Why does a pytest monkeypatch sometimes have no effect, and where should you patch?

Overpatching can make a test pass while bypassing the behavior it claims to exercise. I patch the narrow external boundary, configure a realistic return or exception, and assert the observable result first; interaction assertions are reserved for important effects such as charging once. I use raising=True or a spec-like guard where possible so a renamed attribute fails the test instead of silently creating a fake one. When patching becomes pervasive, I prefer an explicit dependency boundary that application code and tests can both understand.