epic: routing-aware evals #2088

Context

Teams using Switchyard (or similar weak/strong routing) need to systematically measure how routing affects benchmark scores, failures, or cost. Gym can proxy to Switchyard today, but routing policy isn't a first-class eval variable: no comparable runs, no attribution, no guided workflow.

Gym already does this for other knobs (e.g. skills). Routing should work the same way.

Note: These changes may have cross-repo impact: Switchyard + Gym

User outcomes

• Run the same benchmark under different routing conditions and compare scores + token/cost signals fairly.
• Identify whether routing helps or hurts on tasks they care about
• Reproduce routing experiments like any other Gym eval.

Scope Notes

• Enable routing conditions as experimental variables in Gym evals, with outputs that support comparison and diagnosis.
• In scope: eval workflow, provenance/comparability, analysis hooks, docs + ecosystem mention, Switchyard alignment.
• Out of scope: building routers in Gym, replacing Switchyard, RL training paths, switchyard launch interactive workflows, production serving.

Acceptance Criteria

• Documented workflow comparing ≥2 routing conditions on the same benchmark/harness
• Routing condition identifiable in eval outputs across runs
• Docs: when/why Gym + Switchyard
• One end-to-end example published