A transmitter seeded itself from the builtin hash() of its label, and that
is salted per interpreter, so every run produced different synthetic
speech and keying. A test that failed could not be made to fail again --
the one thing needed to fix it. crc32 gives the same content every run.
The hang test then compounded it by listening for a 1 s burst in a 6 s
cycle over three sweeps: it is the release after the transmission that is
under test, but a missed burst failed it just the same, with "nothing
recorded". Now 2 s in 5 s over six sweeps, so only the ending can fail it.
The neighbouring test had the opposite fault: asserting only that nothing
was truncated, it passed whether or not the burst was ever heard. It now
requires the recording it is drawing a conclusion from.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
af51b2657d