Ned and Larry.
Is it allowed to use the patch from issue27358 for fixing a regression in 3.6 and inconsistency in 3.5 for var-keyword arguments? The patch introduces new private function _PyDict_MergeEx() and touches the implementation of PyDict_Merge(). Maybe this can be fixed with smaller patch, but I don't know.
Is it allowed to add new opcode BUILD_TUPLE_UNPACK_WITH_CALL for fixing a regression in 3.6 and inconsistency in 3.5 for var-positional arguments? This is the only way. The bytecode already was changed in a bugfix release of 3.5 for more serious need: fixing potential segfault or security issue (issue27286).