If 50% of people score the exact same type upon retaking the test, I consider that great evidence that there is relatively stable information in the indicator. As a baseline, remember that if the distribution of results was uniform, a null model would expect something on the order of 6.25% if the noise really drowned out all the signal.