Multi-node calibration silently produces false occupancy

Multi-node calibration silently produces false occupancy

Repo: ruvnet/RuView (sensing-server) · Severity: High · Measured: 2026-09-14/15, 3–4 ESP32-C6 nodes

Symptom

A room calibrated with four nodes reports an occupant in an empty room, in every frame.

Root cause

maybe_feed_calibration_frame admits every bound radio into the field model:

// The field model remains single-link, but every radio in the room
// binding may contribute bounded observations on the frozen grid.

The model is single-link: one baseline, one set of amplitude offsets. Averaging several
radios' offsets into it flattens the eigenstructure — hence baseline_eigenvalue_count: 0
— leaving only a scalar energy threshold, inflated by the cross-radio spread (19.96 vs 3.65).

person_count_at (main.rs:2232) already documents the scoring half:

// A single-link model may score only the source node it was calibrated
// against. Applying one node's baseline to a different radio creates
// deterministic false occupancy from hardware-specific amplitude offsets.
let bound_source_node_id = (self.calibration_source_node_ids.len() == 1) …

…but it only engages when exactly one node is bound. With more, it falls through to the
shared mixed history — so the guard never fires in the case its own comment describes.

calibration_start accepts a multi-node binding the model cannot honour.

Fix

Restrict the baseline write, not the admission: only the grid-bound radio calls
maybe_feed_calibration; every bound radio still registers as a contributor and is scored
from the bound radio's history. This mirrors the narrowing bootstrap_baseline::store
already applies (vec![binding.source_node_id] as the frozen model source).

Caution for reviewers: gating the whole feed path instead deadlocks the capture —
calibration_source_nodes_missing requires every bound node to contribute before finalize.
Measured: 43,120 frames over 51 minutes stuck in collecting, missing=[12,13,14], no
error surfaced. The accompanying PR carries a regression test for exactly this.

Fixed in #1936.

Start with calibration_start and maybe_feed_calibration_frame, then read person_count_at in main.rs around line 2232 and compare the source narrowing in bootstrap_baseline::store. The accompanying PR contains the regression test; done means multi-node calibration keeps all bound radios as contributors while writing the single-link baseline only from the grid-bound radio and avoids false occupancy.