feat: add optional logger_prefix config and --logger-prefix CLI flag

[akushonkamen]

Summary

Add an optional logger prefix so log output from concurrent OpenEvolve runs can be told apart.

• New Config.logger_prefix: Optional[str] — set it directly when using the library: OpenEvolve(..., config=Config(logger_prefix="run1")) or openevolve.run_evolution(..., config=Config(logger_prefix="run1")), addressing the "optional prefix" request in Feature Request: Logger prefix #290 (delivered via Config + CLI rather than a run() kwarg, as planned on the issue).
• New --logger-prefix CLI flag, applied to the config before the controller is constructed.
• New LoggerPrefixFilter (openevolve/utils/logging_utils.py): a handler-level logging.Filter installed by _setup_logging() that rewrites record.name (e.g. openevolve.controller → run1.openevolve.controller). Since all module loggers are created at import time via logging.getLogger(__name__), rewriting record.name at the handler level is the one clean point that covers all of them without touching each module.
• The filter is idempotent: the same LogRecord object flows through every handler on the logger, so the prefix is applied exactly once.
• When a prefix is active, the console formatter also gains %(name)s (the file formatter already had it). Without a prefix the output format is byte-for-byte unchanged.

Example:

2026-10-10 23:02:00 - run1.openevolve.controller - INFO - Logging to .../logs/openevolve_20261010_230200.log

Motivation

Fixes #290. As @theahura reported, running multiple openevolve.run() calls in parallel makes it extremely hard to tell which logs belong to which run. @codelion endorsed the idea on 2025-10-13: "This is a good idea we can add a logger prefix." This PR lands exactly that, with the smallest footprint possible.

Scope notes

• Worker-process logging is intentionally out of scope (worker processes don't call _setup_logging); that's the domain of Multiple Openevolve run calls in parallel result in clobbered loggers #295.
• Related open PR Isolate per-run logging from the process root logger #499 isolates per-run handlers via contextvars for Multiple Openevolve run calls in parallel result in clobbered loggers #295; the two changes are complementary. If Isolate per-run logging from the process root logger #499 merges first, this PR will need a small rebase of _setup_logging (the filter itself is independent).
• Within a single process, all records still bubble to the root logger, so a shared root handler setup can't provide strict per-run isolation — this prefix makes concurrent runs' output distinguishable and greppable. Full isolation would need per-run logger hierarchies or a contextvar-based approach (happy to discuss in Multiple Openevolve run calls in parallel result in clobbered loggers #295).

Testing

New fully-offline unittest suite tests/test_logger_prefix.py (7 tests; LLMEnsemble/Evaluator patched, no LLM calls):

• Filter unit tests: rewrites record.name, idempotent under repeated application, empty prefix is a no-op, never swallows records.
• Integration tests: constructing OpenEvolve(config=Config(logger_prefix="run1")) (constructor runs _setup_logging) writes run1.openevolve.controller to both the log file and the console stream; with no prefix, output is unchanged (plain names, no run1.).

Red → green:

• Before implementation: ModuleNotFoundError: No module named 'openevolve.utils.logging_utils' (1 test, FAILED).
• After implementation: Ran 7 tests ... OK.

Full regression (CONTRIBUTING command):

export OPENAI_API_KEY=test-key-for-unit-tests
python -m unittest discover tests
...
Ran 629 tests in 69.690s
OK

Formatting: black --line-length 100 and isort --profile black --line-length 100 applied to all changed files.