docs(adr): ADR-084 — RaBitQ similarity sensor for CSI/pose/memory (proposed)

[ruvnet]

Summary

Adopt RaBitQ-style binary sketches as a first-class cheap similarity sensor at five points in the RuView pipeline, using the already-vendored ruvector-core::quantization::BinaryQuantized (32× compression, NEON+POPCNT-accelerated, 1-bit sign quantization + hamming distance).

The pattern at every site:

dense embedding → RaBitQ sketch → hamming pre-filter → top-K → full precision refinement
                                                       └─→ novelty score → escalation gate

The decision boundary is unchanged at every site — full precision still owns the final answer. The sketch is a sensor that gates which comparisons run, not what they decide.

Five implementation passes (each independently testable)

1. wifi-densepose-ruvector::sketch module — re-export BinaryQuantized with a thin RuView-flavored API (Sketch::from_embedding, SketchBank::topk, sketch_version: u16).
2. AETHER re-ID hot-cache filter — sketch-prefilter to top-K before full 128-d cosine across active tracks × candidates.
3. Cluster-Pi novelty sensor — per-room SketchBank of normal-state sketches; emit novelty_score on the WebSocket update envelope; CNN wake gate uses it.
4. Mesh-exchange compression — inter-cluster broadcasts carry sketch + witness instead of full embedding when novelty is low (composes with ADR-066 swarm bridge).
5. Privacy-preserving event log — log table stores (sketch_bytes, sketch_version, novelty_score, witness_sha256) instead of raw embeddings.

Acceptance test (per source proposal)

• Sketch compare cost reduction: 8×–30× vs full float
• Top-K candidate coverage: ≥ 90% agreement with full-float pass
• End-to-end accuracy regression: < 1 percentage point

Site-by-site rollback if any criterion fails at a given site; remaining sites continue. Acceptance requires the numbers to hold on ≥ 3 of the 5 passes (the load-bearing ones: AETHER re-ID, cluster-Pi novelty, event log).

Composes with ADR-083

The cluster Pi (ADR-083 / PR #428) is the natural home for the sketch bank. Sensor MCU is unchanged by this ADR; sketches happen at the cluster Pi.

Status

Proposed — not committing to implementation. The ADR formalizes the pattern and the implementation home; actual implementation lands as five separate work items, gated by the acceptance numbers above.

Open question

Whether RuView embeddings need a Johnson-Lindenstrauss / RaBitQ-paper randomized rotation before sign-quantization (the published RaBitQ algorithm by Gao & Long, SIGMOD 2024 includes this for theoretical error bounds). Today's BinaryQuantized is pure 1-bit sign — works when embeddings are roughly zero-centered and isotropic. AETHER and CSI spectrogram embeddings need to be benchmarked for this; if either fails the assumption, the rotation pre-pass becomes load-bearing and warrants a follow-up ADR. Decided after pass-1 benchmark, not in this ADR.

Test plan

• ADR references existing ADRs correctly (ADR-024, ADR-027, ADR-029, ADR-030, ADR-066, ADR-076, ADR-081, ADR-082, ADR-083)
• Implementation home points to vendored ruvector-core::quantization::BinaryQuantized (already in tree)
• All path references already on v2/ (rebased on top of chore(repo): rename rust-port/wifi-densepose-rs → v2/ (flatten to one level) #427's rename merge)
• Reviewer accepts the five-site scope and acceptance criteria

🤖 Generated with claude-flow