Fono

A quality-of-life release: two more cloud providers, polish on the
"Pondering" pause UI, headless servers install themselves, and a handful
of papercuts gone.

Added

• Deepgram speech-to-text now actually works. Picking Deepgram in
  fono setup (or running fono use stt deepgram) had been broken
  since v0.8.0 — it offered the option but failed at startup. The full
  pipeline is now wired: both the batch endpoint and a real WebSocket
  for live dictation, with the newer Nova-3 model as the default
  (Nova-2 is still selectable for languages Nova-3 doesn't cover yet).
• Cartesia speech-to-text. Same story — was advertised, now
  delivered. Batch transcription via the ink-whisper family;
  realtime ink-2 will follow in a future release.
• Cartesia text-to-speech now picks a native voice per language.
  Speak Romanian, hear a Romanian voice; switch to English in the
  same session, hear an English voice. No more one-voice-fits-all.
• Auto-stop on silence is now wired end-to-end. If you enable
  "Auto-stop after pause" in the tray, dictation actually stops once
  you've been quiet for the configured time — previously the
  PONDERING label appeared but nothing committed.
• sudo fono install is friendlier on servers. Headless boxes
  (no graphical session, multi-user systemd target) are now detected
  and the systemd lane runs by default — no --server flag needed.
  A new --desktop flag forces the desktop lane on hosts that just
  look headless.
• Server installs auto-enable LAN sharing. fono install --server
  now turns on the Wyoming STT listener on port 10300 out of the box,
  probes that it actually bound, and prints the address so other
  machines on the LAN can discover it immediately. fono uninstall
  on a server also cleans up /var/cache/fono (multi-GB model
  blobs).
• A diagnostic VU bar. [overlay].volume_bar = "advanced" paints
  a dBFS-axis meter with reference ticks for your speaking level and
  the silence threshold — useful for tuning auto-stop without
  guesswork. The default simple bar is unchanged.

Changed

• The "PONDERING" pause indicator is consistent everywhere.
  • It now shows up on the assistant flow (F8) too, in the green
    assistant palette, with the same auto-stop behaviour as
    dictation.
  • It only appears when you've actually enabled auto-stop — no
    more PONDERING under your finger if you've opted out.
  • It works in live (streaming) dictation, not just batch.
  • It doesn't flicker on a single breath, chair creak, or mouse
    click during a real pause.
• Tray "Auto-stop after pause" presets reworked from
  Off / 0.8 s / 1.5 s / 3 s (chat-app numbers) to Off / 3 s / 5 s
  (prose-dictation numbers). Default stays Off.
• Tray "Visualization" picker now turns the VU bar on automatically
  for the Transcript style and off for the others — sensible default,
  still overridable from config.toml.
• fono hwprobe matches what the setup wizard actually picks.
  The recommendation table used to promise large-v3-turbo on
  CPU-only boxes that the wizard would then quietly downgrade. Now
  the report and the wizard agree.
• Hotkey reliability on Wayland. Switching the overlay style from
  the tray now takes effect on the very next hotkey press (no
  restart). GNOME 47's portal hotkey rejection is detected upfront so
  Fono falls back to gsettings/X11 instead of silently dropping
  presses.
• Local Whisper picks better defaults out of the box. Model names
  now resolve through a quality-tested quantization ladder
  (tiny → q5_1, small → q5_1, small.en → q8_0,
  large-v3-turbo → q8_0); CPU threads default to the physical core
  count, which doubles throughput on Zen 3 / Zen 4 SMT systems where
  the previous default over-subscribed logical threads.

Fixed

• Wayland overlay no longer steals keyboard focus, paints as an
  opaque rectangle, or lands top-left on GNOME / Mutter. The overlay
  now runs through a pluggable backend layer: native
  wlr-layer-shell on KDE / wlroots / COSMIC / Hyprland; X11 via
  Xwayland on GNOME (which doesn't implement layer-shell). Set
  FONO_OVERLAY_BACKEND=… to force a specific backend.
• PipeWire audio playback (pw-play) no longer fails on every
  assistant reply — the --raw flag was missing.
• LAN dictation against a Wyoming peer that advertises IPv6 no
  longer fails with EINVAL when the peer's first-listed address is
  a link-local. Discovery now prefers routable IPv4 / IPv6.
• History database rebuilds itself when it carries an older
  schema, instead of warning on every dictation.
• The dictation key held down while pausing no longer flips the
  overlay into PONDERING and (with auto-stop on) no longer ends the
  session out from under you.
• Shipped binaries no longer SIGILL on pre-VNNI / pre-AVX-512 CPUs.
  The release build inherited ggml's GGML_NATIVE=ON default, which
  appends -march=native to the C/C++ compile line. On the GitHub
  Actions Linux runner (AMD EPYC 7763, Zen 3) the C compiler's
  auto-vectoriser baked VPDPBUSD (AVX-VNNI) into the binary, causing
  immediate SIGILLs on users' Kaby Lake, 8th-gen Intel, and earlier
  laptops. The shipped binary now pins an explicit
  AVX2 / FMA / F16C / BMI2 baseline (Intel Haswell ≥ 2013, AMD
  Excavator ≥ 2015) via .cargo/config.toml, so what CI builds is
  what users download — regardless of which CPU GitHub puts in its
  runner pool. A/B verified on Lunar Lake: zero throughput loss
  (±7% noise) because ggml's hand-written VNNI kernels are separately
  gated by GGML_AVX_VNNI (also off by default), so -march=native
  was costing portability without delivering any actual VNNI speedup.
• Hotkeys work immediately after sudo fono install on
  GNOME-Wayland. The post-install autostart spawned the daemon via
  runuser -u $SUDO_USER, which inherited the sudo-scrubbed
  environment: DISPLAY=:0 was preserved but WAYLAND_DISPLAY,
  XDG_RUNTIME_DIR, and DBUS_SESSION_BUS_ADDRESS were not. With
  only DISPLAY set, the daemon's hotkey-backend detector picked
  the X11 listener, the GNOME-gsettings shim never ran, and F7 / F8
  fell through in every native-Wayland app — users only saw working
  hotkeys after logging out and back in (when the XDG autostart entry
  fired with a real session env). The installer now reconstructs the
  graphical-session env from /run/user/$(id -u) inside the spawn
  command — after the user-switch — so the first daemon launched by
  sudo fono install is identical to what the next-login autostart
  would have produced. Drive-by: shutdown_existing_daemon no
  longer panics with "Cannot start a runtime from within a runtime"
  when re-running install while a previous daemon is still alive.

Removed

• 14 inert config keys (the always-warm-mic flag, eight commit-tuning
  knobs, three session-budget knobs, and two more) — all of them
  were silently ignored at runtime. Defaults are unchanged.

Breaking

• [overlay].volume_bar is now "off" | "simple" | "advanced"
  instead of a boolean, and defaults to "off". Existing configs
  need a one-line edit: volume_bar = true → "simple",
  volume_bar = false → "off". The tray picker handles new
  installs automatically.

Full Changelog: https://github.com/bogdanr/fono/compare/v0.8.0...v0.8.1