Fono

Added

• Audio-visualisation overlay + live-dictation VU bar. A new
  waveform cargo feature (default-on, GUI-only) renders a 640-wide
  bottom-centre overlay panel during batch (push-to-talk) recording
  with a selectable style:

  • bars — scrolling RMS amplitude bars; bars glow brighter at
    higher amplitude.
  • oscilloscope — connected-line waveform from raw PCM samples,
    pre-scaled by 1.0 / WAVEFORM_AMPLITUDE_CEILING so a typical
    speaking voice fills a comfortable chunk of the panel; the
    overlay's 5000-sample (~300 ms) ring buffer scrolls slowly
    enough for individual cycles to be visible.
  • fft — real-input spectrum bars from a 4096-pt Hann-windowed
    FFT, aggregated into 300 display bins covering 0–3 kHz with a
    −20 … +30 dB normalisation. Bars are pixel-tiled (no AA gap)
    so the spectrum reads as a continuous gradient.
  • heatmap — rolling spectrogram (frequency on Y, time on X,
    magnitude as colour intensity), backed by a pre-blended pixel
    cache that scrolls leftward by one frame-width per FFT push so
    redraw is a straight blit.

  Configured via [overlay].waveform = true and
  [overlay].style = "bars" | "oscilloscope" | "fft" | "heatmap".
  The standalone overlay is visible during Recording, transitions
  to amber POLISHING while STT runs, and hides on completion or
  cancel.

• Live-dictation VU bar. When [interactive].enabled = true
  the live-dictation panel now grows a thin right-side vertical
  meter that tracks microphone level in real time
  ([overlay].volume_bar = true by default). Drives off the same
  OverlayCmd::AudioLevel pipeline as the bars waveform style,
  so users can see whether their voice is too quiet without
  interrupting the transcript.

• Smoother audio capture for visualisation + streaming. The
  Linux PulseAudio backend now invokes parec --latency-msec=20
  so PCM lands in small frequent chunks (~20 ms). Without this PA
  picked a default fragment of several hundred ms, which made the
  waveform overlay's RMS tail look frozen between chunks and added
  end-of-utterance latency to the streaming pipeline.

Changed

• The pre-existing [overlay] config block (which had unused
  enabled/position/opacity fields) is replaced in place with
  the new waveform / style / volume_bar shape. No other
  consumers existed in the workspace.

Added

• GPU-accelerated release variant. Releases now ship two
  binaries side-by-side: the default fono-vX.Y.Z-x86_64 (compact
  ~18 MB CPU-only build, NEEDED set of 4 universal glibc libs) and
  fono-gpu-vX.Y.Z-x86_64 (Vulkan-enabled ~60 MB build, additionally
  links libvulkan.so.1). Both built from the same source; only
  the accel-vulkan cargo feature differs. Distro packages
  (.deb / .pkg.tar.zst / .txz / .lzm) are CPU-only at this
  release; raw GPU binary + .sha256 ship as release assets.
  Per plans/2026-05-02-fono-cpu-gpu-variants-v1.md slice 1.
  CUDA / ROCm remain build-from-source-only; Vulkan covers ~80 % of
  NVIDIA / ~90 % of AMD perf at zero vendor lock-in.

• Build variant identification. fono doctor and the daemon
  startup log now report which variant is running (cpu /
  gpu). New fono::variant::Variant enum + VARIANT constant
  in crates/fono/src/variant.rs for runtime introspection (and
  for the upcoming GPU upgrade UX).

• Runtime Vulkan probe. fono doctor gains a "Compute backends"
  section that reports the host's Vulkan loader + physical device
  state (e.g. "Vulkan: detected (Intel(R) Iris(R) Xe Graphics,
  llvmpipe (LLVM 22.1.3, 256 bits))"). On a CPU-variant binary
  with a Vulkan-capable GPU detected, an upgrade hint points at
  the fono-gpu release asset. Implemented via ash runtime-loaded
  bindings (Entry::load() → dlopen("libvulkan.so.1")) so the
  CPU variant still has the strict 4-NEEDED-entry allowlist —
  libvulkan never appears in NEEDED. Module lives at
  crates/fono-core/src/vulkan_probe.rs behind the vulkan-probe
  feature; both fono and fono-update opt in. Slice 2 of
  plans/2026-05-02-fono-cpu-gpu-variants-v1.md.

• Auto-variant fono update. Every fono update invocation now
  probes Vulkan on the host and fetches the matching release asset:
  fono-vX.Y.Z-x86_64 when no usable GPU is present, or
  fono-gpu-vX.Y.Z-x86_64 when libvulkan + a physical device are
  available. CPU users on GPU-equipped hardware are switched to
  the GPU build on their next update; if they later move to a
  GPU-less machine, the next update switches them back. No CLI
  flag, no wizard prompt, no config knob — one decision in one
  place. fono_update::check now takes the running binary's
  current asset prefix and treats a prefix mismatch as "update
  available" even at the same version. Slice 3 of
  plans/2026-05-02-fono-cpu-gpu-variants-v1.md.

• Tray "Update for GPU acceleration" entry. On a CPU-variant
  build with a usable Vulkan host, the tray menu surfaces an
  explicit "Update for GPU acceleration" item that triggers the
  same auto-variant apply_update path. Hidden on GPU builds and
  on hosts without Vulkan. New fono_tray::TrayAction::UpdateForGpuAcceleration

  • GpuUpgradeProvider callback type.

• CI gate split. The Binary size & deps audit job now runs as
  a matrix (cpu, gpu), asserting both variants stay within their
  respective budgets and NEEDED allowlists. CPU: ≤ 20 MiB + 4-entry
  allowlist (unchanged). GPU: ≤ 64 MiB + 4-entry allowlist

  • libvulkan.so.1.

• fono install / fono uninstall self-installer. Run
  sudo fono install (or sudo ./fono-vX.Y.Z-x86_64 install from a
  fresh release-asset download) to install fono system-wide on a
  desktop: places the binary at /usr/local/bin/fono, drops a menu
  desktop entry, an /etc/xdg/autostart/fono.desktop entry so the
  daemon launches automatically on next graphical login, the icon,
  and shell completions. Add --server for a headless install
  instead: writes a hardened systemd unit at
  /lib/systemd/system/fono.service running as a dedicated fono
  system user, and enables-and-starts it immediately. --dry-run
  prints the planned actions without touching the filesystem on
  either mode. sudo fono uninstall reads the install marker
  written at install time and removes exactly the files it recorded;
  per-user config and history are never touched. fono doctor now
  reports the install state (self-installed desktop / server,
  package-managed, or ad-hoc on PATH).

What's Changed

• fix(update): distinguish cross-variant switch from version bump in messaging by @bogdanr in https://github.com/bogdanr/fono/pull/7

Full Changelog: https://github.com/bogdanr/fono/compare/v0.5.0...v0.6.0