Incorrect forward mode gradient for truncated normal distribution

[mhauru]

MWE:

julia> import FiniteDifferences, Enzyme; using Distributions

julia> func = (a, b, x) -> logpdf(truncated(Normal(), a, b), x)
#29 (generic function with 1 method)

julia> args = (-0.3, 0.3, 0.1)
(-0.3, 0.3, 0.1)

julia> Enzyme.gradient(Enzyme.Forward, Enzyme.Const(func), args...)
(0.6272640800199636, -2.607264080019964, -1.09)

julia> Enzyme.gradient(Enzyme.Reverse, Enzyme.Const(func), args...)
(1.6172640800199638, -1.617264080019964, -0.1)

julia> FiniteDifferences.grad(FiniteDifferences.central_fdm(4, 1), func, args...)
(1.6172640800151292, -1.6172640800193545, -0.0999999999999462)

On Enzyme v0.13.11.

[wsmoses]

A further reduction of this would be helpful to resolving

[mhauru]

Would like to help, but unfortunately I'm too busy right now to find time to minimise further. Open for grabs for anyone to do it.