fix(calibration): bind the grid whose frames come from one population - #1919

Why

A field model compares a runtime window mean against a calibration reference built
the same way, and it binds exactly one subcarrier grid. That comparison only
means something when both windows come from the same transmitter population.
select_calibration_grid ordered candidates by max gap, then rate, then width, so
a faster but mixed grid could win.

Measurement

Three ESP32-S3 nodes on one access point, 30 s capture per node, frames counted
straight off the UDP socket:

The 306 grid carries the array's own traffic: each node pings the gateway
(CONFIG_CSI_SELF_PING_HZ, documented as "a stable OFDM CSI source") and cannot
hear its own transmission, so what a node hears on that grid is the other
nodes. That population is strong and homogeneous.

With the mixed 192 grid bound, injecting 2000 pps from the host moved the field
model's residual median from 9.4 to 17.4 while the operator sat motionless -
a larger change than the occupant's own effect in the same session (empty room
median 10.4, seated median 11.9). The residual was following the frame population
instead of the room.

Related, and worth knowing when a report looks impossible: with two of the three
nodes powered down, node 3 saw 1.6 Hz on the 306 grid and 0.2 Hz on 192, both at
-58 dBm, and a 600 pps injection from the host raised neither (1.2 Hz and 0.9 Hz).
A one-node array has no steady population to calibrate against.

Change

• RawGridObservation records the frame's reported RSSI.
• CalibrationGridEvidence reports rssi_spread_db and rssi_median_dbm.
• select_calibration_grid orders by RSSI spread first, then max gap, then rate,
  then width. The eligibility gates (10 s span, 2 Hz, 5 s gap) are unchanged, and
  a grid without RSSI evidence falls back to the previous ordering exactly.

Tests

cargo test -p wifi-densepose-sensing-server --bin sensing-server
# 323 passed, including the new
#   calibration_grid_selection_prefers_a_homogeneous_population

Limits

• Measured on one rig (three ESP32-S3 nodes, one Tenda access point, 2.4 GHz
  channel 11) on one day. The spread is used for ordering only, never as a hard
  cutoff, so a rig whose grids all look homogeneous keeps its previous choice.
• The occupant's own effect on the residual is small on this rig (median 10.4
  empty versus 11.9 seated), which is a separate problem: this change removes the
  traffic-driven part of the variance, it does not by itself make the residual a
  validated presence signal.