Lab042: transmit an image over the air, add the air loopback mode

First transmission through a real radio path in the project. A 4756-byte
JPEG went from transmitter to receiver over two antennas 60 cm apart at
225 MHz and reassembled byte-identical, reproduced across three runs.

The cable loopback still needs an SMA male-to-male adapter that has not
arrived, so this air run is a track verification, not the controlled
measurement: room levels are not held the way a cable holds them. Its
purpose was to exercise the device wrapper the models cannot cover.

Four defects the software model passed but the hardware exposed, each
reproduced and fixed:

- a 566 528-sample frame did not fit the 262 144 receive buffer, and the
  sample stream is not continuous across separate reads, so a frame on the
  seam was destroyed. Reception now uses one buffer sized to the frame
- carrier correction was always applied, but with a shared reference the
  true offset is near zero and the 80-symbol marker estimate is noise;
  correcting by it rotated phase along the whole frame, 65% bit errors
  measured. Ported should_apply_cfo_correction from Lab023
- receiver gain of 20 dB left the signal at 0.0003 of full scale. Levels
  set by measurement: RX 50 dB, TX -10 dB give zero bit errors with margin
- the error-vector metric returned zero on a failed receive, reporting a
  perfect signal where there was none, the same class of defect as the
  written-in figures of Lab041. Replaced with an explicit not-computed value

The air TX gain ceiling is now derived from the link budget by antenna
distance rather than a constant. Three loopback modes: software, cable,
air. PROJECT_LOG entry 019 records the run and the defects.

Software and cable results stay reproducible; 125 tests pass, gate passes.

protocol/bpsk_radio.py

@@ -31,6 +31,13 @@ SYMBOL_RATE = 20_000
 CFO_REFINEMENT_HALF_WIDTH_HZ = 200.0
 CFO_REFINEMENT_STEP_HZ = 1.0
 
+# Пороги разрешения частотной коррекции. Значения из Lab023. Смысл в том,
+# что при малом истинном уходе оценка по короткому маркеру состоит почти
+# целиком из шума, и коррекция таким значением портит длинный кадр сильнее,
+# чем отсутствие коррекции вообще.
+CFO_DEAD_ZONE_HZ = 15.0
+MINIMUM_PHASE_CONSISTENCY = 0.55
+
 MAXIMUM_PROTOCOL_PACKET_BYTES = 4096
 
 # Перенесено дословно из Lab023 (tests/lab023_guarded_cfo_correction.py).
@@ -800,3 +807,35 @@ def radio_frame_bit_count(protocol_packet_bytes: int) -> int:
         + RADIO_HEADER_BIT_COUNT
         + protocol_packet_bytes * 8
     )
+
+
+# Перенесено дословно из Lab023 (experiments/lab023_guarded_cfo_correction.py).
+def should_apply_cfo_correction(
+    estimated_frequency_hz: float,
+    phase_consistency: float,
+    coherence_gain: float,
+) -> bool:
+    """
+    Разрешить CFO-коррекцию только при наличии
+    достаточных оснований.
+
+    Требования:
+
+    1. Оценка находится вне мёртвой зоны.
+    2. Межсимвольное вращение достаточно согласованно.
+    3. Компенсация CFO действительно повышает
+       когерентность известного маркера.
+    """
+
+    MINIMUM_COHERENCE_GAIN = 0.02
+
+    return (
+        abs(
+            estimated_frequency_hz
+        )
+        >= CFO_DEAD_ZONE_HZ
+        and phase_consistency
+        >= MINIMUM_PHASE_CONSISTENCY
+        and coherence_gain
+        >= MINIMUM_COHERENCE_GAIN
+    )