feat(v1): add Trace.reward_shaping for trainer reward-shaping terms

[hallerite]

Summary

Adds Trace.reward_shaping: dict[str, float] — additive training-reward terms a consumer's reward shapers record on a trace, keyed by shaper name (e.g. {"length_penalty": -0.1}).

This is a training annotation in the same sense as MessageNode.advantages: verifiers holds the field, the trainer fills it. Trace.reward stays the env's verdict; the trainer scores against reward + sum(reward_shaping.values()). Keeping the two apart means task-success metrics, curricula and evals keep reading the unshaped reward while credit assignment sees the shaped one.

Companion: PrimeIntellect-ai/prime-rl#3376 moves prime-rl's length penalty out of GRPO into a reward-shaping layer and writes its terms here.

🤖 Generated with Claude Code

Note

Add reward_shaping field to Trace model for trainer reward-shaping terms

Adds a reward_shaping: dict[str, float] field to the Trace pydantic model, defaulting to an empty dict. It stores additive training-reward terms keyed by shaper name. The field is included in JSON output from to_record() since it is not excluded.

^{Macroscope summarized e084c75. (Automatic summaries will resume when PR exits draft mode or review begins).}