[hallerite]

Adds opt-in score centering to IcePop (trainer.loss.score_centering = true), using captured top-k sampler probabilities and a proportional tail approximation.

• Bumps verifiers to 512b8f7 for sampler probability transport through traces and training branches.
• Bumps renderers to 70c8204 for top-logprob capture and compact response parsing.

---

Note

High Risk
Touches RL loss gradients, inference logprob transport, and training data validation on the critical policy–inference mismatch path; misconfiguration or missing top-k evidence fails hard but wrong rollout data could skew updates.

Overview
Adds opt-in score centering for IcePop (trainer.loss.type = "icepop" + score_centering, default top-k 128). The launcher enforces untruncated stochastic policy sampling, bumps sampling.logprobs and vLLM processed_logprobs / max_logprobs, and documents the approximate head+tail centering math and metrics.

End-to-end sampler top-k evidence is new: compact binary logprob payloads from vLLM, TopLogprobs on traces/micro-batches, validation in the trainer, and fused LM-head top-k logprobs (with gradients) feeding an IcePop correction that still applies when the sampled token is ratio-rejected.

The Qwen tool example switches agent runtime from Modal to Prime sandboxes (README + tool.toml). Lockfile bumps prime-sandboxes and drops unused Modal/e2b extras; a GLM FLOP test fixture sets dsa_backend="tilelang" for CI.

^{Reviewed by Cursor Bugbot for commit b3d985a. Bugbot is set up for automated code reviews on this repo. Configure here.}