Reject negative-dose samples in mse_optimal_bandwidth (P1)

CI AI review P1: mse_optimal_bandwidth did not enforce the HAD
support requirement D_{g,2} >= 0. Negative-dose samples could
silently pass through both boundary branches: boundary=0 (accepted
via _at_zero even with d_min<0) and boundary=float(d.min()) (accepts
any lower edge). The symmetric nprobust kernel would happily
calibrate a two-sided interior bandwidth while the downstream
one-sided fitter runs on [boundary, boundary+h] -- silent
assumption violation.

Front-door check added: np.any(d < 0) raises ValueError with a
message citing the paper's support assumption.

Two new regression tests:
- test_negative_dose_rejected_boundary_zero: d ~ U(-0.5, 0.5) with
  boundary=0 raises.
- test_negative_dose_rejected_boundary_at_d_min: d ~ U(-1, -0.1)
  with boundary=d.min() raises.

Deferred (P3 same as last round): tri/uni/shifted-boundary golden
parity extension. All three kernels share lprobust_bw, so epa parity
transitively covers kernel dispatch.

177 tests pass (up from 175).

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

Author: igerber

diff_diff/local_linear.py

@@ -622,6 +622,25 @@ def mse_optimal_bandwidth(
     if not np.isfinite(boundary):
         raise ValueError(f"boundary must be finite; got {boundary}")
 
+    # HAD support restriction: de Chaisemartin et al. (2026) Assumption
+    # (dose definition in Section 2) treats ``D_{g,2}`` as the period-2
+    # treatment dose with ``D_{g,2} >= 0``. Negative dose values are
+    # outside the HAD design and would silently calibrate the selector
+    # against a symmetric-kernel two-sided problem while the downstream
+    # fitter remains one-sided. Reject front-door rather than produce a
+    # plausible bandwidth on a malformed input.
+    d_neg = d < 0.0
+    if np.any(d_neg):
+        n_neg = int(d_neg.sum())
+        min_neg = float(d[d_neg].min())
+        raise ValueError(
+            f"Negative dose values detected in d (n_neg={n_neg}, "
+            f"min={min_neg!r}). The HAD estimator (de Chaisemartin et "
+            f"al. 2026) requires the period-2 dose D_{{g,2}} >= 0. "
+            f"Nonnegative-dose data is required for both Design 1' "
+            f"(d_lower = 0) and Design 1 (d_lower > 0)."
+        )
+
     # Boundary-applicability check (Phase 1b scope).
     # The exported wrapper is scoped to the two documented HAD
     # nonparametric estimands:

tests/test_bandwidth_selector.py

@@ -376,6 +376,24 @@ def test_negative_boundary_rejected(self):
         with pytest.raises(ValueError, match="not at a supported HAD estimand"):
             mse_optimal_bandwidth(d, y, boundary=-0.1)
 
+    def test_negative_dose_rejected_boundary_zero(self):
+        """HAD requires D_{g,2} >= 0. Negative-dose samples must be
+        rejected at the wrapper level, even under boundary=0."""
+        rng = np.random.default_rng(2026)
+        d = rng.uniform(-0.5, 0.5, size=1500)  # mix of negative and positive
+        y = d + rng.normal(0, 0.3, size=1500)
+        with pytest.raises(ValueError, match="Negative dose values"):
+            mse_optimal_bandwidth(d, y, boundary=0.0)
+
+    def test_negative_dose_rejected_boundary_at_d_min(self):
+        """Negative-dose samples must also be rejected when the caller
+        tries to pass boundary=d.min() < 0."""
+        rng = np.random.default_rng(2026)
+        d = rng.uniform(-1.0, -0.1, size=1500)  # entirely negative
+        y = d + rng.normal(0, 0.3, size=1500)
+        with pytest.raises(ValueError, match="Negative dose values"):
+            mse_optimal_bandwidth(d, y, boundary=float(d.min()))
+
     def test_mass_point_design_rejected(self):
         """Design 1 mass-point case (boundary > 0, modal fraction > 2%)
         must be rejected with NotImplementedError pointing to 2SLS."""