Route combinatorial ACO/GA parallel dispatch through GenerationRunner

[fundthmcalculus]

Summary

• combinatorial/{aco,aco_mst,ga}.py each hand-rolled their own spawn_streams/use_stream/joblib.delayed dispatch, duplicating what the continuous solvers already factored into core/parallel.GenerationRunner (the copy-fixed-data-once path from the perf work) — they just never adopted it.
• Each solver now builds a fixed tuple of its run-constant data (distance matrix, desirability/config, individuals_per_job) and calls GenerationRunner.run() with only the per-generation varying argument (tau_alpha for the two ACO variants, the genome + its values for GA) instead of re-pickling everything to workers every generation.
• Each per-individual worker (run_ant/run_ant_mst/run_ga) is wrapped in a new module-level _run_*_batch helper so it stays picklable for the "processes" backend.
• Purely a dispatch-mechanism change — the per-individual logic itself is untouched, and GenerationRunner provides the same per-task RNG-stream isolation the old inline spawn_streams/use_stream did.

Test plan

• New 9-test determinism sweep across all three solvers (aco-tsp, aco-mst, ga-tsp) at n_jobs in {1, 3} in tests/test_determinism.py — same seed reproduces identically, different seeds diverge
• Manual smoke test confirming the "processes" backend still pickles cleanly for all three solvers
• flake8 ./src ./tests — clean
• MYPYPATH=src mypy -p optimizers — clean
• black --check . — clean
• Full suite passing (185 tests)

🤖 Generated with Claude Code

https://claude.ai/code/session_011r9W1JZRSoayjyLin4ZBPZ