[recipe] fix: normalize LangGraph chat template token IDs

[le-czs]

What does this PR do?

Fixes #139 by normalizing every tokenizer.apply_chat_template(..., tokenize=True) result used by the LangGraph recipe before it is sliced, appended to, sent to the rollout server, or stored in response metadata. This makes tokenizers returning a Transformers BatchEncoding follow the same path as tokenizers returning list[int].

This is not a duplicate: no open PR matched issue #139 or the langgraph BatchEncoding chat template keywords when checked on 2026-08-20. Open-PR search

AI assistance disclosure: OpenAI Codex was used to inspect the issue, implement the change, run checks, and draft this PR. The PR remains a draft pending human review of every changed line.

Checklist Before Starting

• Search for similar PRs. Query: https://github.com/verl-project/verl-recipe/pulls?q=is%3Apr+is%3Aopen+langgraph+BatchEncoding
• Format the PR title as [{modules}] {type}: {description}.

Test

• pre-commit run --all-files --show-diff-on-failure --color=always: passed (ruff, ruff-format).
• Model Factory CPU regression command including recipe/langgraph_agent/test_chat_model_on_cpu.py: 8 tests passed in the combined agent-loop regression suite.
• Model Factory Qwen3-4B LangGraph GRPO val_only run on 2 H20 GPUs: platform phase 5, container exit code 0, MATH acc/mean@1=0.996, and no occurrence of BatchEncoding/list TypeError in the full job log.

API and Usage Example

No public API or configuration changes.

Design & Code Changes

• Reuse the core verl.utils.normalize_token_ids helper already provided by the verl commit pinned in langgraph_agent/REQUIRED_VERL.txt.
• Normalize initial prompt IDs, the bound system-prompt IDs, and Hermes tool-response IDs at their chat-template boundaries.
• Add CPU regression tests using a tokenizer that returns BatchEncoding for all three paths.

Checklist Before Submitting

• Read the contribution guide and repository agent instructions available in the parent verl repository.
• Apply pre-commit checks.
• Add / Update documentation. Not needed because this fixes internal token normalization without changing usage.
• Add the test to a CI workflow. The repository currently runs pre-commit for this path but has no LangGraph CPU-test workflow; the focused test is included for local and parent-repository execution.
• Request project CI in Slack or Feishu after human review.
• Update the parent verl recipe submodule reference. This PR targets the recipe repository itself; the parent pin can be updated after merge.