Make the test for dependency cycles less brittle.

[cgruber]

Loosens the brittle assumptions of the cycle detection test which implicitly relies on method processing order. We don't actually care that A or C is found first - they're a cycle - any node will do. We just care about the correct error being thrown. Method processing order is simply undefined but happens to be deterministic within a JVM release, and happened to change in Java8.

[cgruber]

Side note - cycle detection is gross on Dagger 1, in that it isn't really a unified thing w.r.t. other analyses. This is partly because the ProblemDetector (which does this) is part of the runtime, so the error returned is not some validation error, it's an IllegalStateException and the processor which uses this just tosses it.

This could be fixed, but doesn't seem worth it given that 2.0 is close. Cycle detection will fit nicely into the ValidationReport system, and be handled more clearly.

[JakeWharton]

I agree. All efforts should be toward 2.0.

[JakeWharton]

And thanks!