Release v0.2.1

What's Changed

• test(lineage): read object-store usage from Ray's public resource acc… by @hijkzzz in #227
• deps: vLLM 0.32.0, AutoModel main ea233c99 by @hijkzzz in #230
• dockerfile: cutlass-dsl 4.8.0, quack-kernels 0.6.5 and tilelang 0.1.1… by @hijkzzz in #232
• docs(skill): train-inference-consistency fixes the rollout performanc… by @hijkzzz in #233
• docs(skill): train-inference-consistency follows main on rollout logp… by @hijkzzz in #234
• docs(skill): train-inference-consistency: GPU architecture in the clo… by @hijkzzz in #235
• fix(chat): keep every message a turn appends after the reply by @MaxFreedomPollard in #237
• fix(tools): tolerate invalid bytes in Python output by @k21993 in #236
• feat(kl): --algo.kl.ref rollout — KL to the rollout policy from vLLM'… by @hijkzzz in #238
• fix(kl): the rollout KL reference only acts when a KL coefficient is set by @hijkzzz in #239
• chore(agents): one skill tree under .agents/, mirrored into .claude/ … by @hijkzzz in #241

Full Changelog: v0.2.0...v0.2.1