Fix false convergence when LBFGSB encounters Inf/NaN at bounds (issue #1094)

[ChrisRackauckas-Claude]

Summary

• Fixes OptimizationLBFGSB.LBFGSB() produces wrong results with retcode: Success #1094 where OptimizationLBFGSB.LBFGSB() would return Success with incorrect results when the objective function has a singularity at the boundary (e.g., log(0))
• The L-BFGS-B algorithm evaluates function/gradient at bounds during its Cauchy step, which can produce Inf/NaN values that corrupt the optimization state
• The fix detects false convergence and returns Failure instead of Success with a helpful warning

Root Cause

The underlying L-BFGS-B Fortran algorithm evaluates the objective and gradient at the exact boundary values during the Cauchy point computation. When a function has a singularity at the boundary (like log(n_exp/n_th) when n_th=0), this produces Inf objective values and NaN gradients. The corrupted values cause the algorithm to incorrectly report convergence.

Changes

1. Track Inf/NaN values: Wrap the loss and gradient functions to detect non-finite values
2. Detect false convergence: When Inf/NaN was encountered AND optimizer claims Success BUT solution hasn't moved from starting point, this indicates false convergence
3. Return Failure with warning: In such cases, override the return code to Failure and emit a helpful warning suggesting users move bounds away from singularities

Test Plan

• Added test case reproducing the issue from OptimizationLBFGSB.LBFGSB() produces wrong results with retcode: Success #1094 (function with log(0) singularity at bound)
• Verify fix returns Failure instead of Success for problematic cases
• Verify normal optimization cases still work correctly
• Run full OptimizationLBFGSB test suite

🤖 Generated with Claude Code