feat: add Qwen3.8-27B (dense) support

[samsja]

What

Adds support for Qwen3.8-27B (the new dense 27B model built on the Qwen3.5 architecture) to prime-rl.

Qwen3.8-27B uses the qwen3_5 / qwen3_5_text model types, so it already resolves to the existing custom Qwen3.5 dense implementation. This PR makes sure the Qwen3.5-family trainer patches run for it and bumps the renderers/verifiers submodules to the new Qwen3.8 renderer + test coverage.

Changes

• src/prime_rl/trainer/model.py — extend the Qwen3.5-family patch gate to also match "Qwen3.8" by name, and broaden the model_type fallback from qwen3_5_moe to any qwen3_5* (covers qwen3_5_text for renamed Qwen3.8 checkpoints). This ensures the position-ids / linear-attention / conversion patches apply to Qwen3.8.
• deps/renderers → 5520d1e — the new qwen3.8 renderer (feat(renderers): add Qwen3.8-27B (dense) renderer renderers#132).
• deps/verifiers → 2c93838a — Qwen3.8 added to renderer e2e/client test matrices (test(verifiers): add Qwen3.8-27B to renderer test matrices verifiers#2416).
• README.md, docs/advanced.md — document Qwen3.8 in the supported-model tables.

Notes

• vLLM already serves the Qwen3_5ForConditionalGeneration architecture (incl. the XML tool parser) natively, so no inference-side change is needed.
• Qwen3.8 checkpoints carry trained mtp.* weights; these are ignored by the non-strict DCP load, same as Qwen3.5-MoE.

---

Note

Low Risk
Small change to when existing Qwen3.5 monkeypatches run; no new modeling or load path. Wrong gating could skip or extra-apply patches, but the logic is a straightforward name/model_type expansion.

Overview
Enables Qwen3.8-27B (dense, same stack as Qwen3.5) by running the existing Qwen3.5 trainer patches for it.

In get_model, the name gate now matches "Qwen3.8", and the config fallback is broadened from qwen3_5_moe* to any model_type starting with qwen3_5 (covers qwen3_5, qwen3_5_text, and renamed checkpoints). Docs list Qwen3.8 as reusing the qwen3_5 VLM/custom path.

^{Reviewed by Cursor Bugbot for commit 6e8a5af. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 6e8a5af. Configure here.}

[cursor]

Missing KL mismatch table

Medium Severity

This PR adds custom-trainer support for Qwen3.8-27B, but it does not include the required mean KL mismatch table across 20 steps on the math environment with batch_size=64 (all entries < 0.015). That validation is required when introducing a new custom model so trainer and inference stay aligned.

Additional Locations (1)

• README.md#L55-L56

 

^{Triggered by project rule: BugBot Instructions}

^{Reviewed by Cursor Bugbot for commit 6e8a5af. Configure here.}