[ChrisRackauckas-Claude]

Preserve the TrustRegion cache's existing raw/despecialized parameter representation during reinit!. Enzyme setup unwraps the inner problem, so an outer AutoDespecialize cache can otherwise pass wrapped parameters into a raw inner cache and fail conversion. The two-line fix normalizes the parameter used by both the cache and its reinitialization calculations.

The regression test exercises public reinit! with explicit ForwardDiff and Enzyme choices for all three AD backends. The Jacobian-cache prerequisite is already merged: #1227 . This PR contains only the TrustRegion follow-up.

Verification

The exact new regression on unfixed code passed the ForwardDiff case (2/2) and errored in the Enzyme case (exit 1):

MethodError: Cannot `convert` an object of type SciMLBase.DespecializedParameters to an object of type Float64
AutoDespecialize trust region reinitialization | Error  Total
                                              |     1      1

With the fix, the same regression passed both cases on Julia 1.12.7 and Julia 1.10.11 (exit 0 on each):

AutoDespecialize trust region reinitialization | Pass  Total
                                              |    2      2
AutoDespecialize trust region reinitialization | Pass  Total
                                              |    2      2

From lib/NonlinearSolveFirstOrder, the complete native group ran with:

GROUP=Core JULIA_NUM_THREADS=1 JULIA_NUM_PRECOMPILE_TASKS=1 \
  julia --startup-file=no --heap-size-hint=8G --project=. \
  -e 'using Pkg; Pkg.test(; julia_args=["--heap-size-hint=8G"])'
GROUP=QA julia --project -e 'using Pkg; Pkg.test()'

Core: 2,724 assertions across 40 reported testsets, exit 0, including TrustRegion 1,176/1,176 and reinitialization 11/11. QA: 20/20, exit 0. Final output:

SciMLOperator Jacobians | 14  14  19.9s
Testing NonlinearSolveFirstOrder tests passed

Runic check, typos on both changed files, and git diff --check passed. Rebase onto the merged prerequisite produced an identical tree to the locally validated commit.

The Catalyst documentation-load-order reproducer failed before this fix with the matching parameter conversion error and passed afterward, printing CATALYST_LOADED_TRUST_REGION_SUCCESS. The complete composed Catalyst documentation build is still running; no full documentation success is claimed. This change touches no public API, docstring, or documentation file. GPU and full downstream suites were not run locally.

The first native Core attempt ended with SIGKILL of unknown cause. The complete successful retry used the GC heap-size hint above, changed no assertions, and observed unchanged OOM counters throughout.

An executable adjacent-commit comparison identified the original introduction: e416641 fails, while its immediate parent c8479f4 passes with the same external dependency versions.

Please ignore this draft until reviewed by @ChrisRackauckas.

🤖 Generated with Codex CLI 0.153.4 (model: gpt-6-astra), local session ID 01a0598f-11b9-72d1-91d9-b2fbb186557d.

[ChrisRackauckas-Claude]

CI is now complete for b87463a66ecf518edca8e667af1cf5061eb4922e: 72 successful checks, 3 skipped, no pending or failed checks. The final ModelingToolkit InterfaceII downstream check passed: https://github.com/SciML/NonlinearSolve.jl/actions/runs/33964662587/job/101302553012.

This is separate from the additional composed Catalyst docs validation, which is still running locally and is not claimed to pass. The draft remains for @ChrisRackauckas review; no merge was performed.

🤖 Generated with Codex CLI 0.153.4 (model: gpt-6-astra; session: local session ID 01a0598f-11b9-72d1-91d9-b2fbb186557d).

[ChrisRackauckas-Claude]

The additional composed Catalyst docs build completed successfully (exit 0), including HTML rendering and nonempty docs/build/index.html and objects.inv. Catalyst source stayed clean at SciML/Catalyst.jl@c01cce6.

Command (Julia 1.12.7):

TMPDIR=/home/crackauc/sandbox/tmp_20260831_204216_89854/.tmp \
JULIA_PKG_PRECOMPILE_AUTO=0 JULIA_DEBUG=Documenter timeout 28800 \
xvfb-run -a julia --project=../.validation/catalyst-docs-pr1574-timeroutputs1-patched \
  --color=no --code-coverage=user docs/make.jl

Final output includes completed HTML rendering, automatic inventory version 16.4.0, and deployment skipped because no deployment environment was detected. Existing large-example/search-index warnings remain. No example error was reported.

This is explicitly composed evidence, not a clean-registry or this-patch-alone claim. Local package sources included this TrustRegion fix; NonlinearSolveBase's initialization correction; SciML/SciMLBase.jl#1574; sumiya11/Groebner.jl#226; pogudingleb/RationalFunctionFields.jl#61; and an unpublished local TimerOutputs closure-capture fix. Groebner and RationalFunctionFields additionally had local TimerOutputs 1 compat allowances. The complete environment is retained locally.

The checked-in Catalyst configuration has doctest = false and warnonly = [:missing_docs, :cross_references]; these settings were unchanged. Therefore this run does not validate doctests or treat those two existing warning categories as errors. The earlier four-hour run timed out and remains recorded; the successful diagnostic run used an eight-hour limit and debug logging, not suppressed checks.

CI on this NonlinearSolve draft previously completed with 72 successful checks and three skipped checks. No merge was performed.

🤖 Generated with Codex CLI 0.153.4 (model: gpt-6-astra; session: local session ID 01a0598f-11b9-72d1-91d9-b2fbb186557d).