When deciding which criteria to use to determine which irrational beliefs you should keep (or whether any should be kept at all, or how to decide that), you need to start from first principles, and not work backwards from foregone conclusions. In this case, it seems to me that your foregone conclusions were "keeping religion is okay" and "approaches to rationality that allow me to keep religious beliefs are better than those that do not" to some premises that allow for those conclusions.

For example, taking your point (1), if "eschewing your religious beliefs makes you feel depressed and you are unable to work productively" is a valid justification, then "eschewing your Aryan Brotherhood beliefs makes you feel depressed and you are unable to work productively" should be an equally valid justification for some people.

If you really think that the 3 criteria of (1) are valid or can be made valid, try this: replace all references to religion and spiritual and God in the criteria in order to make them generic, and then try to derive some undesirable conclusions from the generic criteria. If you can't, please report back the generic criteria. If you can, then it follows that they're not valid in general, so they're not valid for religion unless you are making an exception for religion (in which case you don't need to waste time trying to justify the conclusion at all, since it's assumed to be true).

I didn't mean to imply that working back from a conclusion is not a perfectly valid proof technique. It's used all the time in math. The problem is that in real-life, due to various cognitive biases, it's easy to use it as a form of rationalization, where instead of creating a chain of statements from the conclusion back to the premises that deductively proves the conclusion, we end up creating a chain of rationalizations, each of which is sort of plausible, without verifying that the statements prove what they're supposed to prove and don't also have undesirable implications. It's far easier to start from the generic, which invokes far fewer biases than starting from the specific.