Draw posterior samples for one diagnostic trial, insisting on the full count

sample.nsbi_posterior() returns fewer rows than asked for when a bounded prior and a leaky estimator leave rejection sampling short, and only warns. The diagnostics cannot absorb that quietly. sbc() bins its ranks against n_posterior_samples, so a trial that came back short is scored on a scale it was never drawn on, and the ranks compress toward zero; rescaling that one trial on its own would instead make it incomparable to the others. Either way the run reports a miscalibrated posterior when the real cause is lost draws, so stop and say so.

Usage

diagnostic_draws(post, n, trial)