refactor: propagate shared config via before validator

[mikasenghaas]

Summary

Reorganize all shared-config propagation on RLConfig into a single mode="before" validator (auto_setup_shared_configs), so sub-config validators see final values at construction time.

• Covers model, log, ckpt, wandb, tokenizer, seq_len, max_steps, max_async_level, output_dir
• Collapses eight per-field validate_shared_* after-validators into one validate_shared_configs after-validator
• Removes the fix-up hack in auto_setup_tokenizer that had to correct state written by sub-config validators running before model propagation
• Moves auto_setup_session_headers onto OrchestratorConfig (it only touches orchestrator state, no cross-sub-config logic)
• Deletes RLConfig.max_model_len — declared as a shared field but never read or propagated anywhere

Why

Pydantic constructs nested models before running parent validators. Under the old design, auto_setup_model ran after ModelConfig had already been constructed with default values, so any sub-config validator depending on the final shared value (e.g. parser auto-resolution) saw the wrong input. Moving propagation into a mode=\"before\" validator lets sub-configs see the final shared values on their first validation pass.

This also unblocks parser auto-resolution at config time (PR #2290).

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes Pydantic validation order for RLConfig by moving shared-field propagation to a mode="before" validator, which can subtly alter how defaults/CLI/TOML merges resolve and what sub-config validators see. Moderate risk of breaking existing config edge cases despite added unit coverage.

Overview
Refactors RLConfig shared config propagation to run in a single mode="before" validator (auto_setup_shared_configs), so model/log/ckpt/wandb/tokenizer/seq_len/max_steps/max_async_level/output_dir are filled into trainer/orchestrator/inference before nested models are constructed (with sub-config values taking precedence).

Collapses multiple per-field after-validators into one validate_shared_configs after-validator, extracts the seq_len consistency check into validate_shared_seq_len, moves X-Session-ID header auto-setup onto OrchestratorConfig, removes unused RLConfig.max_model_len, and adds focused unit tests covering propagation/precedence and CLI merge behavior.

^{Reviewed by Cursor Bugbot for commit 1077bcc. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

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

^{Reviewed by Cursor Bugbot for commit a61e1d4. Configure here.}

[mikasenghaas]

@samsja this is the only ugliness in the pr atm which i think we can/should fix at the tyro/pydantic config level. imo we should try to avoid resolving any config more than once