fix(codex): support exec optimizer backend

[TheGameVIX]

Summary

• allow codex_exec as an optimizer backend
• make --backend codex route both optimizer and target through Codex exec
• route optimizer calls through skillopt.model.codex_backend
• update backend config docs and add regression tests

Why

The Codex integration already supports running the target through codex exec, but selecting --backend codex left the optimizer on openai_chat. Users trying to run SkillOpt fully through Codex still had an API-backed optimizer path.

This keeps the existing backend split model and only adds codex_exec to the optimizer side. Claude Code remains target-only.

Validation

• python -m pytest tests/test_codex_optimizer_backend.py -q
• python -m mkdocs build --strict
• python -c "import sys; sys.argv=['train.py','--config','configs/searchqa/default.yaml','--backend','codex']; from scripts.train import parse_args, load_config; cfg=load_config(parse_args()); print(cfg['optimizer_backend'], cfg['target_backend'])" prints codex_exec codex_exec

Note: full python -m pytest -q was also attempted on Windows, but current main has unrelated path-separator assertions in ALFWorld/Sleep tests. This PR intentionally leaves those Windows-support fixes out to keep the change scoped to Codex optimizer support.

[TheGameVIX]

@microsoft-github-policy-service agree

[Yif-Yang]

Thanks for adding Codex as an optimizer backend — this is useful functionality and we would like to preserve the merge under your original PR and commit authorship.

The branch now conflicts with backend registry/configuration work merged after it. Could you please rebase onto the latest main? During conflict resolution, please preserve the existing Azure/OpenAI, Claude, Qwen, and MiniMax paths while adding codex_exec to the optimizer path; do not replace the newer backend sets in skillopt/model/__init__.py or backend_config.py.

We tested a union-style resolution locally and obtained:

pytest tests/test_codex_optimizer_backend.py -q  # 12 passed
pytest tests/ -q                                # full suite passed

Please also keep the target and optimizer routing distinct and verify token accounting is not duplicated. Once rebased and green, we can re-review and merge this through #112 so your contributor record remains intact.

[TheGameVIX]

Rebased onto latest main (57333f3) and force-pushed 0b94f89.

Conflict resolution preserved the newer Azure/OpenAI, Claude, Qwen, MiniMax, and openai_compatible paths, then added codex_exec as the optimizer path. Target and optimizer routing remain distinct, and token summary tests cover Codex + existing backend aggregation without duplicate counting.

Validation:

python -m pytest tests/test_codex_optimizer_backend.py -q
# 12 passed

python -m mkdocs build --strict
# passed

python -c "import sys; sys.argv=['train.py','--config','configs/searchqa/default.yaml','--backend','codex']; from scripts.train import parse_args, load_config; cfg=load_config(parse_args()); print(cfg['optimizer_backend'], cfg['target_backend'])"
# codex_exec codex_exec

Current PR branch is merge-clean against main.