We have to trust our users. There’s so many ways to screw up a Stan program that we can’t really provide much protection.
I’d trust a fair bit more than 0% because some users are careful with software and test as they go.
And as others have noted, we wouldn’t do this without finite diff or autodiff tests.
What we’re likely to add is a way to define a function with gradients, not just define gradients w.r.t. the log density for parameters.
Then users will only define gradients w.r.t. constrained parameters, so we’ll need to chain the transforms and Jacobians for any constrained parameters.