BLE-anchor impairment robustness --- does the RSSI calibration anchor survive a commodity NIC where the CSI phase feature does not?

Question

Count-without-identify rests on a calibration anchor. We use a co-registered BLE/RSSI anchor (c-ble-csi-coregistration). The CSI impairment grid (c-csi-impairment-sim-to-real) showed the CSI occupancy phase feature (Rician-K slope) is fragile under a commodity-NIC measurement model --- it sign-flips / blows up even at 30 dB SNR. Is the BLE RSSI anchor, a narrowband power measurement, robust to the same impairments? If so, that is a positive, physically-motivated reason to anchor calibration on BLE rather than on a second CSI feature.

Framing discipline. This is in-silico motivation for ble-periodic-calibration under a self-authored impairment prior (IP-101 Q3=A) --- not a measured fidelity claim. RSSI is a power measurement, so the physics predicts immunity to the phase terms (CFO, phase noise, IQ) that corrupt CSI phase, with only AWGN touching it (a scale shift, not a rank loss). The honest deliverable is the clean→impaired contrast curve (BLE rank-correlation preserved vs CSI phase feature distorted), with the circularity boundary (self-authored noise) stated. The real defeater is multi-week hardware capture on ≥2 sites.

What we already know

• c-csi-impairment-sim-to-real (sealed 01KTPNH0PV…): the CSI occupancy signature is fragile --- mean blockage attenuation halves and the Rician-K slope sign-flips even at SNR=30 dB. The impairment stage is correct + strictly additive (clean oracle preserved); csi_impaired.hdf5 carries the impaired CFR and domain_metrics.impairment carries the impaired scalars.
• c-ble-csi-coregistration (sealed 01KT9E56YZ…): BLE RSSI is synthesised from the SAME solved CFR; the anchor resolves coarse occupancy where temporal CSI collapses (ρ(N, RSSI) ≤ −0.6) but weakly (−0.115 dB/person on the through-wall rig).
• The runner synthesises BLE RSSI from the clean CFR; the impaired RSSI is therefore computed offline from csi_impaired.hdf5 (narrowband centre-slice power) by the reduction notebook --- no runner change is needed.

What the supervisor does

Fan out exp-csi-static across the 3 × 5 (n_agents × SNR) grid with the fixed phase-impairment base + ble.enabled (15 runs; each carries both the clean and impaired CFR). The reduction (csi_ble_impairment.py, runner python) pools per (frame, link): clean and impaired narrowband BLE RSSI vs n_occluders (rank-correlation + slope), the CSI Rician-K occupancy slope clean vs impaired (from the scalar metrics), and emits ble_impairment_curve.parquet + the csi_ble_impairment figure (occupancy rank-correlation vs SNR: BLE preserved, CSI-K distorted).

Out of scope

• Any BLE-vs-CSI estimator MAE comparison (circular against the self-authored noise model).
• Real hardware impairment fitting / measured fidelity (that is exp-csi-calibration on real capture).
• Moving crowds; SNR below 5 dB / above 40 dB.

Expected interpretation

1. Anchor robust (BLE |Δρ| small across SNR while CSI-K distorts ≥ ×2) → "the BLE anchor's occupancy ordering survives commodity-NIC impairments that corrupt the CSI phase feature --- a positive reason to anchor calibration on BLE." Strength of ble-periodic-calibration stays plausible (motivation, not proof).
2. Anchor also fragile (BLE ρ collapses with SNR) → the anchor needs a noise-robust readout before hardware investment --- a cheap, valuable negative.
3. Criterion 1 fails (missing artefact / NaN) → plumbing finding, fix before interpretation.