nx_shapley.nx

nx_shapley.nx -- game-theoretic feature attribution. Shapley 1953 ("A Value for n-Person Games"): the UNIQUE fair allocation of a coalition's value back to its members. The substrate's third attribution primitive after nx_attribution (top-K of W @ x) and nx_patch_attribution (Cohen's d per axis). phi_i = sum over S subset N\{i} of |S|! * (N-|S|-1)! / N! * [ f(S union {i}) - f(S) ] Closes the WORLD-CLASS attribution gap named in nxc2/docs/TOPDOWN_BOTTOMUP_GAP_ANALYSIS.md. No existing image-gen / LLM stack ships Shapley in typed bits-up form; they all defer to SHAP / KernelSHAP Python libs that hide per-coalition evaluation under floating-point opacity. Two modes: EXACT -- N <= 10 features; full 2^N coalition enumeration. SAMPLED -- N > 10; Castro/Gomez/Tejada 2009 permutation sampling estimator. Unbiased. Caller supplies f(active_mask, n_features, ctx) returning a scalar value in Q10. Substrate computes the contribution vector phi[] (Q10 each) such that sum(phi) + baseline = f(all_on). genealogy_id: shapley_1953_castro_2009 lineage_id: substrate_attribution_v3 5W+H+GLP linkage: each phi[i] is the WHY answer for "feature i contributed how much to output f(all_on)". Caller emits one Nx5whRecord with what=ATTRIBUTION per non-trivial phi[i].