Accept Annotation{<:FunctionWrappersWrapper} in Enzyme rules (closes #48)

[ChrisRackauckas-Claude]

⚠️ Please ignore this PR until reviewed by @ChrisRackauckas.

Summary

Closes #48 (Missing EnzymeRule — MethodError: no method matching forward(...::Duplicated{FunctionWrappersWrapper...}...)).

The existing forward / augmented_primal / reverse rules in FunctionWrappersWrappersEnzymeExt dispatched only on func::Const{<:FunctionWrappersWrapper}. Issue #48's repro (NonlinearSolve + SciMLSensitivity + Enzyme.Forward) drives the rule with func::Duplicated{<:FunctionWrappersWrapper} instead, which fell off to custom_rule_method_error.

This PR relaxes the func argument type to Annotation{<:FunctionWrappersWrapper} across every rule (4 forward × 4 reverse), so Duplicated / BatchDuplicated function annotations dispatch through the same code path as Const.

Why is this safe?

A FunctionWrappersWrapper only carries FunctionWrappers plus cache storage — none of those fields have a meaningful tangent. The function shadow is therefore ignored: we still unwrap func.val (the primal wrapper) and delegate to Enzyme.autodiff(..., Const(f_orig), ...) exactly as the previous Const-only path did. The argument shadows, the return shadow, and the runtime-activity / strong-zero flags continue to flow through unchanged.

Enzyme drives the rule with a Duplicated function annotation when the outer autodiff call is differentiating through a closure that captures an FWW — Issue #48's repro is one such case.

Reproduction (minimal)

using FunctionWrappersWrappers, Enzyme, EnzymeCore
using EnzymeCore.EnzymeRules

f!(residual, u, p) = (residual[1] = u[1]^2 - p[1]; nothing)
fww = FunctionWrappersWrapper(
    f!, (Tuple{Vector{Float64}, Vector{Float64}, Vector{Float64}},), (Nothing,)
)

config = EnzymeRules.FwdConfig{false, false, 1, false, false}()
EnzymeRules.forward(
    config,
    Duplicated(fww, fww),                # <-- the failing dispatch
    EnzymeCore.Const{Nothing},
    Duplicated([0.0], [0.0]),
    Duplicated([2.0], [1.0]),
    Duplicated([1.0], [0.0]),
)

Before: MethodError: no method matching forward(::FwdConfigWidth{1,...}, ::Duplicated{FunctionWrappersWrapper{...}}, ::Type{Const{Nothing}}, ...)

After: residual=[3.0], dresidual=[4.0] — primal and shadow propagate correctly.

Tests

Adds four new testsets exercising each rule branch with Duplicated(fww, fww):

• Enzyme forward, Duplicated FWW annotation — IIP Const return (the issue's failing pattern)
• Enzyme forward, Duplicated FWW annotation — shadow-only return ({false, true} rule)
• Enzyme forward, Duplicated FWW annotation — primal + shadow return ({true, true} rule)
• Enzyme reverse, Duplicated FWW annotation — Const return IIP

Local test results on Julia 1.11 (Enzyme v0.13.147):

	Before	After
Enzyme extension passed	39	50 (+11)
Pre-existing failures	1	1

The 1 pre-existing failure (Enzyme batch forward mode (width > 1) — TypeError: in typeassert, expected Tuple{Float64,Float64}, got @NamedTuple{1::Float64, 2::Float64} inside Enzyme's enzyme_call) reproduces on unmodified main and is unrelated to this change — looks like an Enzyme upstream issue with BatchDuplicated tuple shadows; investigating separately.

Note on the full SciML repro

Issue #48's NonlinearSolve + SciMLSensitivity repro will still hit a different (deeper) Enzyme runtime error after this fix — but that error is no longer the custom_rule_method_error the issue reports. The specific MethodError for forward(...::Duplicated{FunctionWrappersWrapper}...) is what's closed here; remaining downstream issues belong in NonlinearSolveBase / SciMLSensitivity and aren't fixable in this package.

Test plan

• New testsets pass locally (Julia 1.11, Enzyme v0.13.147)
• Existing testsets still pass (no regressions introduced)
• Pre-existing Enzyme batch forward (width > 1) failure reproduces identically on main (not caused by this PR)
• CI green on all jobs

🤖 Generated with Claude Code