Fix differential evolution, brute force, and out-of-range steps

[barrybecker4]

Summary

Evaluation of every OptimizationStrategyType on the eight OptimizerEvalApp problems (seeds 1, 42, 99) found several strategies that do not search. This draft fixes the clear, low-risk defects and leaves the write-up in EVALUATION.md (before/after numbers included).

• Differential evolution called setValue and discarded the new Parameter, so every trial was the target. Sphere fitness stayed around 4.6–18 after 40,041 evaluations (~22 unique). It now builds a real trial vector, samples uniformly, and reaches ~1e-158 on the sphere and 0 on Rastrigin and Rosenbrock.
• Brute force requested findGlobalSamples(Long.MaxValue). The 32-bit grid product wraps to one sample in 2D (sphere fitness 52.4, worse than the start) and to 2.1e9 points in 1D (does not finish). Integer lattices are enumerated directly; continuous spaces throw.
• Integer sampling and epsilon steps could leave [min, max] or assert. That made global hill climbing and global simulated annealing throw on the discrete grid and the 1D line. Both now finish at fitness 0.
• Hill-climbing cache retries grew the neighbor radius past the |r| <= 1.5 assert.
• Genetic duplicate checks compared toCSVString under a global format lock, so DominatingSetProblemSuite did not finish. Value equality brings that suite down to about 4s.
• Optimizer ignored the caller Random for differential evolution, particle swarm, and iterated local search.
• A cancelled SwingWorker could overwrite a finished viewer run with "Cancelled."

./gradlew test: 288 tests, 0 failures. ./gradlew slowTest was still running at 90s and is not part of that count.

Test plan

• ./gradlew test (288 tests, 0 failures), including DominatingSetProblemSuite
• Headless strategy × problem × seed table, before and after (EVALUATION.md)
• OptimizerEvalApp screenshots for all 13 strategies × 8 problems, plus after-fix frames for the repaired cases
• ./gradlew slowTest (did not finish within 90s)