[PIR][WP2] Adopt ruvnet/dream-machine engine as the promotion gate; consolidate 4 implementations (research-gate + Darwin)

Epic: #837 · Cross-repo counterpart: ruvnet/dream-machine#22 · Plan: docs/research/perpetual-intelligence-runtime/03-program-plan.md

This is an adoption and consolidation, not a build

ruvnet/dream-machine is a real repo — engine v0.1.0 on npm, ADR-0001 Accepted, ~2.5K LOC, 85 tests — and its thesis is this program's thesis:

"Freeze the model. Evolve the harness. Evaluation is not promotion — the machine never merges; a human does."

That is invariant 5 plus WP9's frozen-weights constraint, already stated and already implemented elsewhere in the org.

Goal

Adopt the ruvnet/dream-machine engine as PIR's evaluation core, wire it to this repo's existing statistical promotion machinery and to Darwin, and consolidate four independent implementations of the same idea into one:

This repo's contribution to the merge already exists and is CI-wired: scripts/research-gate/ plus .github/workflows/research-*.yml (ADR-282), backed by crates/ruvector-sota-bench/harness — statistics.ts (paired bootstrap significance testing), vetoes.ts (hard vetoes that block promotion regardless of the statistical result), flywheel.ts (Ed25519-signed replay bundles for reproducible re-verification).

Reframe, don't rebuild. An earlier program pass called Dream Machine its largest gap. It is instead its largest duplication: four implementations, no integration.

Acceptance criteria

• The ruvnet/dream-machine engine is adopted as the evaluation core, wired to Darwin's mutation proposals and to this repo's research-gate statistics.
• A documented integration contract states which side owns the paired-bootstrap computation, the hard vetoes, and the signed replay bundle — with no duplicated authority between engine and harness.
• No mutation reaches RVM's proof gate without passing a verdict. Enforced structurally at the API boundary, with a bypass-attempt test.
• "Evaluation is not promotion" holds end-to-end. The verdict is an input to ruvector-proof-gate/rvm-proof, never a bypass; a test asserts an approved mutation still fails promotion when the proof gate rejects it.
• Hard vetoes still override a passing bootstrap, verified by test.
• "Statistically credible improvement" is stated concretely — significance threshold and pre-registered minimum effect size — reusing the existing statistics.ts implementation rather than a new one.
• Verdicts and replay bundles are witness-logged against ruflo ADR-322C's verification contract (JCS, SHA-256, Ed25519 with domain separation, independent statistical recomputation), which PIR adopts program-wide.
• A migration/deprecation position for all four implementations — which consolidate, which stay, and why. Leaving four live is the outcome this WP exists to prevent.
• @metaharness version drift with ruvnet/dream-machine is resolved first (see WP0b, #846): that repo composes darwin 0.9.1 / flywheel 0.1.10; this repo pins 0.8.0 / 0.1.7.
• Cloud Run control plane in project ruv-dev, region us-central1, mirroring mcp-brain-server's deployment as ruvbrain (same project/region, session affinity, reusing its cloudbuild files as a template), with secrets from existing Secret Manager entries.

Evidence

No external paper prescribes this wiring — it is consolidation of in-house subsystems. The credibility bar it must meet is motivated by arXiv:2607.26773 (2026-07-29): end-task performance gains alone do not establish that the mechanism you credit caused them.

Dependencies

WP1 (#848) and WP0b (#846, for the version bump). Paired with ruvnet/dream-machine#22. Blocks WP9 (#841) and transitively WP10, WP12.

Start with docs/research/perpetual-intelligence-runtime/03-program-plan.md, ADR-282, scripts/research-gate/, .github/workflows/research-*.yml, and crates/ruvector-sota-bench/harness. Trace statistics.ts, vetoes.ts, and flywheel.ts alongside the Dream Machine and Darwin integration points, then review dependencies #848 and #846. Done means the four implementations have a documented ownership and migration plan, enforced verdict and proof-gate boundaries, passing tests, and the specified Cloud Run deployment.