[SUPR1M3]

Fixes #288

Problem:

When every iteration fails before producing a candidate program (LLM endpoint down, deployment missing, model returning empty replies), ProcessParallelController.run_evolution logs each error as a warning and moves on. Nothing tracks whether any iteration succeeded, so the run ends exactly like a healthy one:

• the CLI prints Evolution complete! with the seed's metrics and exits with code 0
• best/best_program.py is written as a byte-identical copy of initial_program.py

Scripted pipelines check the exit code and the final banner, so a run that generated nothing looks like a successful run that simply found nothing better than the seed.

What:

• ProcessParallelController.run_evolution counts iterations whose worker produced a candidate program and records the last iteration error (worker error, timeout, or result-processing exception). At the end of the run it sets produced_no_candidates when iterations completed but none produced a candidate, and logs an error. The flag is computed inside run_evolution rather than as a property, because stop() sets the shutdown event during normal cleanup.
• OpenEvolve.run raises a new EvolutionFailedError (exported from openevolve) in that case, before the best program is saved, so the seed is never presented as an evolved result. The message includes the count and the last error, e.g. Evolution failed: 0/10 iterations produced a candidate program. Last error: LLM generation failed: Error code: 404 - ....
• cli.py catches EvolutionFailedError, prints the message without a traceback and returns exit code 1.
• Behaviour is unchanged for intermittent failures (a single successful iteration is enough for the run to finish normally) and for runs interrupted by a shutdown request, which are not treated as failures. Checkpoints are still written as before.
• Add tests/test_all_iterations_failed.py (6 tests): controller counting (all errors, one success among errors, shutdown), OpenEvolve.run raising and not writing best/, and an end-to-end CLI run against a local HTTP server that answers every LLM request with 404 (the reproduction from the issue).
• Existing tests that replace ProcessParallelController with a bare Mock/MagicMock now set produced_no_candidates = False, the same way they already set shutdown_event (test_checkpoint_resume.py, test_initial_program_artifacts.py, test_island_system_messages.py).

"Candidate produced" is used as the success signal rather than "LLM call succeeded", so a run where the model answers but never returns usable output (e.g. No valid diffs found in response on every iteration) also fails. Nothing was evolved in that case either.

Red -> green evidence:

On main, the new tests fail:

ERROR: test_all_errors_produce_no_candidates      AttributeError: 'ProcessParallelController' object has no attribute 'produced_no_candidates'
ERROR: test_shutdown_is_not_a_failure             AttributeError: (same)
ERROR: test_single_success_counts_as_progress     AttributeError: (same)
FAIL: test_exits_non_zero_without_success_banner  AssertionError: 0 != 1
FAIL: test_raises_and_does_not_save_seed_as_best  AssertionError: EvolutionFailedError not raised
FAILED (failures=2, errors=3)

With this PR: Ran 6 tests ... OK.

Testing:

• Unit tests: 887 pass on Python 3.14 (881 on main plus the 6 new tests); no other per-test differences from main

• End to end with the issue's reproduction: examples/function_minimization with only api_base pointed at a local stub, 10 iterations, default retries, 3 workers:

  Endpoint returns	Branch	Exit code	Evolution complete!	best/
  HTTP 404	main	0	printed	copy of the seed
  HTTP 404	this PR	1	not printed	not written
  HTTP 500	main	0	printed	copy of the seed
  HTTP 500	this PR	1	not printed	not written
  HTTP 200, empty completion	main	0	printed	copy of the seed
  HTTP 200, empty completion	this PR	1	not printed	not written

• black 25.12.0 (pre-commit version) clean on all changed files. isort 7.0.0 is clean on all changed files except tests/test_checkpoint_resume.py, whose import json / import time ordering is already flagged on main; this PR does not touch its imports.