migrate config system to pydantic_config

[samsja]

Summary

• Replace the custom pydantic-settings wrapper (BaseSettings, parse_argv, toml_files inheritance) with the external pydantic_config library backed by tyro
• Delete src/prime_rl/utils/pydantic_config.py (302 lines) and replace with a thin 50-line wrapper in src/prime_rl/utils/config.py
• Replace allow_extras pattern in inference server with a vllm_extra: dict field on InferenceConfig
• Remove toml_files inheritance from example SLURM configs (use @ base.toml @ slurm.toml composition instead)

Breaking Changes

Boolean negation flags changed location

The --no- prefix now attaches to the field name, not the full dotted path:

# Before
uv run trainer @ config.toml --no-model.compile

# After
uv run trainer @ config.toml --model.no-compile

vLLM extra arguments use vllm_extra dict instead of passthrough CLI args

The old system captured unknown CLI args and forwarded them to vLLM. The new system uses an explicit vllm_extra field:

# Before (unknown args passed through)
uv run inference @ config.toml --some-vllm-arg value

# After (use vllm_extra in TOML)

# In config TOML
vllm_extra = { some_vllm_arg = "value" }

Test plan

• uv sync --all-extras resolves successfully
• uv run pytest tests/unit/test_configs.py -v — all 58 tests pass
• uv run python -c "from prime_rl.utils.config import BaseConfig, cli" — imports work
• No remaining imports of prime_rl.utils.pydantic_config in codebase
• No remaining imports of pydantic_settings in codebase

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

New vllm_extra field leaks into vLLM namespace

Medium Severity

The new vllm_extra field gets included in to_vllm() via the get_all_fields(self) loop, setting namespace.vllm_extra = {} (the raw dict) as a vLLM namespace attribute. This is unintended — the individual entries from the dict are correctly applied in server(), but the raw dict itself persists on the namespace. Since vllm_extra is not a recognized vLLM argument, this spurious attribute could cause issues if vLLM validates namespace attributes or uses vars(args) downstream.

Additional Locations (1)

• src/prime_rl/inference/vllm/server.py#L320-L324