fix(gui): run the CAPTCHA page in its own process (B195)

On secureblue, WebKit cannot start the sandbox for the verification page:
bubblewrap is kept from the journal socket, xdg-dbus-proxy fails, and
WebKit ends its process with g_error. The webview lived in pdfs-app, so
the whole app crashed after the sign-in (#29).

- gui: `pdfs-app --human-verification` shows the page in its own window,
  reads the URL on stdin and prints the token on stdout; its logs go to
  stderr
- gui: a child that dies or fails without a token ends the sign-in with a
  message that the page could not be opened, instead of "cancelled"
- i18n: translate the new message in every catalog
- docs: B195 in the bug ledger, the cause and a one-time workaround in
  TROUBLESHOOTING.md, the child process in ARCHITECTURE.md, changelog

crates/pdfs-gui/src/app/main.rs

@@ -202,12 +202,20 @@ impl Ui {
 }
 
 fn main() -> glib::ExitCode {
-    tracing_subscriber::fmt()
-        .with_env_filter(
-            tracing_subscriber::EnvFilter::try_from_default_env().unwrap_or_else(|_| "info".into()),
-        )
-        .init();
+    let verify = std::env::args().nth(1).as_deref() == Some(VERIFY_ARG);
+    let logs = tracing_subscriber::fmt().with_env_filter(
+        tracing_subscriber::EnvFilter::try_from_default_env().unwrap_or_else(|_| "info".into()),
+    );
+    // The verification window's stdout carries its token back to the app.
+    if verify {
+        logs.with_writer(std::io::stderr).init();
+    } else {
+        logs.init();
+    }
     i18n::init();
+    if verify {
+        return run_verification_window();
+    }
 
     // The tray launches the app with arguments ("--page locations",
     // "--confirm-stop"), which a running instance must receive too.

crates/pdfs-gui/src/app/pages/login.rs

@@ -150,15 +150,27 @@ pub(crate) fn wire_login(ui: &Rc<Ui>) {
         });
 
         // Same shape for human verification: only fires when the API gates the
-        // sign-in, and the worker is blocked on the token until it does.
+        // sign-in, and the worker is blocked on the token until it does. A
+        // window that never came up is told apart from one the user closed, so
+        // the result below can say why the sign-in stopped.
         let ui_hv = ui.clone();
+        let verification_failed = Rc::new(Cell::new(false));
+        let hv_failed = verification_failed.clone();
         glib::spawn_future_local(async move {
             if let Ok((url, token_tx)) = hv_req_rx.recv().await {
                 ui_hv
                     .login
                     .login_status
                     .set_text(&gettext("Complete the verification to continue…"));
-                prompt_human_verification(&ui_hv, &url, token_tx);
+                // Anything but a token drops the sender, which the worker reads
+                // as a cancelled sign-in.
+                match verify_human(url).await {
+                    Verification::Solved(token) => {
+                        let _ = token_tx.send(token);
+                    }
+                    Verification::Cancelled => {}
+                    Verification::Failed => hv_failed.set(true),
+                }
             }
         });
 
@@ -194,6 +206,9 @@ pub(crate) fn wire_login(ui: &Rc<Ui>) {
                     let _ = gio::spawn_blocking(service::enable_start).await;
                     refresh(&ui);
                 }
+                Err(_) if verification_failed.get() => ui.login.login_status.set_text(&gettext(
+                    "The verification page couldn't be opened, so the sign-in was cancelled. Your system may keep the web view's sandbox from starting; the troubleshooting guide explains what to do.",
+                )),
                 Err(e) => ui.login.login_status.set_text(&e),
             }
         });

crates/pdfs-gui/src/app/pages/verify.rs

@@ -1,4 +1,4 @@
-//! The human-verification (CAPTCHA) dialog.
+//! The human-verification (CAPTCHA) window.
 //!
 //! Proton gates logins it does not recognise — a new IP, a VPN exit — behind a
 //! CAPTCHA. The API answers the login with a challenge instead of a session, and
@@ -8,10 +8,26 @@
 //! The page is hosted in a `WebKitWebView` rather than handed to the system
 //! browser because it reports completion by posting a message to its host, not
 //! by redirecting: a browser has nowhere to post it back to.
+//!
+//! The webview runs in a child `pdfs-app` started with [`VERIFY_ARG`], not in the
+//! app itself. WebKit aborts its whole process when its sandbox cannot start, as
+//! on hardened systems that keep bubblewrap from the journal socket
+//! (`docs/BUGS.md` B195); in a child that costs the verification, not the app.
+//! The URL goes to the child on stdin, so it never shows in a process list, and
+//! the token comes back on stdout.
 
 use crate::*;
+use std::io::{BufRead, BufReader, Write as _};
+use std::process::{ExitStatus, Stdio};
 use webkit6::prelude::*;
 
+/// The argument that makes `pdfs-app` the verification window instead of the app.
+pub(crate) const VERIFY_ARG: &str = "--human-verification";
+
+/// Marks the token among whatever else lands on the child's stdout: GLib and
+/// WebKit write debug output there.
+const TOKEN_PREFIX: &str = "hv-token ";
+
 /// Bridge the verification page's `postMessage` to a handler this side can read.
 ///
 /// The page targets its embedding host (`window.parent`), which in a webview is
@@ -26,15 +42,100 @@ window.addEventListener('message', function (event) {
 });
 "#;
 
-/// Present the challenge and hand the solved token back through `token_tx`.
-///
-/// Closing the dialog without solving drops the sender, which the waiting login
-/// worker reads as a cancelled sign-in — the same contract as the 2FA prompt.
-pub(crate) fn prompt_human_verification(
-    ui: &Rc<Ui>,
-    url: &str,
-    token_tx: std::sync::mpsc::Sender<String>,
-) {
+/// How the verification window ended.
+#[derive(Debug, PartialEq, Eq)]
+pub(crate) enum Verification {
+    Solved(String),
+    /// Closed without solving.
+    Cancelled,
+    /// Never came up, or died: most likely WebKit could not start its sandbox.
+    Failed,
+}
+
+/// Show the challenge in the verification window and wait until it closes.
+pub(crate) async fn verify_human(url: String) -> Verification {
+    match gio::spawn_blocking(move || run_verification_child(&url)).await {
+        Ok(Ok(verification)) => verification,
+        Ok(Err(e)) => {
+            tracing::error!("cannot run the verification window: {e}");
+            Verification::Failed
+        }
+        Err(_) => Verification::Failed,
+    }
+}
+
+fn run_verification_child(url: &str) -> std::io::Result<Verification> {
+    let mut child = Command::new(own_exe())
+        .arg(VERIFY_ARG)
+        .stdin(Stdio::piped())
+        .stdout(Stdio::piped())
+        .spawn()?;
+    if let Some(mut stdin) = child.stdin.take() {
+        // A failed write means the child is already gone; its exit status says
+        // how. Dropping `stdin` closes it.
+        let _ = writeln!(stdin, "{url}");
+    }
+    // Stops reading at the token rather than at EOF: WebKit's helper processes
+    // inherit the pipe and may hold it open a moment after the child quits.
+    let token = child
+        .stdout
+        .take()
+        .and_then(|out| find_token(BufReader::new(out)));
+    let status = child.wait()?;
+    Ok(outcome(token, status))
+}
+
+/// This very binary, so a development build opens its own window. After an
+/// upgrade replaced the file, `current_exe` names a deleted path, and the
+/// installed `pdfs-app` is the next best.
+fn own_exe() -> PathBuf {
+    std::env::current_exe()
+        .ok()
+        .filter(|path| path.is_file())
+        .unwrap_or_else(|| PathBuf::from("pdfs-app"))
+}
+
+fn find_token(out: impl BufRead) -> Option<String> {
+    out.lines()
+        .map_while(Result::ok)
+        .find_map(|line| line.strip_prefix(TOKEN_PREFIX).map(str::to_owned))
+}
+
+/// A token wins over the exit status: the user solved the page, even if
+/// WebKit's teardown crashed after it.
+fn outcome(token: Option<String>, status: ExitStatus) -> Verification {
+    match token {
+        Some(token) => Verification::Solved(token),
+        None if status.success() => Verification::Cancelled,
+        None => {
+            tracing::error!(%status, "the verification window failed");
+            Verification::Failed
+        }
+    }
+}
+
+/// Run as the verification window: read the page's URL from stdin, show it, and
+/// print the token once the user has solved it. Closing the window exits
+/// without a token, which the app reads as a cancelled sign-in.
+pub(crate) fn run_verification_window() -> glib::ExitCode {
+    let mut url = String::new();
+    if std::io::stdin().lock().read_line(&mut url).is_err() || url.trim().is_empty() {
+        tracing::error!("{VERIFY_ARG} expects the verification URL on stdin");
+        return glib::ExitCode::FAILURE;
+    }
+    let url = url.trim().to_owned();
+
+    // Not unique: the app that started it already holds APP_ID.
+    let app = adw::Application::builder()
+        .flags(gio::ApplicationFlags::NON_UNIQUE)
+        .build();
+    app.connect_activate(move |app| build_verification_window(app, &url));
+    // GApplication would refuse VERIFY_ARG as an unknown option.
+    let argv0 = std::env::args().next().unwrap_or_default();
+    app.run_with_args(&[argv0])
+}
+
+fn build_verification_window(app: &adw::Application, url: &str) {
     let content = webkit6::UserContentManager::new();
     content.add_script(&webkit6::UserScript::new(
         BRIDGE,
@@ -53,24 +154,19 @@ pub(crate) fn prompt_human_verification(
         .hexpand(true)
         .build();
 
-    let dialog = adw::Dialog::builder()
+    let body = gtk4::Box::new(gtk4::Orientation::Vertical, 0);
+    body.append(&adw::HeaderBar::new());
+    body.append(&webview);
+    let window = adw::ApplicationWindow::builder()
+        .application(app)
         .title(gettext("Verification"))
-        .content_width(420)
-        .content_height(560)
+        .icon_name(APP_ID)
+        .default_width(420)
+        .default_height(560)
+        .content(&body)
         .build();
 
-    let header = adw::HeaderBar::new();
-    let body = gtk4::Box::new(gtk4::Orientation::Vertical, 0);
-    body.append(&header);
-    body.append(&webview);
-    dialog.set_child(Some(&body));
-
-    // Taken on the first success *or* on close, so the token is sent at most
-    // once and a cancelled dialog always drops the sender.
-    let token_tx = Rc::new(RefCell::new(Some(token_tx)));
-
-    let tx = token_tx.clone();
-    let dlg_weak = dialog.downgrade();
+    let app = app.downgrade();
     content.connect_script_message_received(Some("hv"), move |_, value| {
         let Some(token) = extract_token(&value.to_str()) else {
             // Every `postMessage` on the page reaches here, most of them the
@@ -78,25 +174,17 @@ pub(crate) fn prompt_human_verification(
             // is simply not ours.
             return;
         };
-        if let Some(tx) = tx.borrow_mut().take() {
-            let _ = tx.send(token);
+        let mut stdout = std::io::stdout().lock();
+        if let Err(e) = writeln!(stdout, "{TOKEN_PREFIX}{token}").and_then(|()| stdout.flush()) {
+            tracing::error!("cannot hand the verification token back: {e}");
         }
-        if let Some(dlg) = dlg_weak.upgrade() {
-            dlg.close();
+        if let Some(app) = app.upgrade() {
+            app.quit();
         }
     });
 
-    // Covers the close button, Escape, and the user giving up.
-    dialog.connect_closed(move |_| {
-        token_tx.borrow_mut().take();
-    });
-
     webview.load_uri(url);
-
-    let parent = ui.login.login_button.root().and_downcast::<gtk4::Window>();
-    // Returns immediately; the login worker stays blocked on the channel until
-    // the handler above sends, or the dialog closes and drops the sender.
-    dialog.present(parent.as_ref());
+    window.present();
 }
 
 /// Pull the verification token out of a message posted by the verification page.
@@ -133,7 +221,12 @@ fn extract_token(raw: &str) -> Option<String> {
 
 #[cfg(test)]
 mod tests {
-    use super::extract_token;
+    use super::{Verification, extract_token, find_token, outcome};
+    use std::os::unix::process::ExitStatusExt as _;
+    use std::process::ExitStatus;
+
+    /// The wait status of a process killed by SIGABRT.
+    const ABORTED: i32 = 6;
 
     #[test]
     fn a_completion_message_yields_its_token() {
@@ -163,4 +256,39 @@ mod tests {
             assert!(extract_token(raw).is_none(), "accepted: {raw}");
         }
     }
+
+    #[test]
+    fn the_token_is_found_among_other_output() {
+        let out = "Gtk-Message: something\nhv-token tok-1\nmore\n";
+        assert_eq!(find_token(out.as_bytes()).as_deref(), Some("tok-1"));
+        assert_eq!(find_token("only chatter\n".as_bytes()), None);
+    }
+
+    #[test]
+    fn a_window_closed_without_a_token_cancels() {
+        assert_eq!(
+            outcome(None, ExitStatus::from_raw(0)),
+            Verification::Cancelled
+        );
+    }
+
+    /// WebKit aborts when its sandbox cannot start (B195). That must read as a
+    /// failure the app explains, not as the user giving up.
+    #[test]
+    fn a_window_that_died_fails() {
+        assert_eq!(
+            outcome(None, ExitStatus::from_raw(ABORTED)),
+            Verification::Failed
+        );
+        let exited_1 = ExitStatus::from_raw(1 << 8);
+        assert_eq!(outcome(None, exited_1), Verification::Failed);
+    }
+
+    #[test]
+    fn a_token_counts_even_if_the_window_died_after_it() {
+        assert_eq!(
+            outcome(Some("tok-1".into()), ExitStatus::from_raw(ABORTED)),
+            Verification::Solved("tok-1".into())
+        );
+    }
 }