feat(ptt): full desktop backend ladder + missed-key-up watchdog (gen2 v0.9.3 follow-up)

Lands SDD-081..088 + SDD-092 implementations on top of v1.0.0-rc.3.
The cross-platform pieces — `AudioTransmitGate`, the per-platform
backend ladder, and the missed-key-up watchdog — are wired into the
audio engine lifecycle. Per-platform live verification on Windows
/ macOS / GNOME-Wayland reference hosts is the remaining work
(RR-PTT-001..006/008 in `release-readiness-go-nogo-record.md`).

`chanora_audio::ptt`
--------------------

  * `AudioTransmitGate` now owns an `Arc<AtomicBool>` plus a
    `tokio::sync::watch::Sender<bool>` (SAD-075 / SDD-089). The
    encoder feed reads the atomic on the hot path; the watchdog
    subscribes to the watch channel.
  * `MissedKeyUpWatchdog::spawn(gate, timeout)` watches the gate
    transitions and self-clears `transmit_active` if the
    `false -> true` lifetime exceeds the configured ceiling
    (DEC-028, default 30s). Two unit tests cover the timeout-fires
    and the no-fire-on-normal-release paths.

`chanora_audio::ptt_backends`
-----------------------------

  * `DesktopPttBackend` trait + `PttBinding` value type + `PttInputClass`
    enum + `PttBackendError` (SDD-081). `PttBinding` deliberately
    carries only `input_class` and an opaque `platform_key`
    string; raw key codes never appear in the type surface.
  * `select()` factory (SAD-071): runtime ladder evaluation per
    OS. Windows → Raw Input → low-level hook → Focused; macOS →
    Event Tap → Focused; Linux → GNOME-Wayland portal probe →
    Focused.
  * `FocusedPttBackend` (SDD-087): universal terminal fallback;
    integrates with the existing Flutter Listener-driven PTT.
  * `WindowsRawInputBackend` + `WindowsHookBackend` (SDD-083 /
    SDD-084): three-rung ladder evaluated once at engine start.
    Each backend runs a dedicated worker thread that holds the
    OS-level handle; `start`/`stop` lifecycle is honest. Live
    `RegisterRawInputDevices` / `SetWindowsHookEx` wiring is
    platform-verification work — the scaffolding lets the
    descriptor + watchdog + capability event be exercised
    end-to-end now.
  * `MacOSEventTapBackend` (SDD-085): two-rung ladder with
    explicit `PermissionState` (Granted / Denied / Undetermined).
    `Undetermined` resolves to `L0Focused` so capability
    advertising matches actual runtime behaviour even before
    Input Monitoring is granted. Live `CGEventTap` + `IOHIDCheckAccess`
    wiring is platform-verification work.
  * `LinuxGnomeWaylandBackend` (SDD-086): probes GNOME-on-Wayland
    via `XDG_SESSION_TYPE` + `XDG_CURRENT_DESKTOP`, then verifies
    the `org.freedesktop.portal.GlobalShortcuts` D-Bus interface
    is reachable by reading the `version` property over a
    blocking zbus session. Reports `gnome-wayland-portal` /
    `L2GlobalHoldToTalk`. Other Linux environments fall through
    to the universal Focused backend (DEC-025).

`chanora_audio::engine`
-----------------------

  * Engine now owns `transmit_gate: AudioTransmitGate` and
    threads a `flag_arc()` clone into the existing capture
    state for the cheap hot-path read. `set_transmit_active` /
    `transmit_active()` go through the gate so subscribers see
    every transition.
  * `start_audio` selects the highest-capability backend via
    `ptt_backends::select()`, calls `backend.start(gate, none())`,
    and spawns the watchdog. Both are released in `stop()` and
    on Drop.
  * New `engine.rebind_ptt(binding) -> PttBackendDescriptor`
    drives the binding-capture flow without restarting the engine.
  * New `engine.ptt_descriptor()` returns the privacy-safe
    descriptor for the initial UI render before the first
    capability event arrives.

`chanora_core`
--------------

  * Re-exports `PttBinding` + `PttInputClass`.
  * New `ChanoraSession::set_ptt_binding(binding)` — calls
    `audio.rebind_ptt` and broadcasts the freshly-published
    `SessionEvent::PttCapability` so the UI badge updates live.
  * New `ChanoraSession::ptt_descriptor()` for the initial render.

`chanora_bridge`
----------------

  * New `BridgePttInputClass` enum + `set_ptt_binding(input_class,
    platform_key)` async function. The `platform_key` string is
    opaque to the bridge and never logged.
  * New `ptt_descriptor()` async accessor returning the
    `(level, backend_id, bound_input_class)` triple.

Flutter
-------

  * `_AudioControls` now has a "Configure" button next to the
    capability badge; `_PttBindingCaptureDialog` captures the
    next key press (via `Focus.onKeyEvent`) or mouse side button
    (via `Listener.onPointerDown` filtered to button bitmasks
    `0x08` / `0x10`). The captured value is the platform-neutral
    `LogicalKeyboardKey.keyLabel` or `mouse-side-button:{button}`.
  * The dialog explicitly tells the user that the actual key
    value never leaves it (DEC-027).
  * New ARB keys: `pttConfigureAction`, `pttConfigureTitle`,
    `pttConfigurePrompt`, `pttConfigureWaiting`,
    `pttConfigureCaptured`, `pttConfigurePrivacyNote`,
    `pttConfigureSaveAction` (en + zh-Hans).

Dependencies
------------

  * `chanora_audio` adds (Linux only) `zbus = "5"` with the
    `tokio` runtime selector + `blocking-api` feature for the
    GlobalShortcuts portal probe.
  * `chanora_audio` adds `tokio` `test-util` to dev-deps for
    `start_paused` watchdog tests (the live watchdog tests use
    multi-threaded real time).

Verification
------------

  * `cargo test --workspace` with `CHANORA_DISABLE_KEYRING=1`:
    57 tests green (was 53). chanora_audio rises from 4 to 8.
  * `cargo deny check`: advisories ok, bans ok, licenses ok,
    sources ok.
  * `cargo about generate --offline`: regenerates
    `docs/security/license-inventory.{md,html}`. The crate count
    rises from 364 to 383 with the addition of the zbus tree.
  * `tools/dump_flutter_licenses.sh`: 94 packages, zero without
    LICENSE (unchanged).
  * `flutter analyze`: clean.
  * `cargo build -p chanora_bridge --release` + `flutter build
    linux --release`: clean Linux x86_64 bundle.

Documentation
-------------

  * `docs/release/release-readiness-go-nogo-record.md` flips
    RR-PTT-007 (missed-key-up watchdog) to Done with a pointer
    to the two passing unit tests; bumps to v0.9.4. Live
    per-platform traces (RR-PTT-001..005, RR-PTT-008) remain
    open and are blocked only on platform reference hosts.

Per-platform live verification (Raw Input registration, Event Tap
creation under granted permission, GlobalShortcuts CreateSession +
BindShortcuts) is queued for the platform owners' reference hosts
per `staged-release-plan.md`.
This commit is contained in:
EdisonJwa
2026-05-15 15:38:42 +08:00
parent 7b21916049
commit 5199e3d005
26 changed files with 3291 additions and 40 deletions
+233 -9
View File
@@ -1,20 +1,25 @@
//! Desktop Push-to-Talk capability model (SRS-195 / SRS-196 / SAD-071 /
//! SDD-081 / SDD-082).
//!
//! This module defines the typed `PttCapabilityLevel` enum and a
//! This module defines the typed `PttCapabilityLevel` enum, the
//! lightweight `PttBackendDescriptor` value the audio engine
//! publishes to upstream consumers so the UI can render the live
//! capability badge (SDD-091) and the release verification record
//! can carry per-platform evidence (SysDes-148).
//! publishes to upstream consumers, the `AudioTransmitGate` object
//! that owns the authoritative `transmit_active` flag (SAD-075 /
//! SDD-089), and the `MissedKeyUpWatchdog` task (SAD-079 / SDD-092
//! / DEC-028).
//!
//! Concrete platform backends (`WindowsRawInputBackend`,
//! `MacOSEventTapBackend`, `LinuxGnomeWaylandBackend`) are deferred
//! to a follow-up code milestone; this commit lands the trait shape
//! and the universal `FocusedPttBackend` constant value so the
//! current Flutter-side hold-to-talk widget reports its capability
//! honestly through the bridge.
//! `MacOSEventTapBackend`, `LinuxGnomeWaylandBackend`) live in the
//! sibling `backends` module; this file holds the cross-platform
//! pieces.
use core::fmt;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::Arc;
use std::time::Duration;
use tokio::sync::watch;
use tracing::{info, warn};
/// Detected runtime capability of the active desktop PTT backend.
///
@@ -121,6 +126,172 @@ impl Default for PttBackendDescriptor {
}
}
/// Owns the authoritative `transmit_active` AtomicBool plus a
/// `tokio::sync::watch` channel so subscribers (notably the
/// missed-key-up watchdog) can observe transitions without
/// polling. Per SAD-075 / SDD-089 this is the **only** mutator
/// path on `transmit_active`; the encoder feed and the platform
/// PTT backends call `set` and `load` exclusively.
///
/// Cheap to clone — internally an `Arc` over the atomic and the
/// watch sender.
#[derive(Clone)]
pub struct AudioTransmitGate {
inner: Arc<TransmitGateInner>,
}
struct TransmitGateInner {
flag: Arc<AtomicBool>,
tx: watch::Sender<bool>,
}
impl AudioTransmitGate {
/// Construct a new gate initialised to `initial`.
pub fn new(initial: bool) -> Self {
let (tx, _rx) = watch::channel(initial);
Self {
inner: Arc::new(TransmitGateInner {
flag: Arc::new(AtomicBool::new(initial)),
tx,
}),
}
}
/// Set the transmit flag. When the value changes, every
/// `watch::Receiver` returned by [`Self::subscribe`] observes
/// the new value.
pub fn set(&self, active: bool) {
// We always store on the atomic (it's the hot read path
// for the encoder feed) and we always send on the watch
// channel — the watch implementation deduplicates so a
// repeat `set(true)` does not wake subscribers.
self.inner.flag.store(active, Ordering::Relaxed);
let _ = self.inner.tx.send(active);
}
/// Read the current value. Cheap; no locking.
pub fn load(&self) -> bool {
self.inner.flag.load(Ordering::Relaxed)
}
/// Subscribe to value transitions. The returned receiver
/// observes every distinct transition (the watch channel
/// deduplicates same-value sends).
pub fn subscribe(&self) -> watch::Receiver<bool> {
self.inner.tx.subscribe()
}
/// Live shared handle to the underlying `AtomicBool` for hot
/// paths that need to consult the flag once per outbound audio
/// frame without going through the gate wrappers. Callers must
/// treat the returned Arc as read-only; the gate's `set`
/// remains the sole mutator (SAD-075).
pub fn flag_arc(&self) -> Arc<AtomicBool> {
self.inner.flag.clone()
}
}
/// Background task that clears `transmit_active` after a configured
/// timeout (SAD-079 / SDD-092 / DEC-028).
///
/// Subscribes to the gate's watch channel, notes the timestamp of
/// each `false -> true` transition, clears the timestamp on each
/// `true -> false` transition. If the `true` lifetime exceeds the
/// configured ceiling the task self-clears the gate to false and
/// emits a sanitised diagnostic record (no key data — DEC-027).
pub struct MissedKeyUpWatchdog {
handle: Option<tokio::task::JoinHandle<()>>,
}
impl MissedKeyUpWatchdog {
/// Default timeout per DEC-028 — 30 seconds.
pub const DEFAULT_TIMEOUT: Duration = Duration::from_secs(30);
/// Spawn the watchdog. The returned handle aborts the task on
/// Drop. The watchdog needs a tokio runtime context; the audio
/// engine spawns it inside `start_audio` where the bridge's
/// runtime is available.
pub fn spawn(gate: AudioTransmitGate, timeout: Duration) -> Self {
let handle = tokio::spawn(async move {
run_watchdog(gate, timeout).await;
});
Self {
handle: Some(handle),
}
}
}
impl Drop for MissedKeyUpWatchdog {
fn drop(&mut self) {
if let Some(h) = self.handle.take() {
h.abort();
}
}
}
async fn run_watchdog(gate: AudioTransmitGate, timeout: Duration) {
let mut rx = gate.subscribe();
// Track whether we are currently in a `transmit_active = true`
// window. We don't need the actual timestamp; we just need a
// bounded wait that wakes on either the gate going false or
// the timeout elapsing.
info!(
target: "chanora_audio",
timeout_secs = timeout.as_secs(),
"missed-key-up watchdog spawned"
);
loop {
// Wait for a transition.
if rx.changed().await.is_err() {
// Gate dropped; exit.
return;
}
let is_active = *rx.borrow_and_update();
if !is_active {
// Either the user released, or someone else cleared
// the gate. Either way the watchdog has nothing to
// do until the next press.
continue;
}
// The gate just went true. Wait for either a release
// transition or for the timeout to elapse.
let release_or_timeout = tokio::select! {
r = rx.changed() => r.map(|_| *rx.borrow_and_update()),
_ = tokio::time::sleep(timeout) => {
// Timeout — self-clear and emit the privacy-safe
// diagnostic record. The field set here is
// deliberately limited to the values the gen2
// sanitizer permits (DEC-027): the active
// capability descriptor is owned by the controller
// and is not in scope here, so we emit only the
// watchdog-relevant fact.
warn!(
target: "chanora_audio",
timeout_secs = timeout.as_secs(),
"missed-key-up watchdog fired; clearing transmit_active"
);
gate.set(false);
continue;
}
};
match release_or_timeout {
Ok(false) => {
// Normal release; loop and wait for the next press.
}
Ok(true) => {
// Edge case: same-value double-set. Watch
// deduplicates so this should not actually fire,
// but we keep waiting on the timeout for the
// next iteration.
}
Err(_) => {
// Channel closed; exit.
return;
}
}
}
}
#[cfg(test)]
mod tests {
use super::*;
@@ -158,4 +329,57 @@ mod tests {
let _: &str = d.backend_id;
let _: Option<&str> = d.bound_input_class;
}
#[test]
fn gate_set_and_load_roundtrip() {
let g = AudioTransmitGate::new(false);
assert!(!g.load());
g.set(true);
assert!(g.load());
g.set(false);
assert!(!g.load());
}
#[tokio::test(flavor = "current_thread", start_paused = true)]
async fn gate_subscribe_observes_transitions() {
let g = AudioTransmitGate::new(false);
let mut rx = g.subscribe();
g.set(true);
rx.changed().await.unwrap();
assert!(*rx.borrow_and_update());
g.set(false);
rx.changed().await.unwrap();
assert!(!*rx.borrow_and_update());
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn watchdog_clears_transmit_after_timeout() {
let g = AudioTransmitGate::new(false);
let _wd = MissedKeyUpWatchdog::spawn(g.clone(), Duration::from_millis(80));
// Give the watchdog a tick to subscribe before we press.
tokio::time::sleep(Duration::from_millis(20)).await;
g.set(true);
// Wait past the timeout. The watchdog runs on a separate
// worker so real-time elapses in parallel.
tokio::time::sleep(Duration::from_millis(200)).await;
assert!(!g.load(), "watchdog should have cleared transmit_active");
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn watchdog_does_not_clear_on_normal_release() {
let g = AudioTransmitGate::new(false);
let _wd = MissedKeyUpWatchdog::spawn(g.clone(), Duration::from_millis(80));
tokio::time::sleep(Duration::from_millis(20)).await;
g.set(true);
tokio::time::sleep(Duration::from_millis(30)).await;
g.set(false);
// Wait past what would have been the timeout.
tokio::time::sleep(Duration::from_millis(200)).await;
assert!(!g.load());
// The gate should still be `false` and the next press
// should re-arm the watchdog cleanly.
g.set(true);
tokio::time::sleep(Duration::from_millis(200)).await;
assert!(!g.load(), "watchdog should fire on the second press as well");
}
}