[PIR][WP14] Validate ruvector-perception (3.8K LOC) against real CSI captures, not synthetic-only

Epic: #837 · Cross-repo counterpart: ruvnet/RuView#1657 · Plan: docs/research/perpetual-intelligence-runtime/03-program-plan.md

Goal

crates/ruvector-perception is 3.8K LOC and has only ever been validated against synthetic data. Validate it against real CSI captures, and establish a documented accuracy baseline.

This is the ruvector-side companion to RuView's WP14 (ruvnet/RuView#1657), which closes ADR-178's gaps C (CSI I/Q → pose-semantics) and D (mcp-brain-server-side cluster consumer).

Why it matters to the program

PIR's invariant 6 is "every physical action produces new evidence," and the 30-day acceptance harness (WP12) logs sensing accuracy as one of its four metric families. A perception stack validated only on synthetic data cannot support that metric — the harness would be measuring agreement with a simulator, not with the world.

The ruview-csi-bridge at crates/ruvector-hailo-cluster/src/bin/ruview-csi-bridge.rs is currently transport-only: it parses RuView's ADR-018 UDP CSI frames and posts a header-summary string into the Hailo embed RPC, with a source comment explicitly disclaiming pose embedding. crates/ruvector-mmwave (Seeed MR60BHA2/LD2410 radar, ADR-063) provides a ground-truth comparison channel.

Acceptance criteria

• A documented accuracy baseline for ruvector-perception on real CSI captures, reported separately from its synthetic-data performance so any gap between the two is visible rather than averaged away.
• A real-capture dataset exists and is reproducible — capture conditions, hardware, and chipset recorded, since chipset dependence is the open research question RuView's WP14 addresses.
• Where mmWave/radar ground truth is available, perception output is compared against it and the discrepancy characterized.
• If real-capture accuracy falls materially below synthetic, that is reported as the finding, not treated as a blocker to be tuned away. Knowing the synthetic-to-real gap is the point of this work package.
• The measured baseline is wired into WP12's sensing-accuracy metric if WP14 lands within the acceptance window.

Dependencies

None on the critical path — independent, recommended for separate funding alongside ruvnet/RuView#1657, which supplies the pose-semantic pipeline this validates against.

Start by reading docs/research/perpetual-intelligence-runtime/03-program-plan.md and the crates/ruvector-perception scope. Inspect crates/ruvector-hailo-cluster/src/bin/ruview-csi-bridge.rs and the synthetic validation before planning real CSI captures and the mmWave comparison. Done means a reproducible dataset, separate real-versus-synthetic accuracy baseline, discrepancy report, and WP12 integration if applicable.