You're reading too much into the disproportionate effect mistakes in economics have had on society. Science get things wrong all the time...failure is part of the learning process.
As for physics. The idea of not changing course is exactly what sets apart good theories from bad. If you have a new theory and it does not include the old at a limit then it is a good sign that you are doing something wrong. Newton was not wrong, just inaccurate. Einstein did not invalidate Newton, only corrected the extreme cases.
The social sciences don't have to because they aren't based on the scientific method - mainly the design and independence of repeatable experiments with controlled variables.
Hard science equations have no room for bullshit whereas the social ones do - hence EMH is still taught.
Also...I'm not an economist so I don't really know what's going on in this area...but maybe EMH is still being taught because it's the best model we have at the moment?
Are you sure it's being taught everywhere? Maybe some places have already dropped it? Change takes time.
EMH is taught globally and it will continue to be taught until such a point as people stop believing.
The same is true of many physical sciences - geophysics, oceanography, climate science, astronomy, etc. Your criticism applies to basically any scientific theory which has poorly understood microfoundations (i.e., a lot of them).
The EMH is taught because it's a useful approximation to reality, even if it's imperfect. Or, as the article puts it: Whether markets are efficient or not, and whether P = NP or not, there is no doubt that there will be markets that can allocate resources very close to efficiently and there will be algorithms that can solve problems very close to efficiently.
Incidentally, the EMH claims that financial crises are unpredictable. So the lack of useful predictions of the financial crisis is evidence in favor of the EMH.
[edit: Note, in response to Dn_Ab that 3SAT, the problem considered by the paper, is NP complete.]
Or for markets, aspects of it may invovle solving NP-Hard problems with efficient approximations that are themselves NP-Hard (you are better placed to opine on whether such a possibility is likely).