[RFC] RFC 006: Multi-Component Reward Signals

[adithya-s-k]

Summary

This PR proposes RFC 006 for a standardized multi-component reward representation that remains fully backward compatible with existing scalar reward pipelines.

Scope

• Adds rfcs/006-multi-component-rewards.md
• Design-only RFC PR (no runtime behavior changes)

Why

• Improve reward observability and debugging in complex agentic tasks
• Keep scalar optimization contract unchanged
• Standardize component-level reward schema across environments

Compatibility

• observation.reward remains unchanged
• Component payload is additive in metadata only

Related

Closes #623

[greptile-apps]

PR title or description contains an excluded keyword.

[adithya-s-k]

Closing without merging. The RFC never got ratified and I'm no longer proposing multi-component rewards as an OpenEnv-core change (see #623 for the full reasoning) — the component/weighting need is being met on the training-framework reward contract instead. Leaving the RFC unmerged so rfcs/ reflects only ratified designs. Happy to re-open a scoped version later if OpenEnv needs its own schema.