Capture and playback failures were only written to the log on detached
threads, so a microphone or output that could not be opened still showed a
connected voice room with no sound. Threads now report into a shared
diagnostics sink that the UI polls: errors appear in the status bar, a
room with neither working half is marked failed, and signal counters
(captured, sent, received, decoded, input peak) are logged every ten
seconds. A microphone that delivers only digital silence is reported, and
rejected voice datagrams are logged instead of dropped silently.
Voice activation now tracks the background noise floor and opens just
above it, capped by a much lower default ceiling; the old fixed threshold
(about -26 dBFS) rarely opened for normal speech. Stream configuration only
accepts f32 formats and says so when a device offers nothing else.
Playback gains a mixing queue for short notification effects.
The device pickers showed every ALSA PCM plugin (Discard all samples, Rate
Converter, JACK, Open Sound System, surround/dmix aliases). Enumeration now
goes through a pure filter that keeps the system default, the PipeWire and
PulseAudio servers and one entry per hardware endpoint, with technical tails
removed and a stable sort.
Selections are stored by device id instead of list position, so the filter
and hotplug cannot move a saved choice. Indexes saved by older versions are
migrated by device name; a missing device falls back to the default.
The same preferences change records whether the voice mode was chosen by
the person: a profile that only carries the old push-to-talk default now
starts in voice activation, explicit choices are kept.
Replace the custom X25519/HKDF/ChaCha20 TCP layer with rustls (TLS 1.3 only,
ALPN vnox-lnex/3, no plaintext or fallback). Add a trust store (pin, trust on
first use, public CA), block changed fingerprints and identities, bind the
Ed25519 auth signature to the TLS exporter, add RU/EN trust dialog and
update the interop test and e2e voice tool.