The CAD pre-TX loop previously ran up to 10 retries with a fixed 10 ms
sleep — roughly one max-length packet only at SF7/BW500. At SF10/BW125
the 100 ms total was ~4% of one packet; at SF12/BW125 LBT was a no-op.
Now the per-retry sleep is sized to one time-on-air of an assumed-peer
80-byte packet under the active config, cached on RadioContext and
recomputed on every SET_CONFIG. Retries scale as
clamp(floor(BUDGET/T)+1, 1, 10) with BUDGET=2.5 s, so fast configs
still get up to 10 retries while SF11/12 degrade to single-CAD-and-bail
(logged at config-apply). Sleeps are jittered ±20% per iteration to
decorrelate simultaneously-backing-off contenders.
The ToA math, CR→{1..4} mapping, retry-count derivation, and jitter
helper live in a pure src/lbt.rs module so they're host-testable via
DONGLORA_HOST_TEST=1 cargo test (13 tests).
Validated on-device: 20 iterations of the AI Bot test flow that
previously saw ~2 ChannelBusy errors now see zero at the same SF/BW.
