Add SubprocessEvaluator for process-isolated evaluation - #77

Thanks for this - a subprocess isolation tier between the in-process and containerized evaluators is a real gap, and the wiring through create_evaluator is clean. Before merge, a few items where the subprocess path silently diverges from the in-process Evaluator it is documented to mirror. I reproduced these against the branch.

1. Timeout leaks orphaned processes (high). The child is created without start_new_session=True, and the timeout path calls proc.kill(), which SIGKILLs only the python -c wrapper. Any process the candidate's evaluate() forks (torch DataLoader workers, multiprocessing pools, NCCL) survives and reparents to PID 1. Repro on this branch: a timed-out eval returned {'error': 0.0, 'timeout': True} while its grandchild stayed alive with PPID 1, still holding its resources. This is the exact leak the isolation feature exists to prevent. Fix: pass start_new_session=True and on timeout os.killpg(os.getpgid(proc.pid), signal.SIGKILL) (guard ProcessLookupError) then await proc.wait().

2. cascade_evaluation is ignored (high). EvaluatorConfig.cascade_evaluation defaults to True, and the in-process Evaluator runs evaluate_stage1 -> threshold -> evaluate_stage2 -> merge. The wrapper template only calls mod.evaluate(...), so combined_score silently differs under isolation for the bundled cascade benchmarks (circle_packing, signal_processing, txn_scheduling, gpu_mode/shared_eval). Repro: same program scored 0.9 in-process vs 0.1 under subprocess. Please replicate the stage logic in the child (or share one implementation), or fail loudly on cascade configs rather than downgrading silently.

3. EvaluationResult.artifacts are dropped on the round-trip (high). to_dict nests artifacts under metrics['artifacts'], but from_dict assigns the whole dict to metrics and leaves artifacts={}. Downstream feedback loops (adaevolve/controller.py, gepa_native/controller.py reading artifacts.get('feedback')) go silently empty under isolation. Fix: make from_dict pop 'artifacts' back out so to_dict/from_dict are symmetric.

4. Forced cwd (medium). The subprocess passes cwd=dirname(evaluation_file); the in-process Evaluator keeps the run cwd. Evals that read datasets by cwd-relative path silently fail after toggling the flag. Inherit the parent cwd, or make both paths identical and document it.

5. Image mode broken (medium). is_image_mode is stored but the <temp>.image_path sidecar that the in-process Evaluator writes (and that benchmarks/image_gen/sky_festival reads) is never created, so every image-mode candidate scores 0.0 under isolation. Write and unlink the sidecar mirroring evaluator.py.

6. Fragile JSON extraction (medium). stdout.rfind('\\n{') only tolerates preceding stdout noise that ends in a newline; a non-newline-terminated write immediately before the result (progress bars, print(end='')) turns a valid score into {'error': 0.0}. Prefer a sentinel-delimited result line from the wrapper, or scan for the last balanced JSON object.

Tests. The timeout assertion timeout is True or error == 0.0 is tautological since every failure path sets error=0.0 - it passes even if the timeout mechanism breaks; use assert result.metrics.get('timeout') is True. Also evaluate_batch and concurrent isolation have no coverage. The highest-leverage addition would be a parity test that runs one benchmark through both Evaluator and SubprocessEvaluator and diffs the result - that single test covers items 2, 3, 4, and 5.

Nice work overall - none of these are structural, the design is sound, and it is complementary to the existing evaluators rather than redundant. Formatting (black/isort) and the current suite (164 tests) are green. Happy to re-review once the parity gaps are closed.