Propagate set_runtime_activity through FWW Enzyme forward rules

[ChrisRackauckas-Claude]

Summary

The existing FunctionWrappersWrappersEnzymeExt forward-mode rules
hard-coded plain Forward when delegating to Enzyme.autodiff,
silently dropping the outer caller's set_runtime_activity(Forward)
(and set_strong_zero) flags. Enzyme's static IR-level activity
analysis can't see through the cfunction indirection inside
FunctionWrappersWrapper, so the inner Enzyme.autodiff call then
raised EnzymeRuntimeActivityError at
broadcast_unalias → mightalias inside @. du = … — even though the
user had explicitly set runtime activity on the outer call.

Motivation

Confirmed reproducer from SciML/OrdinaryDiffEq.jl#3518:

Rosenbrock23(autodiff = AutoEnzyme(mode = set_runtime_activity(Enzyme.Forward)))

on any time-dependent in-place ODE RHS that DiffEqBase's
AutoSpecialize wraps via wrapfun_iip (a 4-arg
FunctionWrappersWrapper over (du, u, p, t) returning Nothing).
Dropping the runtime-activity flag through the extension's delegation
broke that end-to-end path.

Bare-f! vs wrapped-f! reproducer matrix (without this fix):

• Bare f!(du, u, p, t) = @. du = p*u, set_runtime_activity(Forward) → OK, ddu == 0.
• Same f! wrapped in FunctionWrappersWrapper, set_runtime_activity(Forward) → EnzymeRuntimeActivityError.

Mechanism / fix

EnzymeRules.FwdConfig{NeedsPrimal, NeedsShadow, Width, RuntimeActivity, StrongZero}
carries the caller's ForwardMode flags as type parameters. The updated
rules extract RuntimeActivity and StrongZero and rebuild the
ForwardMode used for the delegated Enzyme.autodiff call via
Enzyme.set_runtime_activity / Enzyme.set_strong_zero. The reverse
rules' internal forward-mode helpers are updated analogously using
EnzymeRules.runtime_activity(config) /
EnzymeRules.strong_zero(config) accessors on RevConfig.

Numerical check

The added test set (test/enzyme_tests.jl) exercises the DiffEqBase
wrapfun_iip shape end-to-end:

• f!(du, u, p, t) = @. du = p*u (∂/∂t = 0) — wrapped + set_runtime_activity(Forward) → ddu ≈ 0, matches ForwardDiff baseline.
• g!(du, u, p, t) = @. du = sin(t)*u — wrapped + set_runtime_activity(Forward) → ddu ≈ cos(t) .* u, matches ForwardDiff baseline exactly.
• h!(du, u, p, t) = (du[1] = u[1] * t) under set_strong_zero(Forward) → correct tangent, confirms the strong-zero flag is also propagated.

Unrelated test status: the pre-existing "Enzyme batch forward mode
(width > 1)" test errors on this Julia/Enzyme stack before and after
this change — it fails the shadow_result[1]::NTuple{W, T} typeassert
because Enzyme now returns a @NamedTuple for BatchDuplicated. Out
of scope for this PR.

Unblocks

Once a release with this fix is cut, Rosenbrock23(autodiff = AutoEnzyme(mode = set_runtime_activity(Enzyme.Forward))) (and the
same for other Rosenbrock / W-methods) works on any in-place ODE RHS
routed through DiffEqBase's AutoSpecialize / wrapfun_iip.

Test plan

• FWW test suite passes for all the existing Enzyme tests.
• New test set passes end-to-end.
• CI green.

🤖 Generated with Claude Code