fix(dml): drop dead mu_hat line in IRM that referenced undefined names

`_fit_one` referenced `g1_hat_full` / `g0_hat_full` behind a
`if 'g1_hat_full' in locals()` guard that always evaluated False, so
the resulting `mu_hat` was unconditionally `None` and never read by
any caller. flake8 (E9/F7/F82 strict subset in CI) caught the unbound
names as `F821 undefined name`, breaking every Linux/macOS lint job.

Removed the dead branch and merged the leftover comment block into a
single explanation of the residuals we actually expose. No behavior
change — `_last_rep_residuals` already returned `{y_resid, d_resid,
pscore}`, which is exactly what `dml_diagnostics` consumes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: brycewang-stanford

src/statspai/dml/irm.py

@@ -205,13 +205,11 @@ def _fit_one_rep(self, Y, D, X, Z, n, rng_seed, sample_weight=None):
             "weighted": sample_weight is not None,
         }
         # Stash IRM-relevant residuals for sensitivity / diagnostics.
-        # For IRM the natural "Y residual" is Y - μ̂(X, D) and the
-        # natural "D residual" is D - m̂(X). Both are length-n.
-        mu_hat = np.where(D == 1, g1_hat_full, g0_hat_full) if 'g1_hat_full' in locals() else None
-        # We didn't save full per-fold g1/g0; reconstruct using a final
-        # within-fold lookup is non-trivial. Instead, surface ψ_scores
-        # and m_hat_full so dml_diagnostics can build overlap plots, and
-        # provide y_resid/d_resid as defined for sensitivity:
+        # We don't save per-fold g1/g0 — surface ψ_scores and m_hat_full
+        # so dml_diagnostics can build overlap plots, and provide
+        # y_resid/d_resid as defined for sensitivity. The natural "D
+        # residual" is D - m̂(X); for "Y residual" we use ψ - θ̂ (see
+        # below).
         d_resid = D - m_hat_full
         # For IRM, y_resid uses ψ - θ̂ (the score residual) which is the
         # influence function input — appropriate for OVB sensitivity per