The short term fix might be to inherit from TestSafeBrowsingDatabaseManager instead, and ensure that has only methods that call NOTREACHED(). I guess there could still be affects from the ctor/dtor.  Was that the issue?

Longer term, this could definitely be refactored. The current tree looks like this:

SBDbM: 80% is pure virtual, but 20% implements the API Blacklisting code used by desktop+android
\-
  TestSBDbM: Implements all pure-virtual functions with NOTREACHED()
  RemoteSBDbM: Implements a few methods (with calls to GMSCore), and the rest are NOTREACHED().
  LocalSBDbM: Implements desktop V3 DbM.
  V4SBDbM: Implements desktop v4 DbM.

We should find a better way to compose the shared functionality of the Remote/V4 implementations, so the parent class could be a pure interface. We could either use several layers of inheritance (ick), or put the API Blacklisting code in a separate class and have the Remote/V4 impls delegate to that.  That's just a little more code, but much easier to understand.

Yes I believe the issue was in destruction logic. We do already inherit from TestSafeBrowsingDatabaseManager for our fake implementation.