Clarify boundary= contract in mse_optimal_bandwidth docstring (P3)

CI AI review P3: the boundary parameter docs said callers could pass
d_lower (the theoretical support lower bound), but the implementation
requires boundary = float(d.min()) (the sample minimum). Users
following the old docs on e.g. U(1, 2) data with boundary=1.0 would
hit an avoidable ValueError. Docstring now explicitly says use the
sample minimum.

Deferred (P3): extending R golden parity to triangular and uniform
kernels. All three kernels go through the same lprobust_bw code path,
so epa parity transitively covers dispatch; tri/uni parity is a
nice-to-have regression anchor but not a correctness gap.

175 tests pass (unchanged).

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

Author: igerber

diff_diff/local_linear.py

@@ -533,8 +533,15 @@ def mse_optimal_bandwidth(
     y : np.ndarray, shape (G,)
         Outcome values (the first-difference ``Delta Y_g`` in HAD).
     boundary : float, default=0.0
-        Evaluation point ``d_0``. For Design 1' ``d_0 = 0``; for Design 1
-        continuous-near-``d_lower`` pass ``d_0 = d_lower``.
+        Evaluation point ``d_0``. The Phase 1b wrapper accepts only
+        two values (within float tolerance): ``boundary = 0`` for
+        Design 1' or ``boundary = float(d.min())`` for Design 1
+        continuous-near-``d_lower``. Use the sample minimum
+        ``d.min()`` (not a known theoretical lower bound of the
+        support), because the downstream selector operates on the
+        realized data. Any other value -- including
+        ``boundary < d.min()``, interior points, or
+        ``boundary > d.min()`` -- raises ``ValueError``.
     kernel : str, default="epanechnikov"
         One of ``"epanechnikov"``, ``"triangular"``, ``"uniform"``.
     weights : np.ndarray or None, default=None