feat: build KeptTokens from native sampling_mask

[mikasenghaas]

Summary

• Build KeptTokens from vLLM's native sampling_mask payload (one list of surviving vocab ids per completion token; native since vLLM 0.28 via --return-sampling-mask), replacing the custom base64 kept_tokens payload from prime-rl's capture patches (superseded by native support).
• KeptTokens.from_sampling_mask converts to the flat int32 ids/counts arrays; graph attribution validates alignment and attaches the arrays as-is (no base64 decode/copy). Branch.kept_tokens is unchanged for downstream consumers.

Pairs with PrimeIntellect-ai/renderers#144 and prime-rl's native sampling-replay PR.

🤖 Generated with Claude Code

Note

Build KeptTokens from native sampling_mask instead of kept_tokens dict

• Adds KeptTokens.from_sampling_mask classmethod in types.py to construct counts and ids as flat int32 numpy arrays from a list-of-lists sampling_mask.
• response_from_generate in train.py now reads sampling_mask from the generate result and calls from_sampling_mask instead of unpacking a kept_tokens dict.
• _attribute_kept_tokens in graph.py no longer base64-decodes payload.counts/payload.ids; it validates lengths and attaches the payload directly to the node.
• Behavioral Change: response_from_generate no longer accepts a kept_tokens payload; callers must provide sampling_mask instead.

^{Macroscope summarized f0d4154.}