tests: reduce the cloned 20-row sampler-dispatch grids to a coverage-preserving 9-row design

Problem

The slow suite is dominated by a sampler-dispatch smoke matrix with heavy duplication (~43 min of the ~96-min suite; durations from run 29419740259, py3.12):

• The same 20-row (loglik_kind × backend × sampler × step) grid is cloned in tests/slow/test_mcmc.py:19-40, tests/slow/test_missing_data_mcmc.py:46-67, and tests/slow/test_missing_data_and_deadline_mcmc.py:66-88, crossed with 9 test functions (~140 cases; tests/slow/test_choice_only.py adds a 9-row grid × 4 near-identical functions). Assertions are almost entirely isinstance(idata, xr.DataTree).
• Only ~9 of the 20 rows exercise distinct HSSM code paths:
  • step='slice' rows pass pm.Slice through **kwargs straight to PyMC — zero HSSM code (NUTS rows already evaluate logp+dlogp, a superset).
  • explicit sampler='pymc' duplicates the default-resolution branch (src/hssm/base.py:655-662).
  • blackbox (None,None), ('pymc',None), ('pymc','slice') are three near-identical Slice runs — HSSM injects pm.Slice itself (base.py:670-671).
  • 5 rows per grid are pure argument-validation ValueErrors (base.py:664-668, 694-698) that never sample but pay full model construction (incl. ONNX load): 57 cases, 4.1 min total.
• Sampling settings are already minimal (tune=10, draws=10, chains=1, cores=1) — per-test cost is model build + backend compilation, so the lever is grid size, not draw counts.

Plan

1. In test_mcmc.py, replace the 20-row grid with a 9-row covering set: analytical/default, analytical/numpyro, approx_differentiable-pytensor/default, ad-pytensor/numpyro, ad-jax/default (exercises default→numpyro resolution), ad-jax/pymc (the only JAX-op-under-PyMC-NUTS coverage), blackbox/default (auto-Slice injection), blackbox/numpyro→ValueError, analytical/numpyro+slice→ValueError. Keep one real slice row (e.g. analytical/pymc/slice) as PyMC-passthrough smoke.
2. Slim the two missing-data clones to ~4 rows (one per loglik_kind/backend build): their unique value is the missing-net wrapping under NUTS (src/hssm/hssm.py:373-380), not sampler dispatch.
3. Consolidate the 3 validation rules into ~3 fast (non-slow) pytest.raises unit tests on a cheap analytical DDM.
4. test_choice_only.py: run the full grid only on test_choice_only_default_params; run the regression variants (beta / logit1 / combined) on the default row only. Keep both backend variants — jax wrap vs pytensor direct is real dispatch (src/hssm/distribution_utils/dist.py:804-816).
5. Keep the log-likelihood recompute equality checks (test_mcmc.py:72-80, test_missing_data_mcmc.py:91-98) on every surviving sampled row — the only numeric end-to-end validation of that path.
6. Attach a before/after coverage diff to the PR: the coverage job runs the full suite, so pruning can move the tracked codecov number; expected delta ≈ 0, demonstrate it once (see #1076).

Rigor floor — do not prune or mock

• One real-sampling test per distinct dispatch path (the 9 rows above).
• End-to-end VI tests tests/slow/test_vi.py:120-147 (issue #1056 regressions).
• Missing-data/deadline ONNX mixtures under actual NUTS.
• aDDM PPC conditioning/continuation tests.

Acceptance

• ~22–26 min off the summed slow suite (verify via --durations=50).
• Coverage diff shows ~no change.
• One shared grid constant (no more per-file clones drifting apart).

🤖 Generated with Claude Code

Start with the cloned grids in tests/slow/test_mcmc.py, tests/slow/test_missing_data_mcmc.py, tests/slow/test_missing_data_and_deadline_mcmc.py, and tests/slow/test_choice_only.py, then review the referenced dispatch and missing-data paths in src/hssm/base.py, src/hssm/hssm.py, and src/hssm/distribution_utils/dist.py. Preserve the listed sampling, validation, VI, missing-data, and PPC coverage while consolidating the grids and moving validation checks to fast tests. Run the full suite with --durations=50 and attach a before/after coverage diff.