import AllowMissing

AllowMissing("months_since_last_delinquency", delta_miss=2.0)
AllowMissing("bureau_score", delta_miss=3.0, delta_from_miss=1.0)   # asymmetric

• Without AllowMissing, a feature never becomes NaN, and a NaN factual stays fixed (routing follows each node's missing direction).
• With it, the missing state is one more option: flipping value → NaN costs delta_miss (in the feature's normalized units), NaN → value costs delta_from_miss (defaults to delta_miss). There is deliberately no default for these deltas — MAD-based defaults are meaningless for this transition.

Interaction with linear constraints

A Linear constraint that references a missing feature is resolved by its missing_policy:

policy	meaning
"satisfied" (default)	vacuously true — "no delinquency history" satisfies a DPD-consistency rule
"forbid_missing" / "violated"	the counterfactual may not use NaN for the referenced features

Mined constraints (suggest_constraints) also report missingness links (miss(A) => miss(B)) as joint-AllowMissing recommendations.