otbr-agent ran for hours, then died on a burst of Matter traffic:
[W] P-RadioSpinel-: radio tx timeout
[C] P-RadioSpinel-: Failed to communicate with RCP
otbr-agent exited with code 6
The RCP itself was fine -- its frame timestamps showed 18 hours of
uninterrupted uptime across the crash, so nothing on the device reset.
The host simply lost bytes and desynchronised.
Nordic's nrf54l15dk overlay runs uart20 at 1 Mbaud, and this board's
overlay copied that, but the DK pairs it with hw-flow-control. The XIAO
cannot: uart20's pinctrl defines only TX and RX, and the CMSIS-DAP
bridge carries no RTS/CTS. Without back-pressure a single late RX
interrupt drops a byte, and a dropped byte corrupts the HDLC frame
around it.
460800 keeps roughly 4x headroom over peak 802.15.4 throughput while
widening the per-byte service window from 10 us to 21.7 us. The receive
queue goes to 8192 bytes for the same reason -- with no flow control,
whatever does not fit is lost rather than deferred.
The RadioURL on the border router host must set the matching
uart-baudrate=460800.
Two defects, both specific to running this sample on a XIAO rather than a DK,
and both only on the border router side -- which is why the link was so lopsided
and why a Matter device could attach but not sustain SRP updates.
The coprocessor sample sets CONFIG_GPIO=n, which is reasonable on a DK whose RCP
has no GPIO-controlled RF hardware. This board selects its antenna path with
one: P2.03 powers the RF switch and P2.05 picks ceramic or external, declared as
regulator-fixed nodes with regulator-boot-on. With GPIO off, REGULATOR_FIXED
cannot build (it depends on GPIO), so those nodes got no driver, neither pin was
ever driven, and the border router ran with an unpowered antenna switch. The
devicetree looks correct throughout, which is what made this hard to see.
Separately, Zephyr defaults OPENTHREAD_DEFAULT_TX_POWER to 0 dBm and the sample
never overrides it, while Nordic's Matter samples use 8 -- so the border router
was transmitting 8 dB weaker than the devices talking to it. That asymmetry hits
beacon responses during a joiner's active scan, which is exactly what kept
failing.
Measured: -93 dBm at close range before, -71 dBm after, against a -101 dBm
sensitivity floor.
