User-facing opt-out: nondifferentiable

The problem

Part of a computation is sometimes structural rather than estimated — a grid of integration nodes, a clamp location, any fixed hyperparameter that is where to compute rather than what to estimate. primal and primal_distribution already give this discipline to the package's own internals: strip an AD wrapper to its underlying value so the quantity built from it stays off the tape. What is missing is a deliberate, user-facing entry point that applies the same discipline to a function or constructor the caller names, so intent is expressed at the call site instead of threading primal through downstream code by hand.

The fix

nondifferentiable(f) returns a callable that strips every argument to its primal via primal, calls f on the stripped arguments, and strips the RESULT the same way, so the call contributes exactly zero derivative on every supported backend regardless of what f computes internally:

using EpiAwareADTools

window_midpoint(lo, hi) = (lo + hi) / 2
frozen_midpoint = nondifferentiable(window_midpoint)

frozen_midpoint(0.0, 1.0)

0.5

A struct's own constructor is itself callable, so wrapping it the same way — nondifferentiable(QuadratureGrid) — holds construction out of differentiation too, once the struct's own type has a primal method (the same pattern primal_distribution follows for a distribution's parameters). There is deliberately no generic reflection-based primal fallback for an arbitrary struct: isstructtype is true for Dict, Module and every concrete function type — a closure, or typeof(sin) — as well as a user's own type, so a blanket fallback over every struct type would silently mishandle values this package was never asked to touch.

CAUTION — a captured value, not just an explicit argument, is also held constant: if f is a closure over a live differentiated value, that captured contribution is silently dropped too, consistently across every backend. This is the correct, deliberate consequence of "everything in this call is a constant" — never close over a value you still want differentiated.

This machinery is thin. NonDifferentiable and its per-backend rules are documented in full on the Internal API page; nondifferentiable itself is the public entry point.

Upstream target

nondifferentiable generalises primal's discipline to an arbitrary function, so it shares primal's target: no single owner exists for a cross-backend "hold this constant" primitive. This entry is deleted alongside primal/primal_distribution once a shared stop-gradient primitive exists that marks a value or a function call non-differentiable uniformly across ForwardDiff, ReverseDiff, Enzyme, and Mooncake.