chore: rename deprecated orchestrator config keys

[mikasenghaas]

Summary

• Rename [orchestrator.sampling] → [orchestrator.train.sampling] and [[orchestrator.env]] → [[orchestrator.train.env]] across all configs
• Rename max_tokens → max_completion_tokens in sampling sections
• Drops reliance on the deprecated auto-translation emitted by the orchestrator config validator

Validation

• Ran uv run rl @ <config> --dry-run on all 38 modified RL configs — no deprecation warnings
• Validated the two orchestrator-only partial configs (configs/debug/orch.toml, configs/ci/integration/rl_multi_run/orchestrator.toml) via direct OrchestratorConfig.model_validate — no deprecation warnings

🤖 Generated with Claude Code

---

Note

Low Risk
Low risk: this is a mechanical rename of TOML config keys/fields to match the current orchestrator schema, with no functional code changes. Main risk is mis-typed keys causing configs to be ignored or validation to fail at runtime.

Overview
Updates training configs across configs/ and examples/ to stop using deprecated orchestrator keys.

Specifically renames [orchestrator.sampling] to [orchestrator.train.sampling], [[orchestrator.env]] to [[orchestrator.train.env]] (and similarly for orchestrator-only partial configs), and replaces max_tokens with max_completion_tokens in sampling sections.

^{Reviewed by Cursor Bugbot for commit 6ce2707. Bugbot is set up for automated code reviews on this repo. Configure here.}