Propagate initial values to reduced variables for diag attribute (#154)

When CvxAttr2Constr creates reduced variables for dimension-reducing
attributes (e.g. diag=True), the initial value was not propagated to
the reduced variable, causing NLP solvers to fail. This mirrors the
existing value propagation already done for parameters.

Also handle sparse diagonal matrices in lower_value() by using
.diagonal() instead of np.diag() which doesn't accept sparse input.

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: Transurgeon

cvxpy/reductions/cvx_attr2constr.py

@@ -120,7 +120,7 @@ def lower_value(variable, value=None) -> np.ndarray:
     if attributes_present([variable], SYMMETRIC_ATTRIBUTES):
         return value[np.triu_indices(variable.shape[0])]
     elif variable.attributes['diag']:
-        return np.diag(value)
+        return value.diagonal() if sp.issparse(value) else np.diag(value)
     elif variable.attributes['sparsity']:
         if full_size:
             return np.asarray(value)[variable.sparse_idx]
@@ -217,6 +217,8 @@ def apply(self, problem):
 
                     reduced_var = Variable(n, var_id=var.id, **new_attr)
                     reduced_var.set_leaf_of_provenance(var)
+                    if var.value is not None:
+                        reduced_var.value = lower_value(var)
                     id2new_var[var.id] = reduced_var
                     obj = build_dim_reduced_expression(var, reduced_var)
                 elif new_var: