Can you expand on this point? This is one aspect of TDD, which seems to focus most intensely on unit testing, I've never understood. Why wouldn't full-stack, acceptance testing be just as effective?
Also full stack acceptance tests take ages to run (by definition they will be doing some complex computation). As mentioned above, if their test suite runs in under 5s, it gives a huge advantage. Imagine having light on your computer that goes red any time you make a mistake. They effectively have that.
I also think writing the test before the feature will cause you to write better assert statements. If you write the test afterwards you'll tend to either assert everything (which leads to every test breaking on a small change), or to be horribly confused about what to assert and what not.