chore(deps): bump renderers (offset-tokenizer config-build fallback for Laguna)

[hallerite]

Bumps the deps/renderers submodule 89ab3f0 (v0.1.8.dev35) → 35c2407 (v0.1.8.dev37), pulling in PrimeIntellect-ai/renderers#75.

Why

#2657 bumped renderers to pick up the original Laguna KeyError fallback (PrimeIntellect-ai/renderers#72), but minh kept seeing the same crash on hosted-rl whenever renderers were enabled:

Rollout failed in group N (reverse-text) -
  ModelError() -> KeyError("Missing required keys in `rope_parameters`
                            for 'rope_type'='default': {'rope_theta'}")

Root cause: #72 only wired the _load_fast_tokenizer_directly fallback into _load_tokenizer_via_auto (i.e. into load_tokenizer). _get_offset_tokenizer — the helper hand-coded renderers (LagunaXS2, Qwen35, …) call from attribute_text_segments to get an offset-aware tokenizer — called AutoTokenizer.from_pretrained directly to keep the fastokens patch out of that path, and so bypassed the fallback entirely. That path fires on every rollout for hand-coded renderers, which is why disabling renderers made the symptom go away.

PrimeIntellect-ai/renderers#75 routes the offset-tokenizer load through _load_tokenizer_via_auto instead. Same vanilla (no-fastokens) path, but the config-build fallback now applies.

Also pulls in PrimeIntellect-ai/renderers#74 (message_tool_names field for per-message tool attribution).

Verification

Reproduced end-to-end on a single H200 node against poolside/Laguna-XS.2 with the reverse-text env. Without the fix every rollout failed with the KeyError above; with it, the first two RL steps completed cleanly:

• Step 0: reward 0.3553, 97.9 tok/sample
• Step 1: reward 0.3116, 110.8 tok/sample

The run then OOMed on the trainer side at step 2 (model.layers.33.mlp.experts.w3 FSDP all-gather on a single GPU with 256 experts) — that's a config-sizing issue with the 256-expert Laguna model on dp_shard=1, unrelated to the renderers bug.

🤖 Generated with Claude Code

---

Note

Low Risk
Dependency-only submodule bump with a targeted tokenizer-load bugfix; no prime-rl source changes beyond the pin.

Overview
Advances the deps/renderers submodule from v0.1.8.dev35 to v0.1.8.dev37 (PrimeIntellect-ai/renderers#75 and #74). There is no in-tree application code change beyond the pinned commit.

The important fix routes offset-tokenizer loading through the same path as load_tokenizer, so the existing rope_theta / config-build fallback applies when hand-coded renderers (e.g. Laguna) load tokenizers for attribute_text_segments. That closes rollout KeyError failures that still happened after an earlier fallback that only covered the main tokenizer load.

Also picks up message_tool_names for per-message tool attribution in renderers.

^{Reviewed by Cursor Bugbot for commit c7e5e0a. Bugbot is set up for automated code reviews on this repo. Configure here.}