Native GEPA program optimization

GEPA is one of the thirteen optimizer families compiled by the Rust epistemic-graph eg-program authority. Agent Utilities does not carry a Python GEPA implementation or a second provider stack. It maps ephemeral training data to opaque program, corpus, evidence, budget, and promotion contracts, then submits one durable ProgramOptimize job.

The native compiler emits a bounded model-transport plan for GEPA reflection. The plan contains only opaque input/output references, explicit dependencies, a maximum operation count, policy scope, and all observed modalities. Existing governed model transport executes the plan; the optimizer never creates a duplicate LLM client.

Every candidate remains propose-only until typed evaluation evidence proves the promotion policy. No raw prompt, response, identity, endpoint, credential, or local path is representable in the durable contract.

Current implementation:

• agent_utilities/harness/optimization_backend.py — privacy-safe request mapping for all fourteen modalities and all thirteen optimizer families.
• agent_utilities/harness/program_optimization.py — target registry, evaluation, and promotion boundary.
• epistemic-graph/crates/eg-program — optimizer compiler, plans, candidates, and promotion rules.
• graph_evolution action=optimize_component — the sole GraphOS on-demand surface.