Add run_config passthrough for orchestrator settings

[rasdani]

Summary

• Expose [run_config] TOML section for passing arbitrary orchestrator CLI args to hosted training
• The backend already supports runConfig as a freeform dict converted to CLI args — this just wires it up in the CLI
• Enables settings like use_token_client, max_concurrent, filters, advantage, etc. that aren't first-class CLI fields

Usage

model = "Qwen/Qwen3-4B-Thinking-2507"
batch_size = 256

[run_config]
use_token_client = false
max_concurrent = 512

Test plan

• prime rl run with [run_config] section parses without error
• Verify runConfig appears in API payload
• Verify orchestrator receives the extra CLI args

🤖 Generated with Claude Code

---

Note

Medium Risk
Adds a freeform run_config dict to the RL run config and forwards it to the /rft/runs payload, which can change orchestrator behavior and potentially cause unexpected run failures if misconfigured.

Overview
Adds support for a freeform run_config section in prime rl run configs. The CLI now parses run_config, prints it in the training config preview, and passes it through to RLClient.create_run so it is included as run_config in the /rft/runs request payload.

^{Reviewed by Cursor Bugbot for commit b6d3de4. Bugbot is set up for automated code reviews on this repo. Configure here.}