…tions

The Const-return (IIP) reverse rule re-runs `Enzyme.autodiff(Reverse, ...)`
on the unwrapped function during the reverse pass, reading the arguments'
current `.val`s. When a caller mutates those arguments after the forward
call -- e.g. an ODE integrator stepping `u` after every RHS call -- the VJP
was taken about the post-mutation state, giving a silently wrong gradient.
In the full nested EnsembleProblem adjoint this surfaced as the Julia-1.11
`_dispatch_ensemble_solve` GC-root-rewrite segfault (SciMLSensitivity.jl#1424).

`augmented_primal` now snapshots the call-time argument values into the tape,
and `reverse` temporarily restores them before recomputing the VJP (then puts
the live values back so the surrounding reverse pass is undisturbed).

Adds regression tests: a single wrapped IIP call with the argument mutated
afterwards, and an 8-step integrator loop, both checked against analytic /
finite-difference references.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BZwV7CBdLPqVpJrvRAWcv6

…ation

The snapshot/restore safety net only needs to copy arguments that are
actually mutated between the forward and reverse passes. Gate it on
`EnzymeRules.overwritten(config)` (a conservative over-approximation, so it
can't miss a needed snapshot): a non-mutating call now copies nothing, and an
integrator-style call copies only the overwritten arrays (`du`/`u`), not `p`
or the scalar `t`.

Also fixes two pre-existing hand-built `RevConfig`s in the tests whose
`Overwritten` tuples were one entry short (it is indexed `(func, args...)`).

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BZwV7CBdLPqVpJrvRAWcv6

The Const-return reverse rule returned `ntuple(i -> args[i] isa Active ?
zero(...) : nothing)`, which (1) was union-typed — Enzyme rejects a
`Tuple{Union{Nothing,Float64},…}` return with `ReverseRuleReturnError`, it
requires exact per-slot types — and (2) zeroed every Active arg's gradient.
This broke any wrapped call with a mix of Duplicated and Active args, e.g. a
time-dependent in-place rhs `f!(du, u, p, t)` where `t` flows in as Active.

`Enzyme.autodiff(Reverse, Const(f_orig), Const, args...)[1]` already returns
the per-argument gradient tuple in the exact form a reverse rule must hand back
(`nothing` for Const/Duplicated, the real gradient for Active). Forward it,
rebuilt through `map(_revslot, args, raw)` so dispatching on each argument's
concrete annotation type gives an exactly-typed result (the raw tuple is
`Any`-typed inside the rule). Adds a mixed Duplicated/Active regression test.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BZwV7CBdLPqVpJrvRAWcv6