fix(calibration): PresenceSpecialist reads empty room as present when occ_var < empty_var (#1440)

[ruvnet]

Addresses part of #1440.

Scope — what this fixes and what it doesn't

Issue #1440 is a rigorous controlled reproduction with two distinct findings:

1. Raw-CSI amplitude-only presence detection sits at chance (AUC ≈ 0.576 once time-of-day/PHY confounds are controlled). This is a genuine sensing-difficulty finding, not a code bug — already addressed by the existing maintainer comment pointing to the per-room calibration path (ADR-151) as the intended production route.
2. The calibration path itself misbehaves: after room calibration (train_empty_room/train_still_presence/train_walking/train_active_movement), a known-empty room read "present" 31/31 frames, and training accuracy was 0.45 on 4 classes. This part is a code bug, and this PR fixes it.

Root cause

PresenceSpecialist::train() (v2/crates/wifi-densepose-calibration/src/specialist.rs) sets its variance decision threshold at the midpoint between the empty-room baseline variance and the mean occupied-anchor variance, assuming occupied windows are always noisier than empty ones. A still, quiet occupant (train_still_presence) can measure less variance than an empty room's own ambient/interference noise floor — in that case the midpoint sits below the empty-room baseline itself, so a genuinely empty room reads "present" on every frame, at nonzero confidence. This is consistent with the reporter's 31/31 symptom and a training accuracy barely above chance for 4 classes.

Fix

Only trust the variance channel when occupied variance exceeds the empty baseline by a margin (VARIANCE_SEPARATION_MARGIN = 5%); otherwise disable it (threshold = +inf), mirroring how the mean-shift channel already goes inert when the anchors' means don't separate (existing mean_dist_threshold: Option<f32> pattern). Also fixed a latent confidence-scoring bug this exposed: an infinite threshold previously drove the variance confidence term to a spurious 1.0 via (x - inf).abs() / span, which would have overstated confidence in a disabled channel's contribution.

Testing

• New test presence_inverted_variance_never_reports_empty_room_as_present, isolating the exact scenario (occ_var < empty_var, identical means so there's no mean-shift fallback) that reproduces the bug pre-fix and confirms the fix post-fix.
• cargo test -p wifi-densepose-calibration --no-default-features: 63 unit + 1 integration test, 0 failed, 0 regressions (all pre-existing presence/mean-shift/backward-compat tests still pass).
• No new clippy warnings.

What's still open

The reporter's deeper questions (is phase the right lever? is ruview-unified the intended future path? placement guidance?) are legitimate open research questions this fix doesn't answer — flagging so the issue isn't fully closed by this alone, just the concrete regression within it.

Co-Authored-By: claude-flow ruv@ruv.net