[PalmarHealer]

Four independent changes, each usable on its own.

1. Program categories for process actions

A slider or button can target Browsers or Games instead of one named
executable, so a newly installed game or browser needs no config change.

Membership comes from three places:

• the bundled assets/app_groups.json (executable-name stems),
• path heuristics — anything under a Steam/Epic/GOG/Riot library counts as a
  game, whatever it is called. This is the part that keeps the category useful
  without an impossible list of every game ever made,
• the OS itself — browsers registered under StartMenuInternet, i.e. the same
  list the Windows "Default apps" page shows.

Users can extend any group (or add their own) with an app_groups.json in the
config directory; it is merged on top of the bundled file. Targets are stored as
@group:<id>; WASAPI resolves membership per audio session, using the process
image path as well as its name. Launchers and crash handlers are vetoed, so
steam.exe sitting inside the game library is not treated as a game.

2. Re-apply volumes after the default output device changes

Per-app volumes live on the audio session, and a session that follows the
default device to a new endpoint arrives carrying that endpoint's remembered
level — so the physical slider and the actual volume drift apart until the next
nudge. That jump is what this fixes.

The actuator remembers the last volume it set per target, polls the default
endpoint id (new AudioBackend::default_output_id) every 400 ms, and replays
twice after a switch (350 ms and 1500 ms) for streams that migrate late. Mute
state is deliberately untouched: switching outputs should not unmute anything.

Works for switches Slidr makes itself and for ones made in Windows.

3. Two new LED condition sources

• Music playing (src/media.rs) — the OS media session (WinRT SMTC), so it
  covers Spotify, browser tabs, and any player that publishes transport
  controls, with no login and no API key. Optional source filter; empty = any
  player.
• Discord (src/discord.rs) — self-mute, deafen, or in-a-voice-channel, read
  from Discord's local RPC socket. Nothing else reports Discord's self-mute:
  the OS mic endpoint stays unmuted while Discord mutes you internally.

Discord needs a one-off OAuth application (Client ID + Secret, entered under
Settings → LEDs; setup steps are in the module header). The secret is only read
until a token is cached, and SLIDR_DISCORD_CLIENT_SECRET can supply it instead
of the settings file. A half-typed Client ID never triggers a connection.

Both sources are served from a cache filled by a background worker, so the LED
tick never blocks on WinRT or a socket.

4. Tray bugfix

After an autostart --hidden launch the tray menu read "Minimize Slidr" while
clicking it opened the window. The label was hard-coded at construction time.

It is now seeded from the startup state and re-derived from the real window
state on every poll, so it cannot drift — including when the window is minimized
from the taskbar or restored outside our own code paths. Double-click (and the
menu entry) now also restore a window minimized to the taskbar, raising it to
the foreground.

Verification

• Windows: cargo check --all-targets clean, 25 tests pass.
• Linux (CI): cargo check passes — this is the first Linux compile of
  discord.rs's Unix-socket path, app_groups.rs's /proc walk and
  media.rs's non-Windows branch.
• Not verified: the interactive UI paths and the Discord flow against a real
  Discord client (that needs a personal OAuth application).

⚠️ CI is red, and not because of this branch

cargo test fails at the link step with unable to find library -lxdo.
That is main's pre-existing enigo configuration (the default xdo backend
needs libxdo.so at link time), which #9 fixes by switching to the pure-Rust
x11rb backend. Every branch off current main fails the same way.

Deliberately not worked around here: adding libxdo-dev to CI would paper over
the bug #9 fixes properly, and copying #9's one-line enigo change would conflict
with it.

Merge order: #9 first, then rebase this branch — it goes green on its own.
The only conflict between the two is in src/main.rs, where both branches
insert a block at the same spot after the settings load; the resolution is
keep-both.

Follow-up

default_output_id() has a None default, so change 2 is Windows-only for now
and Linux simply no-ops. The PulseAudio implementation is five lines on top of
#9's rewritten pulse.rs and is tracked there.

🤖 Generated with Claude Code