Fix NoMethodError when a caller location has no path

[koic]

Thread::Backtrace::Location#path returns nil for the Enumerator frames (e.g. "Enumerator::Generator#each") that appear in caller_locations during an external iteration of Enumerator (e.g. Enumerator#next). When a TracePoint callback fires inside such an iteration, PowerAssert.internal_file? raises NoMethodError and the callback reports:

power_assert: [BUG] Failed to trace: NoMethodError: undefined method 'start_with?' for nil

This happens in practice when a block-style assertion drives Capybara, whose Capybara::Result iterates matched elements with Enumerator#next.

While #31 originally suspected a CRuby issue, the type signature of Thread::Backtrace::Location#path in ruby/rbs allows nil (() -> String?), and this nil case is observable at least on Ruby 2.6.10, 2.7.8, 3.1.5, 3.3.10, 3.4.10, and 4.0.6, so it seems reasonable for power_assert to tolerate nil either way.

Treat locations without a path as internal ones since they can never be the assertion's source location. Also guard app_context?, which walks caller locations the same way.

Closes #31.

[k-tsj]

internal_file? returns true when the given file belongs to the power_assert library. I think it should return false when file is nil.

[k-tsj]

Except for parser-related code, this project generally does not test individual methods directly. Instead, tests check that the expected power assert message is generated.

For this issue, I think it would be better to add the following reproduction test:

    t do
      obj = []
      def obj.foo; self; end
      enum = Enumerator.new {|y| y << obj.foo }
      def enum.inspect; '#<Enumerator>'; end
      assert_equal <<END.chomp, assertion_message {
        enum.next.foo
        |    |    |
        |    |    []
        |    []
        #<Enumerator>
END
        enum.next.foo
      }
    end

[k-tsj]

it seems reasonable for power_assert to tolerate nil either way.

Agreed. I left a few review comments. Could you take a look?

[koic]

Thanks, I applied both suggestions.

One thing I noticed: the &. in app_context? is not covered by the test suite (removing it still passes all tests). The following test covers it by yielding on a non-target thread, which keeps the :call
TracePoint enabled while the Enumerator fiber runs. It fails with NoMethodError if the &. is removed:

t do
  enum = Enumerator.new {|y| y << 1 }
  message = ::PowerAssert.start(-> { enum.next }) do |pa|
    Thread.new { pa.yield }.join
    pa.message
  end
  assert_equal '', message
end

Should I add this test, or would you prefer to keep the &. as an untested defensive guard?

[k-tsj]

Good catch, could you also add it?

[koic]

Thanks for your review. I've added the test.