v0.5.0 — le micro qu'on porte, la fenêtre qu'on lit, la carte qu'on prête

Local push-to-talk dictation for Fr-En code-switchers. Everything on your own silicon.

Réglages › Langues	Réglages › Décodage

Three fixes, each with a different cause than its symptom suggested.

• 🎤 Your Bluetooth mic is finally in the list. PortAudio on a PipeWire desktop enumerates only pulse and default — a bluez source is not an ALSA card, so the rescan the dialog already did could never have surfaced one. Sources now come from the sound server, with human labels (Bose QC Ultra 2 HP) and live hotplug.
• 🪟 Réglages scrolls instead of squashing. A flat layout past the window manager's cap compresses every row below its minimum, and word-wrapped labels print over each other. Now a sidebar of seven categories, each page in its own scroll area — and it carries its own stylesheet, because nothing in the package set one.
• ⚡ « Préférer le CPU » lends the GPU back. For when someone else wants the card. It never constructs the CUDA engine rather than building it and falling back: 0 MiB held when set before launch, against 2316 MiB before.

---

Sprint 39 — 2026-09-09 · 0.5.0 — le micro qu'on porte, la fenêtre qu'on lit, la carte qu'on prête

Two reports from one session on Bluetooth headphones: the mic picker did not
list the headset that pavucontrol showed plainly, and Réglages rendered as
overlapping text — "this feels 1996, not 2026". Neither cause was the one the
symptom suggested. Full write-up:
docs/research/2026-09-09-mic-enumeration-and-settings-ui.md.

Fixed

• The mic list now comes from the sound server, not PortAudio (new
  pulse.py). PortAudio here has only the ALSA host API, and on a PipeWire
  desktop that means it enumerates exactly pulse and default — before and
  after the rescan the dialog was already doing. A Bluetooth mic is a PipeWire
  node, never an ALSA card, so no amount of rescanning could ever surface it:
  the picker was offering a choice between two aliases for "whatever the desktop
  picked". Sources now come from pactl (JSON, with a text parser for pactl
  < 15), monitors filtered out, which also buys human labels
  (Bose QC Ultra 2 HP rather than
  alsa_input.pci-0000_00_1f.3-platform-sof_sdw.HiFi__hw_sofsoundwire_4__source)
  and hotplug with no rescan at all, since pactl is queried live.
• Réglages squashed instead of scrolling. One flat QVBoxLayout,
  thirty-five widgets, no scroll area, a natural height past 1000 px. The window
  manager caps the window, Qt compresses every row below its minimum, and
  word-wrapped labels — which need a definite width to know their height — print
  one clipped line over the next. Now a sidebar of seven categories, each page
  its own QScrollArea(widgetResizable=True).

Added

• « Préférer le CPU » — leave the GPU to whoever else needs it (#137). Not a
  fallback, a choice: when someone wants the card for training a model, a game
  or a render, dictation should step aside rather than compete for VRAM. The
  daemon was holding 2316 MiB while idle.

  The distinction that makes it work: with the preference on, the CUDA engine is
  never constructed. GpuEngine() preloads the CUDA libs and allocates VRAM
  in its constructor, so "build it, then fall back" would still take the memory
  the user is trying to free. Measured:

  when the preference is set	VRAM held by TuParles
  before launch	0 MiB — libcudart/libcublas never even mapped
  toggled mid-session	2266 → 186 MiB

  The 186 MiB is the CUDA primary context, which no library call can destroy
  short of exiting the process; the ~2.1 GB of model weights are genuinely
  returned. So the honest claim is "the weights come back immediately, the last
  ~190 MB at the next restart" — not "it frees the card".

  Lives in ResilientEngine, deliberately, rather than load_engine() returning
  a bare CPU engine: a bare engine has nowhere to notice the setting changing
  back, so the toggle would have been one-way until a restart. It also tracks
  why it is on CPU — "preference" vs "failure" — because un-falling-back a
  card that actually died would break every take after. tuparles transcribe
  honours it too on --device auto, ahead of the ctranslate2 probe (so the
  preferred path never imports it); an explicit --device cuda still wins,
  because a flag you just typed beats a preference you once set.

• Réglages has a look of its own (new theme.py). Nothing in the desktop
  package set a stylesheet, and this box resolves no Qt6 platform theme at all
  (it exports QT_QPA_PLATFORMTHEME=qt5ct, the Qt5 bridge, which Qt6 ignores) —
  so the dialogs were bare Fusion. They now bring their own paint, in the
  bubble's palette, derived for light and dark from the running QPalette.
  Hand-rolled rather than a theme dependency: identical on every box, and scoped
  per-window so it can never repaint the bubble or the ribbon, which paint
  themselves.

• Settings re-homed by meaning. The old "Microphone" section had accreted
  eight checkboxes about decode behaviour, tray animation and the clipboard.
  Now: Micro, Langues, Bulle, Écriture, Décodage, Vie privée, Dev.

Changed

• A chosen PipeWire source is opened by pinning PULSE_SOURCE and
  PIPEWIRE_NODE (libpulse honours the first; pipewire-alsa, which serves
  default, only the second) around the InputStream constructor — that is
  where the ALSA plugin connects; by start() the source is already chosen. The
  pin is scoped and restored, because a leaked pin at a headset that just walked
  out of range would break the fallback open too, turning one failure into two.
• Stored mic names resolve pulse source → PortAudio name → system default, so
  settings written before this keep working. No migration.
• Resolving a mic at take start calls pulse.source_names(), not
  list_sources(): marking which mic the desktop would have chosen costs a
  second pactl call, and the take-start path does not need to know. Measured
  ~19 ms → ~11 ms, of a ~54 ms Recorder.start() (n=5, 35–58 ms — small
  sample, wide spread; the stream open itself dominates either way).

Infra

• tests/conftest.py, at last: an autouse fixture points XDG_CONFIG_HOME
  and XDG_CACHE_HOME at a tmp dir for every test. The suite had been reading
  whatever box it ran on — survivable while no engine consulted a setting at
  construction, and not survivable the moment prefer_cpu reached
  ResilientEngine.__init__: ticking a box in Réglages could have turned a green
  suite red, the worst kind of flake because it looks like a code failure.
• ResilientEngine.engine_name now names the rung that actually decoded
  (whisper.cpp or qwen) instead of hardcoding QwenCpuEngine. That was a rare
  white lie while CPU meant "the GPU died"; with CPU as a deliberate choice it
  would have been the common case. One test asserted the lie and was corrected.
• tests/test_settings_layout.py imposes the reported 738×578 and asserts no
  visible widget on any page is laid out shorter than its minimumSizeHint()
  (offscreen has no window manager, so the test inflicts the cramp itself),
  plus mic id-vs-label round-tripping and cross-page save coverage. Pulse
  parsing, monitor filtering, the PortAudio fallback and the env scoping are
  covered from synthetic fixtures — a real Bluetooth address is
  hardware-identifying and stays out of the repo. 990 tests, ruff and mypy
  clean.

Doctrine

• "Free the resource" and "don't take the resource" are different features.
  A fallback path already existed and was useless here: falling back happens
  after the constructor has taken the VRAM. The lever was never the decode, it
  was construction — and the discriminating test is "the factory was never
  called", not "it decoded on CPU", which a built-then-abandoned GPU also passes.
• State the residue. ~190 MB of CUDA context survives a mid-session release
  and no API can reclaim it. Measuring it turned a clean marketing sentence into
  a true one, and the true one is still a good feature.
• "No devices found" is a question about who is doing the enumerating. The
  picker was not buggy; it was asking the wrong subsystem, and it had a rescan
  bolted on that could never have helped. Before hardening a retry, check that
  the thing you are retrying can ever return the answer.
• A missing icon is not a visible failure. Qt stylesheet url() resolves
  files and Qt resources only: a data: URI loads nothing, with no error, no
  warning and no broken-image box. Measured both ways rather than assumed
  (file SVG: 20 tick pixels; data URI: 0). When a declarative layer can fail
  silently, verify it by rendering and counting pixels, not by reading it.