v0.16.0

Added

• Windows support (experimental). Fono now runs on Windows: a single
  fono.exe alongside the Linux and macOS builds, with the notification-area
  tray icon and its menu, push-to-talk dictation and the voice assistant
  (F7 and F8 by default, Escape to cancel), text typed straight into your
  apps with a clipboard fallback, the floating recording overlay, focused-app
  awareness, local speech-to-text, local text polishing, local text-to-speech
  and wake word, and every cloud provider. Like the macOS build there is one
  download, not a CPU/GPU choice: it uses your graphics card to speed up
  transcription when a driver is present and quietly falls back to the
  processor when it isn't — and it starts fine on a fresh machine or virtual
  machine that has no graphics driver yet. fono install copies the app into
  your user folder and starts it at login with no administrator prompt,
  fono uninstall reverses that while keeping your settings and history, and
  fono update downloads, verifies, and swaps in the new version in place.
  This is an early port: most of it was built and exercised on a remote
  Windows machine rather than daily-driven, so expect rough edges and please
  file an issue with whatever you hit. Nothing changes for Linux or macOS
  users — their binary is byte-for-byte identical, with no new dependencies.

• Glass Cortex overlay (opt-in, experimental). A new recording-overlay
  style that shows a live view of the on-device AI while it works — the model
  thinking through a reply, then speaking it back. Pick it as your
  visualisation style to try it; it is off by default and only has anything
  to show when you are running a local (on-device) model. It is still rough
  on real replies and is due for a rework, so treat it as a preview.

Fixed

• An interrupted model download no longer leaves a broken model behind
  (all platforms). If a model download was cut short — a dropped
  connection, the machine sleeping, or the app closing mid-download — Fono
  could keep the half-finished file and then fail on every later start with a
  "model corrupted or incomplete" error that never cleared on its own.
  Downloads now finish into a temporary file and are only moved into place
  once the whole file is present and its checksum verifies; a bad download is
  re-fetched automatically.

Full Changelog: v0.15.0...v0.16.0