fix: preserve Qwen3.5 broadcast weight names

[samsja]

What changed

• Preserves Qwen3.5 HF hub weight names during live broadcast and checkpoint export.
• Routes NCCL/filesystem broadcast and weight checkpoint export through a shared helper.
• Updates the training monitor skill with the Qwen3.5 skipped-weight failure mode.

Why

Qwen3.5 VLM checkpoints already use model.language_model... HF hub naming. Calling transformers.core_model_loading.revert_weight_conversion on those keys rewrites them into the wrong namespace for vLLM live reload. In the failing runs, vLLM logged skipped keys like language_model.language_model..., so inference was not receiving updated LM/linear-attention weights after trainer updates.

With the bypass, raw HF keys are sent for model_type = "qwen3_5"; vLLM maps them to its internal language_model.model... keys and the weight updates load.

Validation

• uv sync --all-extras
• uv run ruff check src/prime_rl/trainer/weights.py src/prime_rl/trainer/ckpt.py src/prime_rl/trainer/rl/broadcast/nccl.py src/prime_rl/trainer/rl/broadcast/filesystem.py src/prime_rl/trainer/models/qwen3_5_moe/modeling_qwen3_5_moe.py
• Slurm run 23166 reached trainer step 9 with Mismatch KL between 0.0005 and 0.0010; old skipped-weight logs were gone.

W&B: https://wandb.ai/primeintellect/wordle/runs/6f6cfafdf0274166ad038e7e79375f29

---

Note

Medium Risk
Transformers version and weight-export/broadcast behavior affect Qwen3.5 inference sync; dependency pin changes the whole training stack’s HF behavior.

Overview
Pins Transformers to 5.6.2 on PyPI (replacing the git pin and >=5.1.0.dev0 override) so training, checkpoints, and vLLM live reload share a single release aligned with Qwen3.5 fixes.

In the diff, broadcast/checkpoint paths still call revert_weight_conversion for non–PrimeRL models; only redundant inline comments were removed in ckpt.py and filesystem.py. Qwen3.5 workaround docstrings in model.py now say they can drop once an official Transformers release includes the fixes (not a specific git commit).

The start-run skill adds steps to verify verifier env packages import under uv run and how to wire missing local envs into pyproject.toml before rl launches.

^{Reviewed by Cursor Bugbot for commit 1143117. Bugbot is set up for automated code reviews on this repo. Configure here.}

[S1ro1]

is this intended?

[S1ro1]

lgtm boss