verl backend dispatches on method name and silently drops most TrainConfig fields

[thedataengineer]

CLAUDE.md states the core invariant: "Backends dispatch on the spec's fields, never on the method name." The verl backend breaks it, and also breaks the "ignored with a log line" convention for TrainConfig.

1. Name dispatch. backends/verl.py:142 does if config.method != "grpo": raise. It should check spec.trainer != "grpo". As written, a user-registered method with trainer="grpo" (the documented methods.register(...) extension point) is rejected by verl even though it is a valid GRPO method.

2. Silent TrainConfig drops. verl.py:43-57 maps roughly 7 of ~35 fields (grpo_max_completion_length, max_seq_length, learning_rate, beta, grpo_group_size, derived total). Everything else is discarded with no log line: weight_decay, max_grad_norm, lr_scheduler_type, optim, warmup fields, batch/accumulation, seed, save_steps/eval_steps/logging_steps, report_to, packing, use_rslora, resume_from_checkpoint, and the whole LoRA family (verl GRPO is full-parameter). The SHADOWLM_VERL_OVERRIDES escape hatch at verl.py:64 is undiscoverable from a run log. The convention says each ignored field gets a log line; mlx already implements the pattern at backends/mlx.py:349-358.

3. Stub load, missing surface. verl.py:133-137 load() only records the model name and silently discards load_in_4bit and max_seq_length. generate() raises (verl.py:198, with an actionable message), but chat() and save() fall through to the generic backends/base.py:96,99 raises with no actionable message, so backend="verl" cannot round-trip the SDK surface at all.

4. Minor: verl.py:50 formats config.learning_rate:g, which is a TypeError when None. Both live paths resolve the method default first (models.py:196, serve.py:148), but a direct Backend.finetune() call hits it.

[thedataengineer]

Fixed in 1aca641: spec.trainer dispatch, _ignored_fields logging for the ~20 unmapped TrainConfig fields, load() noting load_in_4bit/adapter, chat()/save() pointing at backend='torch', and the None learning_rate TypeError. 9 new tests in tests/test_verl_contract.py, no verl install needed.