Stay in xarray in do() producers; simplify with native broadcasting

[drbenvincent]

Summary

Closes #356.

The do() producers (run_do_pymc, run_do_panel_unified) now assemble labelled xr.DataArrays end-to-end. The leftover .to_numpy() → reshape → _da_chain_draw* round-trips from the flat-dict era are gone, and so is the adapter machinery that existed only to re-wrap numpy back into labelled arrays.

Rather than swap the old numpy wrappers for new single-use ones, the producers lean on xarray's native broadcasting and named reductions so they read as label algebra:

• Constants by broadcasting. A coord-carrying ones = _chain_draw_ones(post) turns every intervention/exogenous fill into ones * value — no manual dims/coords or np.full/np.broadcast_to.
• Mean path inlines inverse_link → isel(subgroup) → .mean(obs_dims); the g-computation standardization is one expression.
• Predictive path just renames the observation dim to unit; _stack_sample flattens it at the public .draws() boundary.
• Panel uses _panel_unit_mean (average units → (time, chain, draw)) and _spread_over_time for scalar/per-time interventions.
• _apply_inverse_link dropped its dead numpy branch and dispatches np.exp/pt.exp on the input type, operating directly on DataArrays (no xr.apply_ufunc).

Helpers removed: _da_chain_draw, _da_chain_draw_unit, _da_time_chain_draw, _const_chain_draw, _const_time_chain_draw, _select_subgroup, _mean_to_chain_draw, _predictive_to_chain_draw, _panel_to_time_chain_draw, _extra_dims, _chain_draw_coords. Kept/added only small label-only helpers used across multiple call sites: _obs_dims, _chain_draw_ones, _spread_over_time, _panel_unit_mean, _fill_missing_posterior_rvs, _predictive_source, _exog_value.

Small cleanups: _model.py prob() / sensitivity() use public .draws(); repr sample counts come from dim sizes.

Size: simulate.py is 79 lines smaller than main (985 vs 1064), with no numpy bridges remaining in the producers.

Test plan

• uv run pytest tests/test_xarray_storage.py tests/test_estimand_result_api.py tests/test_reprs.py tests/test_panel_by_time.py tests/test_causal_queries.py tests/test_att_atu.py
• Full make test: all pass, 87.8% coverage
• make lint (ruff, ruff-format, mypy, license checks)
• Grep gate: no .to_numpy() / numpy re-wrap on PyMC/ArviZ outputs in producers
• Added coverage for time-varying (array-valued) panel interventions — the (n_times,) set path through run_do_panel_unified was previously exercised only with scalar values. New tests assert the intervened variable's per-time draws track the supplied array exactly.

Notes for reviewers

• No tests referenced the deleted helpers or the old xr→np→xr machinery, so none were removed; the inverse-link mean path (bernoulli/poisson) is covered by test_bernoulli.py / test_families.py.
• _panel_unit_mean and the predictive path identify the observation dims positionally (obs[0] = time, obs[1] = unit) because the scan compiler names them generically (mu_sales_dim_0 / _dim_1). This (time, unit) ordering was confirmed against a live scan model; if the compiler ever changes it, this is the spot that would need a coord-name lookup instead.

[drbenvincent]

@BugBot review

[cursor]

PR Summary

Medium Risk
Touches core causal do() propagation for cross-sectional and panel models (g-computation standardization, predictive unit flattening, time-varying interventions); behavior is intended to match prior semantics but regressions would affect ATE/CATE/ATT outputs.

Overview
do() simulation (run_do_pymc, run_do_panel_unified) now assembles labelled xr.DataArrays directly instead of .to_numpy() → reshape → manual wrappers. Interventions and exogenous fills use _chain_draw_ones broadcasting; the mean path does inverse link → subgroup isel → .mean(obs_dims) on arrays; panel paths use _spread_over_time and _panel_unit_mean.

Removed the old _da_chain_draw* helpers and consolidated shared logic into _fill_missing_posterior_rvs, _predictive_source, _exog_value, and _sample_count (repr draw counts). _apply_inverse_link works on DataArrays and PyTensor via a single dispatch. PathModel.sensitivity() and prob() call public .draws() instead of ._draw().

New panel tests assert array-valued set interventions track the supplied per-time values through by_time().

^{Reviewed by Cursor Bugbot for commit 858bfb9. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 858bfb9. Configure here.}