Rationale:

• without this users have to manually specify test targets in order to skip the package with the offending test suite. It gets tiring.
• sometimes test suites are badly written, in which case a non-deterministic failure appears in some package. At this point it's not so helpful for stack to refuse to process the dependent test suites, when in previous invocations it would have.
• it's consistent with the behaviour of most test frameworks, which run all tests in a test suite regardless of failures.

Variations:

• don't make this the default, but provide-k and --keep-going flags, à la make.
• don't process dependent packages, but do continue processing test suites for packages that don't depend on the package whose tests don't pass. Then -k would mean blindly test everything no matter what (provided it builds).