ReshapeReshape could rewrite multiple matches that share the same shape initializer ([2, -1]) into distinct concrete shapes, but reuse the same initializer name. This caused initializer map clobbering and left rewritten Reshape inputs pointing to non-registered values after name-fixup.

• Rewrite rule change (_basic_rules.py)

  • Stop reusing the incoming shape value name for synthesized shape initializers.
  • Generate a per-match initializer name from the matched second reshape node context, then create the new initializer with that unique name.
  • This prevents cross-match overwrite in graph.initializers when multiple chains share one original shape input.

• Regression coverage (_basic_rules_test.py)

  • Add a focused test with two Reshape -> Reshape chains sharing shared_shape=[2, -1], where -1 resolves differently per chain.
  • Run the same optimizer sub-pipeline used in the repro and assert rewritten reshape shape-inputs are distinct and registered initializers.

# before: both rewrites reused shape.name (e.g. "shared_shape")
new_shape = op.initializer(ir.Tensor(self._new_shape, name=shape.name))

# after: per-match unique shape initializer name
self._new_shape_name = f"{context.nodes[1].name or shape.name}/shape"
new_shape = op.initializer(ir.Tensor(self._new_shape, name=self._new_shape_name))