For most types of problems, the feedback is clear - if it works and gives you the expected result, then it's (at least mostly) correct. For implementing cryptographic systems, a passing test suite doesn't mean anything, you'd need extensive expert review to tell you where you went totally wrong, and without it you'd just likely learn untrue things.
I prefer this to the more common 'Just don't do it', as you actually state why. The reason is something I knew intuitively but couldn't really (And had actually never devoted any time to) put[ting] it into words. Thank you.
I would actively discourage developers from writing their own encryption software, in case they're tempted to use their work in a production app. Encryption is sufficiently complicated that you will get it wrong, and sufficiently important that getting it wrong will be very bad. Leave encryption to cryptographers.
That would be bad, so maybe an obfuscator instead?
Some kind of game.
Edit: Something that uses an API.