nx_id_param_gate.nx

nx_id_param_gate.nx -- THE IDENTITY-PARAMETER CORE AND ITS FIRST CONSUMER, GATED. /compare/procgen C36 (id_param_core) and C33 (gn_palette_sample). Subjects: nx_id_param_core, nx_id_palette. WHAT MAKES THIS GATE NON-VACUOUS. A "correlated sampler" is the easiest thing in the world to fake: return the master for every axis and every correlation is perfect. So the teeth here are the properties the WRONG implementations provably cannot all satisfy at once, and BOTH wrong implementations are kept RUNNABLE in the subject rather than described in a comment: * the INCUMBENT (independent uniform draws, loading 0) -> dies at T3, T6 * a COLLAPSED palette (every axis IS the master) -> dies at T7, keeps T3/T6 -- which is exactly why T7 has to exist * a correlation bought by narrowing the spread -> dies at T4 * loadings that drifted from their ledger -> dies at T8 T6 IS THE LOAD-BEARING TOOTH AND ITS BAR IS NOT A CONSTANT. A one-factor model predicts every PAIRWISE correlation from the two loadings alone -- corr(a,b) = load_a*load_b/UNIT -- a non-obvious three-way prediction nothing but that structure produces. It is measured with a real Pearson estimator over the whole cast, against a bar DERIVED from the declared loadings. THE TOLERANCE IS DERIVED, NOT PICKED. The standard error of a Pearson r is about (1-r^2)/sqrt(n). At n=512 and the weakest declared pair (r=0.45) that is ~35 permil, so IPG_CORR_TOL=60 leaves under 2x headroom -- enough for sampling noise and integer rounding, not enough to admit a different model. Stated so a later reader can tighten it against a bigger cast rather than guess. license_tier: ORIGINAL expect_exit: 0 No hw writes (Rule 26).