What

Fixes tascCODA always returning insignificant results, plus a run_hmc crash.

1. NaN node-effect medians (the main bug)

The arviz 1.0 migration (#911) recomputed the posterior median in summary_prepare via:

summ["median"] = posterior[var_names].median(dim=["chain", "draw"]).to_dataframe().stack().reindex(summ.index)

.to_dataframe().stack() produces a tuple MultiIndex (e.g. ("A", "x", "n0", "alpha")) that never matches arviz's string row labels (e.g. "b_tilde[x, n0]"), so reindex(summ.index) fills the entire median column with NaN.

tascCODA decides node credibility with |median| > delta (__complete_node_df), so NaN medians made every effect non-credible regardless of the data — i.e. "always insignificant".
scCODA is unaffected because it selects effects via inclusion probabilities, not the median, which is why only tascCODA broke.

The fix computes the medians positionally, which matches az.summary's row order exactly (rows ordered by var_names, then C-order of each variable's dims — identical to DataArray.values.flatten()).

2. run_hmc rng-key TypeError

run_hmc converted rng_key to a JAX key before handing it to set_init_mcmc_states, which seeds a NumPy generator and therefore needs an int — raising TypeError: SeedSequence expects int ....
It now keeps the int seed for the NumPy generator and derives the JAX key separately, mirroring run_nuts.

Not included here

While verifying this, I found a deeper, separate problem: even with medians fixed, the (correct) spike-and-slab model recovers no credible effects because the global theta collapses to its prior under numpyro's samplers. That needs a model-level change and is tracked in #1015 with the full analysis.

Tests

No new test is added: the meaningful assertion (tascCODA recovers known credible effects) is blocked by #1015, and asserting only that medians are non-NaN isn't worth pinning.
Existing tests/tools/_coda/test_tasccoda.py passes, and run_hmc now runs end-to-end.