feat(config): add checkpoint_on_improvement option to save only on new best

[mvanhorn]

Summary

Optimization runs that span many iterations without improvement currently produce checkpoints anyway, since checkpoint_interval fires on a fixed cadence. This PR adds an opt-in checkpoint_on_improvement config flag that triggers the checkpoint callback when a new best program is found, in addition to the existing interval-based trigger.

Why this matters

• Issue: #434 - "Add option to checkpoint only when improvement is detected" (label: enhancement, help wanted). Reporter notes that interval-only checkpointing wastes I/O when slow-progress runs save the same best program over and over.

Changes

• openevolve/config.py:407 adds checkpoint_on_improvement: bool = False next to checkpoint_interval. Default preserves existing behavior.
• openevolve/process_parallel.py:387 wires the new field into the worker config dict so worker processes see it.
• openevolve/process_parallel.py:670 splits the trigger into interval_hit and improvement_hit and fires the checkpoint callback when either is true. Logs distinguish "Checkpoint interval reached" vs "Checkpointing new best solution".
• tests/test_process_parallel.py:156 adds test_checkpoint_on_improvement_only_fires_for_new_best. Sets checkpoint_on_improvement=True and checkpoint_interval=10000 so only the improvement path can fire, then drives one new-best iteration and one not-best iteration. Asserts checkpoint_calls == [1].

Testing

Full test suite passes locally on Python 3.10 via uv:

Ran 371 tests in 66.808s
OK (skipped=1)

Default checkpoint_on_improvement=False keeps existing runs identical; the new path only activates when users opt in.

Fixes #434

AI was used for assistance.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}