Fix silent lcons/ucons length mismatch and undef constraint buffers

[ChrisRackauckas-Claude]

Summary

When cons! writes fewer entries than length(ucons), the DifferentiationInterface out-of-place constraint wrappers allocated Vector{T}(undef, num_cons), leaving garbage in unwritten slots that Ipopt/IPNewton can turn into silent Failure/NaN results. Those wrappers now use zero-initialized buffers (dense and sparse paths). Separately, mismatched lcons/ucons lengths were accepted by SciMLBase and then silently broadcast or crashed opaquely in solvers; _check_opt_alg now throws a clear ArgumentError when both are provided with different lengths (both-nothing and existing one-missing checks unchanged).

Regression evidence

• Before (src stashed, tests present): OptimizationBase Core 61 passed, 6 failed — _check_opt_alg missing lcons/ucons 4 fail; zero-init buffer test failed with J[3,:] == [2.0e-323, 6.9e-310] style garbage. Logs: opt-cons/before.txt.
• After: OptimizationBase Core 67 passed (Testing OptimizationBase tests passed). Logs: opt-cons/after.txt.
• Julia 1.10.12 x64: regression tests (buffer + length mismatch) passed in an isolated temp env.

Test-group tails

• OPTIMIZATION_TEST_GROUP=Core / Pkg.test(OptimizationBase): Test Summary: Core | Pass 67 Total 67 — tests passed
• OPTIMIZATION_TEST_GROUP=QA on OptimizationBase: Test Summary: QA | Pass 20 Broken 1 Total 21 — tests passed
• Root GROUP=QA / Pkg.test(Optimization): Test Summary: QA | Pass 21 Total 21 — tests passed
• Runic --inplace + typos on touched files: clean

What was NOT verified

• Full Ipopt/IPNewton/LBFGSB end-to-end re-run of auditor a3_safety.jl after the fix (unit regressions cover the two root causes).
• OptimizationBase AD group / GROUP=All.
• Whether SciMLBase should eventually own the length check (intentionally left in OptimizationBase per task).
• SymbolicAnalysis Array{Num}(undef, num_cons) path (different type; not the float-garbage AD path).

Reviewer pushback

• Is zero-filling the right semantic when num_cons exceeds what cons! writes (masks a user bug as constraint value 0), or should we also warn/error on write undercount?
• Should length mismatch be rejected at OptimizationProblem construction in SciMLBase instead of only on solve/init?
• Sparse DI path has the same zero-init change but no dedicated sparse regression beyond Core coverage.

Please ignore until reviewed by @ChrisRackauckas.

Risk assessment

Independent review: Claude Code (claude-opus-5-5) MERGE, risk low (confidence medium-high): every valid constraint path (inequality, equality, mixed; MOI-Ipopt, OptimizationIpopt, IPNewton, AugLag, LBFGSB; unconstrained) gives identical results to master; zero-fill has no measurable cost. Deliberate tightening: the lcons/ucons length check now rejects mismatched lengths even when they were previously tolerated (e.g. lcons length 1 with ucons length 2 through MOI-Ipopt, or bounds given without a constraint function); both are malformed inputs and no in-repo usage relies on them.

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

Made with Cursor