Spike-and-slab null fraction under selection hinges on slab shape

Refitting a selection model (point null + slab, p<.05 publication filter) to the 73 Fisher-z RPP originals: with a moment (non-local) slab the originals alone give pi0 ~0.87-0.92, with a N(0,tau) slab ~0.21 (LR CI 0-0.8), and the normal slab has the higher likelihood. pi0 here is the null share among all TESTED hypotheses, an extrapolation to unpublished tests; the null share among PUBLISHED originals is ~0.35-0.58 in the joint fits. Before quoting a '90% nulls' figure, report which population and which slab, and check a shape-robust quantity such as P(|effect|<eps). Quadrature over an effect grid with cached phi matrices (they depend only on data and fixed h) made each fit ~25x faster.