DYODE 1.4.2

Fixed

• Large folder transfers failing part-way (from about 1 GB, at high
  bitrate) on receiving boxes with plenty of RAM and an encrypted (LUKS)
  disk. The cause was disk writeback, not the network. Linux wrote ~30 s of
  received data at once, and that burst stalled udp-receiver long enough to
  overflow its socket buffer, so udpcast aborted the file. install.sh now
  writes /etc/sysctl.d/90-dyode-receive.conf on receiving boxes (v1, and v2
  --side out), with continuous writeback (vm.dirty_background_bytes) and a
  larger UDP receive buffer (net.core.rmem_*). Verified in the field: 1.3 GB
  and 5 GB transfers at 900 Mbit/s that had always failed now arrive with
  identical SHA-256. --tuning-only adds it to an existing box; --no-tuning
  skips it.

Changed

• A stalled transfer's error now says where the packets went, from the
  kernel's UDP RcvbufErrors counter. "N UDP packets were dropped on this box
  because a receive buffer was full" means the receiver stalled (the case
  above). "No receive-buffer overflow on this box" means the packets were lost
  before reaching it (NIC, cabling, diode).

Documentation

• CAVEATS.md rewritten with the confirmed cause. The 1.4.1 entry blamed the
  receive-side NIC; swapping it changed nothing. The entry now has the field
  results, the diagnostic procedure that found the cause, the fix and how to
  undo it, and other tuning for other hardware.