[PIR][WP24] Zetta-pattern frozen-policy embodied critic/recovery evolution (ADR-327, stretch, new context)

阅读 docs/research/perpetual-intelligence-runtime/08-wave3-evidence-review.md 和 09-wave3-program-plan.md，然后检查 crates/ruvector-sota-bench/harness 以及现有的 SHAPER 循环。使用 Zetta-Embodiment 仓库作为模拟 harness 的参考。完成标准是：冻结 policy 的三时间尺度循环满足所述验收标准，并通过使用独立重新计算的 research-gate delta 进行的留出 seed 验证。

Epic: #837 · Wave 3, Phase W3-4 · Proposed ADR-327 · Paper: Zetta ζ — An Efficient Closed-Loop Embodied Harness for Self-Evolving skills (arXiv:2608.16590, Grade A, all figures exact-match; reference implementation live at github.com/air-embodied-brain/Zetta-Embodiment, 224 stars, confirmed author repo).

🧭 EXPLICIT STRETCH — sequenced last, no hard blocking dependency, and NOT a physical-robotics commitment. Read the scope boundary below before treating any part of this as a robot-actuation deliverable.

Scope boundary — read first

No repo in the ruvnet org currently operates physical robots or a robot-rollout harness. RVM is an agentic runtime / hosting substrate (no robotics); RuView is RF sensing (no actuation). Zetta therefore has no existing bounded context to extend — ADR-327 proposes a new one (Physical Embodiment), it does not expand an existing context.

Accordingly, this slice is the timescale-separated evolution-harness PATTERN only, on a SYNTHETIC environment:

• The three-timescale critic/recovery evolution structure — ms-scale critic / seconds-scale recovery / minutes-scale skill-evolution — over a frozen base policy, validated against held-out seeds before promotion.
• Landed in crates/ruvector-sota-bench/harness against a synthetic environment.
• Physical / RuView actuation is EXPLICITLY DEFERRED to a future ADR + hardware. Nothing in this WP commits the program to operating physical robots. Any real-world path is sensing-side coordination with RuView only — never actuation.

This is structurally the same kind of first-party stretch bet as Wave 1's WP14 (universal CSI vocabulary) — except WP14 at least had RuView as a plausible eventual home; Zetta's embodied-harness mechanism has none today.

Repo placement — DECIDED (coordinator, 2026-08-21)

The first slice hosts in crates/ruvector-sota-bench/harness, beside WP9's SHAPER loop — Zetta's frozen-base + timescale-evolving critics is a timescale-structured variant of that same loop. NOT a new crate, NOT a sibling-repo epic for the slice. The physical/RuView actuation target stays deferred to a future ADR + hardware as stated above. ADR-327's author has this scoping.

Goal

Stand up a minimal simulated embodied-harness surface (LIBERO-Pro / RoboCasa-class rollout infrastructure, simulated) and a frozen-policy critic/recovery evolution loop, adapting the released air-embodied-brain/Zetta-Embodiment implementation as the starting reference rather than building rollout infrastructure from nothing.

Extends: nothing merged — new bounded context (Physical Embodiment), stretch-only.

Open placement question (coordinator to decide — this issue does NOT decide it)

Where WP24 lands is deliberately left open: a new crates/ruvector-embodiment inside ruvector, or a new sibling-repo epic (mirroring how Wave 1 treated LatentMesh / rvm / autogenous / RuView as separate epics), if the physical-robotics scope is judged too far outside this repo's remit. WP21–23 have unambiguous ruvector-internal homes; WP24 does not. Flagging the question, not resolving it.

Acceptance criteria

• Three timescales are separately governed (action-frequency governance): ms critic, seconds recovery, minutes skill-evolution.
• Base policy is frozen; only critic/recovery/skill layers evolve.
• Promotion is validated against held-out seeds the evolution loop never trains against.
• Environment is synthetic; the issue and ADR state prominently that physical/RuView actuation is deferred to a future ADR + hardware.

Preprint-reproduction rule

Zetta's reported numbers (SOTA on LIBERO-Pro / RoboCasa) are hypotheses, not the acceptance bar. A live reference repo changes how much code must be written — not what counts as evidence. Promotion requires an independently recomputed research-gate paired-bootstrap delta over the pre-WP baseline.

Depends on

None hard. Soft-coordinates with ruvnet/RuView only if/when a real-world (not simulated) path is pursued, on sensing/state-estimation only, never actuation.

Sources: 08-wave3-evidence-review.md (evidence grade + artifact status), 09-wave3-program-plan.md (WP/ADR mapping). Filed by the issue steward after coordinator approval of the Wave-3 breakdown.