[PIR][WP9] Frozen-weight skill/harness evolution loop via @metaharness/darwin (SHAPER pattern, arXiv:2608.11350)

Epic: #837 · Plan: docs/research/perpetual-intelligence-runtime/03-program-plan.md

Goal

Implement the physical-intelligence evolution loop with foundation-model weights frozen throughout: planner and optimizer are the same frozen model, and only skills, context, and the execution harness evolve.

Darwin's mutation surfaces already exist at:

• harness/src/darwin.ts (in crates/ruvector-sota-bench/harness, via @metaharness/darwin)
• examples/mragent — scorePolicy
• crates/sona/src/darwin_guard.rs

Evidence

arXiv:2608.11350 — "Self-Evolving Embodied Agents via Skill-Harness Evolution" (SHAPER), submitted 2026-08-11. Grade A. A train-free framework keeping model parameters frozen while evolving reusable skills and a context/code harness through target-environment rollouts, with the same frozen model acting as planner and optimizer. Evaluated on VLABench and ESI-Bench. No name collision — "SHAPER" is specific to this paper.

Acceptance criteria

• A CI check proves no promotion-reachable code path can write to foundation-model weight files. The check fails the build if any mutation surface reachable from the promotion pipeline imports a training or fine-tuning API. Policy statements are explicitly insufficient — this must be mechanically enforced, and it is the acceptance test's central constraint.
• A VLABench/ESI-Bench-style evaluation shows skill improvement without weight changes.
• Skills, context, and harness are separately versioned mutation surfaces mapped onto Darwin's existing surfaces; weights are not a mutation surface.
• Foundation-model weights are hash-verified at loop start, and the hash is recorded in the witness chain.
• The same frozen model instance serves as planner and optimizer (SHAPER's structural claim), verified by configuration test.
• Every evolution cycle terminates in a WP2 research-gate verdict — no promotion path bypasses it.

Dependencies

• WP0b (#846) — the MetaHarness hard-dependency non-compliance and the ruvllm HTTP-307 download bug both block this. Darwin's local mutator backend (ADR-259) cannot be live-serve tested end-to-end until the 307 bug is fixed, and this WP depends on @metaharness/darwin.
• WP2 (#838) — the promotion gate this loop terminates in.

Blocks WP10 (#842) and WP12 (#843).

Start with docs/research/perpetual-intelligence-runtime/03-program-plan.md and trace the named mutation surfaces in harness/src/darwin.ts, examples/mragent (scorePolicy), and crates/sona/src/darwin_guard.rs. Review the WP0b and WP2 dependencies before attempting evaluation. Done means the frozen-weight policy, hash verification, shared planner/optimizer instance, evaluation, versioned surfaces, and research-gate termination are mechanically verified.