[PIR][WP5] Build new crates/rvAgent/latentmesh-* crates (greenfield — no existing ruvector wiring)

Epic: #837 · Cross-repo counterpart: ruvnet/LatentMesh#3 · Plan: docs/research/perpetual-intelligence-runtime/03-program-plan.md

Goal

Build new crates under crates/rvAgent/ (alongside rvagent-a2a, ADR-159) implementing network transport, RVF packaging, and rvm-cap admission enforcement against LatentMesh's latentmesh-core / latentmesh-align / latentmesh-gate wire format.

This is greenfield — scope it that way

Two facts that set expectations:

• ruvector has zero existing LatentMesh integration. The only trace is an unmerged origin/docs/link-latentmesh branch stub.
• LatentMesh itself is ~1,407 LOC across 23 tests and contains no network transport crate. Its wire vocabulary (LatentFrame, F32/F16/Int8 encodings) and training-free orthogonal Procrustes/SVD alignment are real and tested; transport, RVF packaging, and RVM enforcement are not present anywhere.

So adopting ADR-009 gives us a design contract, not a codebase. This work package is near-greenfield engineering, not integration, and is the single largest source of the program's revised top risk (cross-repo coordination overhead).

Acceptance criteria

• New crates/rvAgent/latentmesh-* crates implement real network transport for LatentFrame — not an in-process channel or test harness. Wire format, versioning, and failure/retry semantics documented.
• Latent frames are packaged through RVF, carrying as artifact metadata (per LatentMesh ADR-008): model identity, applied transform, permitted recipients, provenance, and witness history.
• Admission is enforced via rvm-cap capability tables — a frame whose recipient is not permitted is rejected at the gate, demonstrated by test, not filtered downstream by the consumer.
• A live multi-agent task runs LatentMesh's causal-edge verification (ADR-003) end-to-end for the first time, from ruvector. Per LatentMesh ADR-009, that machinery has never been run against a live task. Report what it finds — including if it shows the channel carries less than assumed. That is a valid and valuable result.
• Every frame transmission emits an RVM witness record, satisfying invariant 3 ("every agent communication is attributable").
• LatentMesh maintainers sign off on wire-format compatibility (tracked in ruvnet/LatentMesh#3).
• If direct P2P needs a rendezvous/relay, host it as a Cloud Run service in us-central1 with session affinity — and per rvm ADR-285, do not claim isolation strength a hosted relay does not have.
• npx @claude-flow/cli@latest security scan clean — this touches capability tables.

Dependencies

WP1 (#848) and WP4 (#840) — the RVF packaging format and witness schema must be settled before packaging against them. Blocks WP6 (#850) and WP7 (#853).

Start with docs/research/perpetual-intelligence-runtime/03-program-plan.md, ADR-009, and the LatentMesh wire-format implementations in latentmesh-core, latentmesh-align, and latentmesh-gate. Inspect the settled interfaces from WP1 (#848) and WP4 (#840) before designing crates/rvAgent/latentmesh-*. Done means real transport, RVF metadata, gate-level rvm-cap rejection, witness records, end-to-end causal-edge verification, compatibility sign-off, and a clean security scan.