ci(gui): close the driven window when the walk ends instead of idling to the smoke deadline (#376)

* ci(gui): close the driven window when the walk ends

The three injected drive legs waited out BDINFO_GUI_SMOKE_MS, a sleep the
app starts at BOOT rather than at the end of the walk, so each leg burned
roughly 100 s of pure sleep after its last assert.

Inject the platform close gesture instead — WM_CLOSE, xdotool windowclose,
System Events — and wait a 15 s grace. This needs no app code:
exit_on_close_request is off and close_requests() is subscribed
unconditionally, so the request routes through save-and-close and the
process exits 0 once its last window is gone.

The smoke deadline stays as the backstop with its original failure text, so
a close that does not land costs exactly what it costs now. Each leg logs
which path ended the app. The macOS gesture sits outside the geometry block,
which that leg skips whenever Accessibility refuses the lookup.

* docs(wasm): correct the published payload sizes

The figures had drifted from the build they describe: the optimized .wasm is
~535 KB, not ~501 KB, and the shipped JS ~51 KB, not ~44 KB. The tracked
wasm-size-budget.txt already recorded the larger figure, so the package
README contradicted it. Name the optimization level the numbers are measured
at, which is what makes them move.

Author: agentjp

.github/scripts/gui-drive-inject-linux.ps1

@@ -100,8 +100,32 @@ $after = Get-FileHash (Join-Path $Gallery '04-settings-open.png') -ErrorAction S
 $landed = $before -and $after -and ($before.Hash -ne $after.Hash)
 Write-Host "==> settings-click pixel change: $landed"
 
-Write-Host '==> waiting for the smoke-deadline exit'
-$exited = $proc.WaitForExit($SmokeMs)
+# End the app on the walk's clock, not the boot clock — the same change as the
+# Windows leg, with xdotool's windowclose as this platform's close gesture.
+# BDINFO_GUI_SMOKE_MS is a sleep the app starts at BOOT, so waiting it out
+# burns ~100 s of pure sleep after the walk has finished. No new debug hook is
+# needed: `exit_on_close_request` is false and `iced::window::close_requests()`
+# is subscribed unconditionally, so an OS close request routes CloseRequested
+# -> close_with_geometry -> SaveAndClose -> window::close and the process exits
+# 0 once its last window is gone. The smoke deadline stays as the BACKSTOP, so
+# a close that does not land costs exactly what it costs today.
+$CloseGraceMs = 15000
+Write-Host '==> asking the window to close (xdotool windowclose)'
+# Best-effort: a refused close must fall through to the backstop, not abort the
+# run. `$ErrorActionPreference = 'Stop'` plus PowerShell 7.4's native-command
+# error mapping would otherwise make a non-zero xdotool exit terminating here,
+# unlike the walk's own xdotool calls, where a failure genuinely is fatal.
+try { & xdotool windowclose $wid } catch { Write-Host "!! xdotool windowclose failed: $_" }
+$exited = $proc.WaitForExit($CloseGraceMs)
+if ($exited) {
+    Write-Host '==> app ended by: the injected close'
+}
+else {
+    Write-Host '!! the injected close did not land — falling back to the smoke deadline'
+    Write-Host '==> waiting for the smoke-deadline exit'
+    $exited = $proc.WaitForExit($SmokeMs)
+    if ($exited) { Write-Host '==> app ended by: the smoke deadline' }
+}
 if (-not $exited) { Stop-Process -Id $proc.Id -Force; throw 'the app never hit its smoke deadline' }
 Write-Host ("==> app exit code {0}" -f $proc.ExitCode)
 # The app's per-launch diagnostics log rides along in the gallery.

.github/scripts/gui-drive-inject-macos.ps1

@@ -119,8 +119,37 @@ else {
     if (-not $landed) { $failures += 'the Settings click changed no pixels (missed or blocked)' }
 }
 
-Write-Host '==> waiting for the smoke-deadline exit'
-$exited = $proc.WaitForExit($SmokeMs)
+# End the app on the walk's clock, not the boot clock — the same change as the
+# other two legs, with System Events' close button as this platform's close
+# gesture. BDINFO_GUI_SMOKE_MS is a sleep the app starts at BOOT, so waiting it
+# out burns ~100 s of pure sleep after the walk has finished. No new debug hook
+# is needed: `exit_on_close_request` is false and
+# `iced::window::close_requests()` is subscribed unconditionally, so an OS
+# close request routes CloseRequested -> close_with_geometry -> SaveAndClose ->
+# window::close and the process exits 0 once its last window is gone.
+#
+# Deliberately OUTSIDE the `$geomOk` block above: that block is skipped
+# whenever the geometry lookup is refused, and this leg still wants its exit
+# shortened when the walk never ran. Best-effort throughout — Accessibility is
+# exactly what this leg exists to probe, so a refused close is expected here
+# and must fall through to the smoke deadline rather than fail the run.
+$CloseGraceMs = 15000
+Write-Host '==> asking the window to close (System Events)'
+try {
+    & osascript -e 'tell application "System Events" to click button 1 of window 1 of (first process whose name is "bdinfo-rs-gui")' 2>&1 |
+        ForEach-Object { Write-Host "    $_" }
+}
+catch { Write-Host "!! the close gesture was refused: $_" }
+$exited = $proc.WaitForExit($CloseGraceMs)
+if ($exited) {
+    Write-Host '==> app ended by: the injected close'
+}
+else {
+    Write-Host '!! the injected close did not land — falling back to the smoke deadline'
+    Write-Host '==> waiting for the smoke-deadline exit'
+    $exited = $proc.WaitForExit($SmokeMs)
+    if ($exited) { Write-Host '==> app ended by: the smoke deadline' }
+}
 if (-not $exited) { Stop-Process -Id $proc.Id -Force -ErrorAction SilentlyContinue }
 $code = if ($exited) { $proc.ExitCode } else { 'killed' }
 Write-Host "==> app exit code $code"

.github/scripts/gui-drive-inject-windows.ps1

@@ -13,8 +13,10 @@
 # probe with a visible effect). PrintWindow screenshots after every event are
 # the evidence; one hard automated assert (the Settings click must change the
 # rendered pixels vs the shot before it) proves input actually reached the
-# app. The app exits by the BDINFO_GUI_SMOKE_MS deadline — exit 0 means the
-# event loop survived the whole injected walk.
+# app. The walk ends by injecting the platform's close gesture and the app
+# exits 0 through its own save-and-close path, with the BDINFO_GUI_SMOKE_MS
+# deadline kept as the backstop — exit 0 either way means the event loop
+# survived the whole injected walk.
 #
 # Every click re-verifies GetForegroundWindow() is still the app first —
 # injected input lands wherever the OS says focus is, never click blind.
@@ -410,8 +412,34 @@ $after = Get-FileHash (Join-Path $Gallery '04-settings-open.png') -ErrorAction S
 $landed = $before -and $after -and ($before.Hash -ne $after.Hash)
 Write-Host "==> settings-click pixel change: $landed"
 
-Write-Host '==> waiting for the smoke-deadline exit'
-$exited = $proc.WaitForExit($SmokeMs)
+# End the app on the walk's clock, not the boot clock. BDINFO_GUI_SMOKE_MS is
+# a sleep the app starts at BOOT, so waiting it out burns ~100 s of pure sleep
+# after the walk has finished. The app is closable without any new debug hook:
+# `exit_on_close_request` is false and `iced::window::close_requests()` is
+# subscribed unconditionally, so an OS close request routes CloseRequested ->
+# close_with_geometry -> SaveAndClose -> window::close, and the process exits 0
+# once its last window is gone. WM_CLOSE is the same primitive Send-Click uses
+# on blocker windows above.
+#
+# The smoke deadline stays as the BACKSTOP: if the close does not land, this
+# falls through to the original wait and the original failure text, so a lost
+# close costs exactly what it costs today and reddens nothing new. Closing this
+# way also exercises the save-and-close path, which writes gui.conf into the
+# isolated config dir the walk provisioned — the log collection below reads the
+# same directory and is unaffected.
+$CloseGraceMs = 15000
+Write-Host '==> asking the window to close (WM_CLOSE)'
+[void][GuiInput]::PostMessage($hwnd, 0x0010, [IntPtr]::Zero, [IntPtr]::Zero)
+$exited = $proc.WaitForExit($CloseGraceMs)
+if ($exited) {
+    Write-Host '==> app ended by: the injected close'
+}
+else {
+    Write-Host '!! the injected close did not land — falling back to the smoke deadline'
+    Write-Host '==> waiting for the smoke-deadline exit'
+    $exited = $proc.WaitForExit($SmokeMs)
+    if ($exited) { Write-Host '==> app ended by: the smoke deadline' }
+}
 if (-not $exited) { Stop-Process -Id $proc.Id -Force; throw 'the app never hit its smoke deadline' }
 Write-Host ("==> app exit code {0}" -f $proc.ExitCode)
 # The app's per-launch diagnostics log (it names the debug hooks the boot

crates/bdinfo-rs-wasm/web/README.md

@@ -21,10 +21,11 @@ Firefox.
 npm i @bdinfo-rs/wasm
 ```
 
-The published payload is **~501 KB of WebAssembly + ~44 KB of JS**. Only the
-main-thread entry you import (~9 KB) loads up front; the scan Worker (~3 KB)
-and the wasm-bindgen glue (~32 KB) that hosts the `.wasm` are fetched lazily
-inside the Worker, and nothing past the entry loads at all until the first scan.
+The published payload is **~535 KB of WebAssembly + ~51 KB of JS**, measured on
+the optimized build (`wasm-opt -Oz`). Only the main-thread entry you import
+(~12 KB) loads up front; the scan Worker (~4 KB) and the wasm-bindgen glue
+(~34 KB) that hosts the `.wasm` are fetched lazily inside the Worker, and
+nothing past the entry loads at all until the first scan.
 
 ## Usage