fix: forward all top-level config fields to worker processes (#516)

[SUPR1M3]

Fixes #516

Problem:

ProcessParallelController._serialize_config built the dict sent to worker processes from a hand-written list of top-level Config fields. _worker_init rebuilds the worker's config with Config(**dict), so any field missing from that list silently fell back to its default inside workers.

Two of the missing fields are read only in workers, so setting them in the config had no effect during a run:

• enforce_evolve_blocks: true: edits outside EVOLVE-BLOCK-START/END were kept and evaluated
• custom diff_pattern: replies were parsed with the default regex, so every iteration failed with "No valid diffs found in response"

The existing worker tests patch _worker_config directly, so the serialize → _worker_init round trip was never exercised.

What:

• Build the top-level part of the worker dict from dataclasses.fields(Config) instead of a hand-written list, so new fields reach workers automatically. The fields come from the Config class rather than the instance, so a user-defined Config subclass with extra fields still initialises workers.
• Send evolution_trace as a dict and rebuild it in _worker_init, like prompt, database and evaluator.
• Add tests/test_worker_config_serialization.py: round-trips a config with a non-default value for every top-level field through _serialize_config → _worker_init and asserts each one matches. It also fails if a future field is added without a non-default value in the test.

The other six previously dropped fields (logger_prefix, evolution_trace, early_stopping_*, max_tasks_per_child) are only read in the main process, so forwarding them changes nothing in workers.

Red -> green evidence:

On main, the new test fails on exactly the eight dropped fields:

FAIL: test_all_top_level_fields_round_trip (field='enforce_evolve_blocks')  False != True
FAIL: test_all_top_level_fields_round_trip (field='diff_pattern')  '<<<<<<< SEARCH...' != '<<<< SEARCH...'
FAIL: ... (field='logger_prefix', 'evolution_trace', 'early_stopping_patience',
           'convergence_threshold', 'early_stopping_metric', 'max_tasks_per_child')
FAIL: test_diff_pattern_reaches_worker
FAIL: test_enforce_evolve_blocks_reaches_worker
FAILED (failures=10)

With this PR: Ran 3 tests ... OK.

Testing:

• Unit tests: 667 pass on Python 3.10 and 3.13 (664 on main plus the 3 new tests); per-test results are otherwise identical to main
• End to end with a real process pool, both fork and spawn, 3 workers:
  • enforce_evolve_blocks: true: outside edits kept in 6/6 children on main, 0/6 with this PR
  • custom diff_pattern: 6/6 iterations fail on main, 0/6 with this PR
  • both settings at their defaults: identical to main
• CLI with a YAML config, including checkpoint and resume: settings take effect and persist after resume
• Integration suite against a local OpenAI-compatible stub: same results as main
• black 25.12.0 / isort 7.0.0 (pre-commit versions) clean; Frame scan reports no findings

Notes:

Open PR #480 adds these two keys to _serialize_config as part of a much larger feature; this PR is a standalone fix with a regression test.

[CLAassistant]

All committers have signed the CLA.