The more logic we push out of band into sidecars, the harder application issues become to debug. For example, let's say an Envoy config change is made centrally and all of sudden my app breaks because an HTTP header has stopped being set. Before I would easily be able to write a unit test to fix such a thing. Now I would need to replicate the envoy config in a test environment etc.
The sidecar model is different to integrating with a 3rd party API as it is designed to operate transparently. My Integration test might be passing but when running with the sidecar, traffic can be mutated etc