[faresobeid]

Stacked on #3861 (refactor/trainer-vlm: fold multimodal into trainer/vlm.py; one VLM registry). That PR removes the last readers of these fields. Merge it first.

Summary

Breaking config change.

• [model.vlm] loses vision_encoder_attr and language_model_attr. Only freeze_vision_encoder is left. The trainer now finds the modules by convention (model.visual, and model.language_model with a fallback to model.model). The only trainable VLM families (Qwen3.5 dense and MoE) already used exactly these paths.
• orchestrator.model.vlm is removed. Nothing in the orchestrator read it, because images are detected from message content parts. vlm moves from BaseModelConfig to the trainer ModelConfig. The shared RL [model.vlm] now propagates only to trainer.model.vlm.
• Updated to match:
  • examples/extra/vlm/{sft,sft-moe}.toml and configs/ci/nightly/multimodal_color_codeword.toml: the color-codeword config now has a bare [model.vlm].
  • docs/advanced.md: the family table loses the attr columns, and the wrong "prefix derived from language_model_attr" sentence is gone.
  • tests/unit/test_configs.py.
• Removed test_orchestrator_vlm_requires_renderer. test_orchestrator_renderer_none_rejected already covers the renderer=None rejection, and the VLM part no longer exists.

Migration: delete vision_encoder_attr / language_model_attr lines and any [orchestrator.model.vlm] block. Old keys now fail with "Extra inputs are not permitted". No compatibility shim, by design.

Behavior change

Config-schema only, as described above. Resolved configs were dumped with model_dump(mode="json") for every TOML under configs/, examples/, k8s/ (91 entries, with overlays composed on sibling bases). Every difference is a removed key:

• orchestrator.model.vlm (always null)
• the two attr strings in the 3 VLM TOMLs

No value changed.

Checks

• pytest tests/unit/test_configs.py tests/unit/orchestrator tests/unit/train/test_multimodal.py -m "not gpu": 292 passed
• ruff check, ruff format --check: clean
• Old keys and [orchestrator.model.vlm] are rejected at parse time (checked by hand).

GPU checks (not run): the nightly configs/ci/nightly/multimodal_color_codeword.toml RL run, and examples/extra/vlm/sft.toml for a few steps. These are the same model paths that the base PR already validates on GPU.

Conflicts

Text-only conflicts:

• refactor(configs): single source of truth for RLConfig shared fields #3774 (utils/validation.py: the model.vlm propagation)
• refactor: orchestrator as hook-wired components with per-component configs #3652 (configs/shared.py)
• chore: remove Kubernetes deployment path (Helm chart, trainer metrics server) #3767 (shared.py, trainer.py, test_configs.py)
• chore: remove legacy config options #3770 (trainer.py, validation.py, test_configs.py)
• refactor(configs): one trainer-common config base for TrainerConfig and SFTConfig #3849 (trainer config base: trainer.py)

No logical conflicts.