OptimizationOptimJL: default extended_trace to false

[ChrisRackauckas-Claude]

Summary

__map_optimizer_args always passed extended_trace = true into Optim.Options, so every Optim iteration built a metadata Dict with copies of x, g(x), and (for BFGS) the full n×n ~inv(H) — even though OptimizationOptimJL never reads that trace (callbacks use optim_state fields). This defaults extended_trace to false while still allowing opt-in via solve(...; extended_trace=true) through the existing kwargs passthrough. The docs already say extended_trace "Defaults to false" (docs/src/optimization_packages/optim.md:45); this change makes the code match the docs.

User-visible change: with store_trace=true, each sol.original.trace entry's metadata now keeps only "time" by default. In addition to losing x / g(x) / ~inv(H), BFGS also loses "Current step size", and NelderMead loses "centroid" and "step_type". Passing extended_trace=true restores everything. The printed show_trace=true / verbose=true output changes the same way (e.g. BFGS no longer prints ~inv(H)).

Performance (BenchmarkTools minima, JULIA_NUM_THREADS=1 taskset -c 96)

Rosenbrock + AutoForwardDiff, same machine (amdci2).

Julia 1.12.4

case	before	after
LBFGS n=2	0.09 ms / 0.05 MB	0.08 ms / 0.04 MB
BFGS n=10	0.22 ms / 0.22 MB	0.18 ms / 0.10 MB
BFGS n=100	18.07 ms / 41.98 MB	13.06 ms / 0.61 MB

Julia 1.10.12

case	before	after
LBFGS n=2	0.10 ms / 0.06 MB	0.08 ms / 0.04 MB
BFGS n=10	0.24 ms / 0.23 MB	0.18 ms / 0.10 MB
BFGS n=100	18.46 ms / 41.56 MB	13.87 ms / 0.62 MB

Matches auditor A/B (BFGS n=100: 18.26 ms / 42.0 MB → 13.05 ms / 0.61 MB on 1.12).

Regression test

New @testset "extended_trace defaults to false" in lib/OptimizationOptimJL/test/core_tests.jl:

• Unfixed (git stash of src): 5 pass / 3 fail (metadata still has x/g(x)/~inv(H)).
• Fixed: 8/8 pass. Verified on 1.12.4 and 1.10.12.

Callback parity

Captured OptimizationState (u, objective, grad, loss) for LBFGS / BFGS / Newton / IPNewton over 10 iters: identical master vs this PR.

Test-group tails

• OPTIMIZATION_TEST_GROUP=Core (1.12.4): Core | 6977 Pass / 6977 Total; tests passed
• OPTIMIZATION_TEST_GROUP=Core (1.10.12): Core | 6977 Pass / 6977 Total; tests passed
• OPTIMIZATION_TEST_GROUP=QA (1.12.4): Quality Assurance | 54 Pass, 1 Broken / 55 Total; tests passed (broken is pre-existing ExplicitImports)
• Runic + typos on touched files: clean (files were Runic-clean on master)

Not verified

• Root monorepo GROUP=QA / GROUP=All (never run per policy).
• Downstream packages that may have depended on extended metadata appearing under store_trace=true alone.
• Absolute µs on a busy shared host have run-to-run noise; memory (MB) is the stable signal for BFGS n=100.

Reviewer pushback

• Is anyone relying on store_trace=true implicitly getting extended metadata (x/g(x)/~inv(H), BFGS "Current step size", NelderMead "centroid"/"step_type") or the matching show_trace/verbose lines without extended_trace=true? Docs already said extended_trace defaults to false.
• Should we omit extended_trace from mapped_args entirely (let Optim's default apply) instead of setting false explicitly? Behavior is the same today because kwargs override.

Please ignore until reviewed by @ChrisRackauckas.

Risk assessment

Independent review: MERGE (Claude Code, Opus 5.5) — low risk; callback parity holds; memory 42 → 0.6 MB. Pending @ChrisRackauckas.

---

🤖 Generated with Cursor Agent 2026.10.01-14929f9 (model: unknown (Cursor auto)), transcript /home/crackauc/sandbox/goals/performance/jobs/opt/opt-trace/log.txt on amdci2; orchestrated by Claude Code (claude-opus-5-5[1m]) https://claude.ai/code/session_01LPHREnnonfLg1VcE1EJovv