Zygote autodiff fails on remake tests with DifferentiationInterface

[ChrisRackauckas-Claude]

Summary

When using DifferentiationInterface.gradient with AutoZygote() backend on remake-based loss functions (using SciMLSensitivity with ZygoteVJP()), the tests fail.

Context

This was discovered during the migration from direct AD backend usage to DifferentiationInterface in PR #1206. The following test patterns fail with Zygote:

• Basic remake autodiff: gradient of sum of solution w.r.t. u0 and p
• Symbolic indexing autodiff: gradient of sum(soln[x]) w.r.t. u0 and p
• Symbolic indexing observed autodiff: gradient of sum(soln[o, i]) w.r.t. u0 and p

Affected Files

• test/downstream/remake_autodiff.jl

Current Status

Tests are marked as @test_broken in PR #1206 until this is resolved.

Related

These tests use BacksolveAdjoint(autojacvec = ZygoteVJP()) from SciMLSensitivity, so this may be related to how DifferentiationInterface interacts with SciMLSensitivity's Zygote extension.

[ChrisRackauckas]

This is likely not fixable with MTK v11.