[barrybecker4]

Phase A of the new strategies plan: a constraint benchmark

This branch includes the plan (#36), so merging this makes #36 unnecessary.

What it adds

• Three constraint problems in scala-test/.../benchmark, each with a known solution and fitness 0 as the goal:
  • NQueensProblem: a permutation, with an allocation free swap delta in evaluateMove.
  • GraphColoringProblem: integer variables, built around a hidden coloring so a solution exists.
  • AllIntervalProblem: a permutation with many plateaus; the check that a strategy is not tuned to queens.
• Solve rate reporting: ./gradlew benchmark -Pargs="--solve-rate=true" runs them with 20 seeds and 2 s each, stops a run at its first solution, and writes docs/solve-rate-results.md (solved runs of 20, median time and evaluations of the solved ones, median broken constraints of the rest). The ordinary benchmark is unchanged.
• A strategy that rejects a problem when it starts (global annealing and global hill climbing cannot sample 200 integer parameters) is skipped rather than ending the benchmark.
• Tests for each problem (known solutions, delta equals full evaluation for every swap), the limiter, the report, the options and the run.

Baseline (two processors, 20 seeds, 2 s)

Problem	Annealing	Genetic search	Hill climbing, ILS
queens-100 / 400	20/20	0/20	0/20
coloring-200	0/20	0/20	0/20
all-interval-14	1/20	10/20	0 to 5/20

So the benchmark separates the strategies: annealing is already good on swap problems, nothing handles integer variables, and genetic search wins only the all-interval series. Full table in docs/solve-rate-results.md. Notes on those results and other problems worth adding (3-SAT, Costas arrays) are in the plan.

Testing

562 tests pass in my manual build; the 6 failures are the same known ones as before. I could not run Gradle here.

---

Generated by Claude Code