diff --git a/.github/scripts/interrupt-install.ps1 b/.github/scripts/interrupt-install.ps1
new file mode 100644
index 0000000000..5bbff97d45
--- /dev/null
+++ b/.github/scripts/interrupt-install.ps1
@@ -0,0 +1,211 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# Windows counterpart of interrupt-install.sh: run install.ps1 and kill it partway
+# through, reproducing a user quitting the desktop app mid-install.
+#
+# Windows has no process groups (hence the app's windows_job.rs), so this kills the whole
+# process TREE: killing only the leader leaves uv/python children to finish the dep pass.
+#
+# Usage:
+#   pwsh -File .github/scripts/interrupt-install.ps1 -Marker 'studio deps' `
+#        -LogPath logs/install.log -InstallArgs '--tauri --no-torch --local'
+[CmdletBinding()]
+param(
+  [string]$Marker = '',
+  [string]$LogPath = 'logs/install.log',
+  [string]$InstallArgs = '',
+  [int]$KillAtSeconds = 900
+)
+
+$ErrorActionPreference = 'Continue'
+New-Item -ItemType Directory -Force -Path (Split-Path -Parent $LogPath) | Out-Null
+Set-Content -Path $LogPath -Value '' -Encoding utf8
+
+# Stand in for the desktop app, which writes this before spawning the installer
+# (install.rs). We kill install.ps1 directly, so without it #7490's marker is absent for an
+# unrelated reason -- exactly what the Windows legs reported. Both locations: Rust
+# hardcodes ~/.unsloth/studio, CI overrides UNSLOTH_STUDIO_HOME. Never cleared by design.
+foreach ($dir in @($env:UNSLOTH_STUDIO_HOME, (Join-Path $HOME '.unsloth\studio'))) {
+  if ([string]::IsNullOrWhiteSpace($dir)) { continue }
+  try {
+    New-Item -ItemType Directory -Force -Path $dir -ErrorAction Stop | Out-Null
+    Set-Content -Path (Join-Path $dir '.desktop-install-in-progress') -Value '' -ErrorAction Stop
+  } catch { Write-Host "[interrupt] could not seed install marker in ${dir}: $_" }
+}
+
+# Its own host, so stdout can be redirected to the log while we poll. That host is WINDOWS
+# PowerShell 5.1, not pwsh, with install.rs:325-339's exact flags: the only host a real
+# desktop install ever uses, while every other Windows job in .github runs install.ps1
+# under pwsh 7, leaving 5.1 behaviour (.NET Framework, OEM/ANSI console encoding, different
+# native-command and OSArchitecture reporting) covered by nothing. The driver stays under
+# pwsh; only the installer child and the repair re-run change.
+$argList = @(
+  '-NoLogo', '-NoProfile', '-NonInteractive',
+  '-WindowStyle', 'Hidden',
+  '-ExecutionPolicy', 'Bypass',
+  '-File', 'install.ps1'
+)
+if ($InstallArgs) { $argList += $InstallArgs.Split(' ') }
+$proc = Start-Process -FilePath 'powershell.exe' -ArgumentList $argList `
+  -RedirectStandardOutput $LogPath -RedirectStandardError "$LogPath.err" `
+  -PassThru -NoNewWindow
+Write-Host "[interrupt] installer pid=$($proc.Id) marker='$Marker' deadline=${KillAtSeconds}s"
+
+# Proof that the signal was DELIVERED, not merely attempted. The installer can also fail
+# on its own between the last HasExited check and Stop-Tree, and a natural failure carries
+# a non-zero exit code just like a kill does, so the exit status alone cannot separate the
+# two on Windows. Stop-Process throws on a process that has already gone, so this flag is
+# false exactly when there was nothing left to interrupt.
+$script:rootKilled = $false
+
+function Get-Descendants([int]$RootId) {
+  # Depth-first, deepest first. CIM gives the parent link Windows does not expose via
+  # process groups.
+  $ids = @()
+  foreach ($k in @(Get-CimInstance Win32_Process -Filter "ParentProcessId=$RootId" -ErrorAction SilentlyContinue)) {
+    $kid = [int]$k.ProcessId
+    $ids += Get-Descendants $kid
+    $ids += $kid
+  }
+  return $ids
+}
+
+function Stop-Tree([int]$RootId) {
+  # Snapshot the whole tree BEFORE killing anything. Once a parent is gone its children are
+  # orphaned and there is no ParentProcessId left to walk, so the walk has to happen first.
+  $descendants = @(Get-Descendants $RootId)
+  # Then the ROOT, ahead of its children. install.ps1 watches the child it is waiting on:
+  # in staging run 30424366953 it had already printed "unsloth studio setup failed (exit
+  # code -1)" by the time Stop-Process reached it. Killing children first therefore races
+  # the leader's own exit, and a leader that wins that race makes Stop-Process throw over
+  # an interruption the driver really did deliver, failing the leg for nothing. Dead first,
+  # it cannot react to anything, and it cannot respawn what we are about to kill either.
+  try {
+    Stop-Process -Id $RootId -Force -ErrorAction Stop
+    Write-Host "[interrupt] killed installer pid=$RootId"
+    if ($RootId -eq $proc.Id) { $script:rootKilled = $true }
+  }
+  catch { if ($RootId -eq $proc.Id) { Write-Host "[interrupt] installer pid=$RootId was already gone: $_" } }
+  foreach ($id in $descendants) {
+    try { Stop-Process -Id $id -Force -ErrorAction Stop; Write-Host "[interrupt] killed pid=$id" }
+    catch { }
+  }
+}
+
+# A leg can be aimed at either kind of phase the installer prints, and only one of them is
+# a line. install.ps1 prints "[TAURI:STEP] <name>" lines, while the dependency pass rewrites
+# ONE physical line with \r (install_python_stack.py:2499), so its sub-steps are
+# CR-separated SEGMENTS. Splitting on \r is what makes a sub-step's END observable at all.
+$SubRe = '\[[=-]+\]\s*\d+/\d+\s'
+
+function Get-PhaseLines([string]$Path) {
+  $raw = Get-Content -Path $Path -Raw -ErrorAction SilentlyContinue
+  if (-not $raw) { return @() }
+  return @(($raw -replace "`r", "`n") -split "`n")
+}
+
+function Get-LastPhase([string]$Path) {
+  $p = @(Get-PhaseLines $Path | Where-Object { $_ -match '^\[TAURI:STEP\]' -or $_ -match $SubRe })
+  if ($p.Count) { return $p[-1] }
+  return ''
+}
+
+# True when the phase the marker named is no longer the running one. A sub-step marker is
+# judged against the running sub-step, a step marker against the running step -- a step is
+# not "over" because the sub-steps beneath it advanced.
+function Test-MarkedPhaseOver {
+  if (-not $Marker) { return $false }
+  $lines = @(Get-PhaseLines $LogPath)
+  $subs = @($lines | Where-Object { $_ -match $SubRe })
+  if ($subs | Where-Object { $_ -match $Marker }) {
+    $last = Get-LastPhase $LogPath
+    return -not ($last -match $SubRe -and $last -match $Marker)
+  }
+  $steps = @($lines | Where-Object { $_ -match '^\[TAURI:STEP\]' })
+  if ($steps | Where-Object { $_ -match $Marker }) {
+    return ($steps[-1] -notmatch $Marker)
+  }
+  return $false
+}
+
+$killed = $false
+$reason = ''
+# Fifth-of-a-second slices, matching the POSIX driver: every phase label prints BEFORE its
+# work starts, so this delay IS the whole distance between the label and the signal, and it
+# is the only thing that can push the kill past the end of a short phase. It slept 500ms
+# while the comment claimed a fifth, so it carried 2.5 slices of overshoot the POSIX side
+# does not.
+for ($i = 0; $i -lt ($KillAtSeconds * 5); $i++) {
+  if ($proc.HasExited) { $reason = 'exited-before-marker'; break }
+  if ($Marker) {
+    $hit = Select-String -Path $LogPath -Pattern $Marker -SimpleMatch:$false -ErrorAction SilentlyContinue
+    if ($hit) {
+      # Same as the POSIX driver: signal at detection, never after a delay. The label
+      # prints before the work, so the kill is inside the phase the moment the line
+      # appears, and any wait is a bet on the phase outlasting it that staging runs
+      # 30419729244 and 30426111484 both lost.
+      #
+      # The installer can still exit on its own between the match and the signal, which
+      # would record marker-hit over an install that interrupted nothing.
+      if ($proc.HasExited) { $reason = 'exited-before-signal'; break }
+      $reason = 'marker-hit'
+      $killed = $true
+      break
+    }
+  }
+  Start-Sleep -Milliseconds 200
+}
+if (-not $killed -and -not $proc.HasExited) { if (-not $reason) { $reason = 'deadline' }; $killed = $true }
+
+if ($killed) {
+  Write-Host "[interrupt] killing process tree of $($proc.Id) ($reason)"
+  Stop-Tree $proc.Id
+  # Any straggler uv/python that reparented away from the installer. The old sweep matched
+  # nothing: UNSLOTH_STUDIO_HOME arrives as `D:\a\r\r/.studio-home` (github.workspace joined
+  # with a forward slash) while Process.Path is all backslashes, so the literal -like missed
+  # even the venv's own python. Hence the separator normalisation, and uv by name (it lives
+  # outside the studio home, and the ephemeral runner has no other uv). Under --tauri there
+  # is no UNSLOTH_STUDIO_HOME, so fall back to install.ps1's root or the sweep only sees uv.
+  $studioRoot = if ([string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { Join-Path $HOME '.unsloth\studio' }
+                else { $env:UNSLOTH_STUDIO_HOME }
+  $homeNorm = if ([string]::IsNullOrWhiteSpace($studioRoot)) { $null }
+              else { ($studioRoot -replace '/', '\').TrimEnd('\') }
+  foreach ($p in @(Get-Process -Name 'uv', 'python', 'pythonw' -ErrorAction SilentlyContinue)) {
+    $path = $null
+    try { $path = $p.Path } catch { }
+    $inHome = $homeNorm -and $path -and ($path -like "$homeNorm\*")
+    if ($p.ProcessName -eq 'uv' -or $inHome) {
+      try { Stop-Process -Id $p.Id -Force; Write-Host "[interrupt] swept $($p.ProcessName) pid=$($p.Id)" } catch { }
+    }
+  }
+}
+
+try { $proc.WaitForExit(30000) | Out-Null } catch { }
+$rc = if ($proc.HasExited) { $proc.ExitCode } else { 'running' }
+Write-Host "[interrupt] installer exit=$rc reason=$reason killed=$killed root_killed=$($script:rootKilled)"
+Write-Host '[interrupt] last log lines:'
+Get-Content $LogPath -Tail 15 -ErrorAction SilentlyContinue
+
+if ($Marker -and -not (Select-String -Path $LogPath -Pattern $Marker -ErrorAction SilentlyContinue)) {
+  Write-Host "::warning::marker '$Marker' never appeared -- killed at the deadline, not the intended step"
+}
+# Where the signal actually landed. A phase that ended before the poll saw the marker sends
+# the kill into a LATER phase, and the leg then duplicates whichever leg owns that phase
+# while its own label claims otherwise.
+$lastPhase = Get-LastPhase $LogPath
+Write-Host "[interrupt] phase at kill: $lastPhase"
+$mismatch = Test-MarkedPhaseOver
+if ($mismatch) {
+  Write-Host "::warning::killed in '$lastPhase', not the marked phase -- that phase was already over"
+}
+# Lower-cased so the workflow can compare it the same way on every platform, and only
+# simple values: the POSIX side sources this file.
+@(
+  "interrupt_reason=$reason"
+  "interrupt_killed=$killed"
+  "interrupt_root_killed=$(if ($script:rootKilled) { 'true' } else { 'false' })"
+  "installer_exit=$rc"
+  "interrupt_phase_mismatch=$(if ($mismatch) { 'true' } else { 'false' })"
+) | Set-Content -Path (Join-Path (Split-Path -Parent $LogPath) 'interrupt.env') -Encoding utf8
+exit 0
diff --git a/.github/scripts/interrupt-install.sh b/.github/scripts/interrupt-install.sh
new file mode 100755
index 0000000000..7e73d12c27
--- /dev/null
+++ b/.github/scripts/interrupt-install.sh
@@ -0,0 +1,164 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# Run install.sh and SIGTERM it partway through, reproducing a user quitting the desktop
+# app mid-install: main.rs cleanup_child_processes() -> install::stop_install() kills the
+# installer PROCESS GROUP (install.rs:798-807). Group, not leader: `uv`/`python` children
+# would otherwise finish the dep pass and the leg would prove nothing.
+#
+# Usage: bash .github/scripts/interrupt-install.sh "<marker>" "<logfile>" [-- install args]
+#   <marker>  log regex to wait for before killing, e.g. "studio deps"; "" kills at deadline.
+# Env: KILL_AT_SECONDS deadline (default 900), KILL_GRACE grace before SIGKILL (default 10)
+set -uo pipefail
+
+MARKER="${1:-}"
+LOG="${2:-logs/install.log}"
+shift 2 || true
+[ "${1:-}" = "--" ] && shift
+KILL_AT_SECONDS="${KILL_AT_SECONDS:-900}"
+KILL_GRACE="${KILL_GRACE:-10}"
+
+mkdir -p "$(dirname "$LOG")"
+: > "$LOG"
+
+# The desktop writes this before spawning the installer (install.rs); we kill the installer
+# directly, so without it #7490's marker is absent for an unrelated reason. Both roots: Rust
+# hardcodes ~/.unsloth/studio, CI overrides UNSLOTH_STUDIO_HOME. Never cleared by design.
+for _marker_dir in "${UNSLOTH_STUDIO_HOME:-}" "$HOME/.unsloth/studio"; do
+  [ -n "$_marker_dir" ] || continue
+  mkdir -p "$_marker_dir" 2>/dev/null || continue
+  : > "$_marker_dir/.desktop-install-in-progress" 2>/dev/null || true
+done
+
+# Job control puts the child in its own group, so $! is the pgid and `kill -- -$!`
+# reaches every descendant, as the Rust side does.
+set -m
+bash install.sh "$@" > "$LOG" 2>&1 &
+PID=$!
+set +m
+echo "[interrupt] installer pid/pgid=$PID marker='${MARKER}' deadline=${KILL_AT_SECONDS}s"
+
+# A leg can be aimed at either kind of phase the installer prints, and only one of them is
+# a line. install.sh prints "[TAURI:STEP] <name>" lines, while the dependency pass rewrites
+# ONE physical line with \r (install_python_stack.py:2499), so its ten sub-steps are
+# CR-separated SEGMENTS. Splitting on \r is what makes a sub-step's END observable at all:
+# without it the "studio deps" leg of staging run 30419729244 killed at "7/10 data designer
+# deps" with backend_ok=true, having installed the structlog it exists to remove.
+SUB_RE='\[[=-]+\][[:space:]]*[0-9]+/[0-9]+[[:space:]]'
+phase_lines() { tr '\r' '\n' < "$LOG" 2>/dev/null || true; }
+
+# True when the phase the marker named is no longer the running one. A sub-step marker is
+# judged against the running sub-step, a step marker against the running step -- a step is
+# not "over" because the sub-steps beneath it advanced. A marker naming neither is not
+# judged. Results go through variables, never a `| grep -q`, which can report SIGPIPE
+# through pipefail on a long log.
+marked_phase_over() {
+  [ -n "$MARKER" ] || return 1
+  local lines steps subs last
+  lines="$(phase_lines)"
+  steps="$(printf '%s\n' "$lines" | grep -aE '^\[TAURI:STEP\]')" || true
+  subs="$(printf '%s\n' "$lines" | grep -aE "$SUB_RE")" || true
+  if [ -n "$subs" ] && [[ $subs =~ $MARKER ]]; then
+    last="$(printf '%s\n' "$lines" | grep -aE "^\[TAURI:STEP\]|$SUB_RE" | tail -1)" || true
+    [[ $last =~ $SUB_RE && $last =~ $MARKER ]] && return 1
+    return 0
+  fi
+  if [ -n "$steps" ] && [[ $steps =~ $MARKER ]]; then
+    last="$(printf '%s\n' "$steps" | tail -1)"
+    [[ $last =~ $MARKER ]] && return 1
+    return 0
+  fi
+  return 1
+}
+
+killed=false
+reason=""
+# Fifth-of-a-second slices: every phase label prints BEFORE its work starts, so the poll
+# delay is the whole distance between the label and the signal.
+for i in $(seq 1 $(( KILL_AT_SECONDS * 5 ))); do
+  if ! kill -0 "$PID" 2>/dev/null; then
+    reason="exited-before-marker"
+    break
+  fi
+  if [ -n "$MARKER" ] && grep -qE "$MARKER" "$LOG" 2>/dev/null; then
+    # Signal at detection, never after a delay. Every label prints BEFORE its work starts,
+    # so the kill is inside the phase the moment the line appears, and any wait at all is a
+    # bet on how long that phase runs. The bet loses: a flat 3s wait put 5 of the 12 legs
+    # of staging run 30419729244 into a LATER phase, and in 30426111484 it carried the
+    # macOS torch leg from "Installing PyTorch" into "Installing Unsloth" -- a step the
+    # workflow called minutes long finished in under three seconds.
+    #
+    # Between the grep and the signal the installer can still exit on its own, which would
+    # record marker-hit over an install that interrupted nothing.
+    if ! kill -0 "$PID" 2>/dev/null; then
+      reason="exited-before-signal"
+      break
+    fi
+    reason="marker-hit"
+    killed=true
+    break
+  fi
+  sleep 0.2
+done
+
+if [ "$killed" != "true" ] && kill -0 "$PID" 2>/dev/null; then
+  reason="${reason:-deadline}"
+  killed=true
+fi
+
+if [ "$killed" = "true" ]; then
+  echo "[interrupt] SIGTERM to process group -$PID ($reason)"
+  kill -TERM -- -"$PID" 2>/dev/null || kill -TERM "$PID" 2>/dev/null || true
+  for _ in $(seq 1 "$KILL_GRACE"); do
+    kill -0 "$PID" 2>/dev/null || break
+    sleep 1
+  done
+  # Unconditional, and to the GROUP: the leader exits on SIGTERM while a uv or python
+  # descendant does not, and gating on `kill -0 "$PID"` left that child finishing the dep
+  # pass under the probe. Signalling an empty group is a no-op.
+  echo "[interrupt] SIGKILL to process group -$PID"
+  kill -KILL -- -"$PID" 2>/dev/null || kill -KILL "$PID" 2>/dev/null || true
+fi
+
+wait "$PID" 2>/dev/null
+rc=$?
+
+# Only after the reap: an unreaped leader is still a member of its own group, so this
+# poll would report it alive forever. No installer may still run when the probe starts.
+if [ "$killed" = "true" ]; then
+  for _ in $(seq 1 "$KILL_GRACE"); do
+    kill -0 -- -"$PID" 2>/dev/null || break
+    kill -KILL -- -"$PID" 2>/dev/null || true
+    sleep 1
+  done
+  if kill -0 -- -"$PID" 2>/dev/null; then
+    echo "::warning::processes from installer group -$PID outlived SIGKILL"
+  fi
+fi
+echo "[interrupt] installer exit=$rc reason=$reason killed=$killed"
+echo "[interrupt] last log lines:"
+tail -15 "$LOG" || true
+
+# A leg that never reached the target step must be visible, not quietly green.
+if [ -n "$MARKER" ] && ! grep -qE "$MARKER" "$LOG" 2>/dev/null; then
+  echo "::warning::marker '$MARKER' never appeared -- this leg killed at the deadline, not at the intended step"
+fi
+# Where the signal actually landed. A phase that ended before the poll saw the marker sends
+# the kill into a LATER phase, and the leg then duplicates whichever leg owns that phase
+# while its own label claims otherwise.
+_last_phase="$(phase_lines | grep -aE "^\[TAURI:STEP\]|$SUB_RE" | tail -1)" || true
+echo "[interrupt] phase at kill: $_last_phase"
+mismatch=false
+if marked_phase_over; then
+  mismatch=true
+  echo "::warning::killed in '$_last_phase', not the marked phase -- that phase was already over"
+fi
+# Only simple values: the workflow sources this file, so the phase text stays out of it.
+{
+  echo "interrupt_reason=$reason"
+  echo "interrupt_killed=$killed"
+  echo "installer_exit=$rc"
+  echo "interrupt_phase_mismatch=$mismatch"
+} > "$(dirname "$LOG")/interrupt.env"
+exit 0
diff --git a/.github/scripts/interrupted_install_probe.py b/.github/scripts/interrupted_install_probe.py
new file mode 100644
index 0000000000..d302a38d2c
--- /dev/null
+++ b/.github/scripts/interrupted_install_probe.py
@@ -0,0 +1,313 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""After an install is interrupted, decide whether the desktop app WOULD report the
+resulting venv as healthy -- reproducing the Tauri preflight probes so the regression
+is testable without building the app.
+
+The reported bug: quitting the app during the dependency pass SIGTERMs the installer
+(install.rs stop_install). Landing in the "studio deps" step drops
+studio/backend/requirements/studio.txt, where structlog is declared. Preflight then
+probes `unsloth -h` (managed.rs:419) and `studio desktop-capabilities` (managed.rs:318);
+both SUCCEED because typer/click/rich are core, so the app reports ManagedReady with
+can_auto_repair=false and the backend dies on `import structlog`.
+
+ONE implementation for all three platforms. The Windows leg was briefly a bespoke inline
+PowerShell probe that ran only `-h` and `desktop-capabilities`, so it could not observe
+`studio_install_ok`, `verify-install` or `desktop-runtime-check` and would have failed
+the very PRs that add them. A probe that cannot see the fix is worse than no probe.
+
+Verdicts:
+  HEALTHY      the backend boots AND desktop-capabilities reports the install
+               complete -- preflight would report ManagedReady and be right
+  REPAIRABLE   the backend is broken AND a probe the DESKTOP consumes reports it,
+               so the app can offer a repair
+  FALSE_READY  the backend is broken and every probe says ready -> THE BUG
+
+Exit: 0 for HEALTHY/REPAIRABLE/NO_CLI, 1 for FALSE_READY, 2 for a usage error.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import socket
+import subprocess
+import sys
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+
+def run(cmd: list[str], timeout: int = 120) -> tuple[int, str, str]:
+    """Returns (rc, stdout, stderr). Kept SEPARATE: preflight pipes stdout and sends
+    stderr to /dev/null (managed.rs:358), so anything folded into stdout here is text
+    the desktop never sees."""
+    try:
+        p = subprocess.run(cmd, capture_output = True, text = True, timeout = timeout)
+        return p.returncode, p.stdout or "", p.stderr or ""
+    except (subprocess.TimeoutExpired, OSError) as e:
+        return 127, "", f"{type(e).__name__}: {e}"
+
+
+def merged(rc_out_err: tuple[int, str, str]) -> str:
+    """Both streams, for artefact logs only -- never for parsing."""
+    return rc_out_err[1] + rc_out_err[2]
+
+
+def has_subcommand(bin_path: str, args: list[str]) -> bool:
+    """Whether the CLI understands a subcommand at all: older builds lack the newer
+    verify commands, and 'absent' must not be read as 'reported failure'."""
+    rc, _, _ = run([bin_path, *args, "--help"], timeout = 60)
+    return rc == 0
+
+
+def free_port() -> int:
+    with socket.socket() as s:
+        s.bind(("127.0.0.1", 0))
+        return int(s.getsockname()[1])
+
+
+def main(argv: list[str]) -> int:
+    ap = argparse.ArgumentParser(description = __doc__)
+    ap.add_argument("bin", help = "path to the unsloth CLI")
+    ap.add_argument("--port", type = int, default = 0, help = "0 picks a free port")
+    ap.add_argument("--out", default = "probe", help = "directory for probe artefacts")
+    ap.add_argument("--boot-timeout", type = int, default = 120)
+    a = ap.parse_args(argv)
+
+    binp = a.bin
+    if not Path(binp).exists():
+        print(f"::error::unsloth bin not found: {binp}")
+        return 2
+    out = Path(a.out)
+    out.mkdir(parents = True, exist_ok = True)
+    port = a.port or free_port()
+    facts: dict[str, object] = {}
+
+    def say(k: str, v: object) -> None:
+        facts[k] = v
+        print(f"[probe] {k:28} = {v}")
+
+    # ── the two probes Tauri preflight actually runs ─────────────────────────
+    # The DESKTOP's deadline, not a generous CI one: preflight times each call out after
+    # 10s (managed.rs:337 for `-h`, :390 for desktop-capabilities) and reports Stale
+    # (managed.rs:471, :521). A longer timeout would call a slow torn venv HEALTHY and skip
+    # the re-run assertion; run() reports a timeout as a non-zero rc, the same REPAIRABLE arm.
+    PREFLIGHT_TIMEOUT = 10
+
+    t0 = time.time()
+    r = run([binp, "-h"], timeout = PREFLIGHT_TIMEOUT)
+    (out / "cli-h.log").write_text(merged(r), encoding = "utf-8", errors = "replace")
+    say("cli_h_ok", r[0] == 0)
+    say("cli_h_seconds", round(time.time() - t0, 2))
+
+    t0 = time.time()
+    caps_rc, caps_out, caps_err = run(
+        [binp, "studio", "desktop-capabilities", "--json"], timeout = PREFLIGHT_TIMEOUT
+    )
+    (out / "desktop-capabilities.json").write_text(caps_out, encoding = "utf-8", errors = "replace")
+    (out / "desktop-capabilities.stderr.log").write_text(
+        caps_err, encoding = "utf-8", errors = "replace"
+    )
+    say("capabilities_ok", caps_rc == 0)
+    say("capabilities_seconds", round(time.time() - t0, 2))
+
+    # Parse EXACTLY as the desktop does: managed.rs:414 hands the whole stdout buffer to
+    # serde_json, which rejects leading or trailing non-JSON, and stderr was already
+    # discarded at managed.rs:358. Folding stderr in made one warning line enough to fail
+    # the parse and report FALSE_READY over an install the real app offers to repair.
+    # "absent" (studio_install_ok predates the install-manifest work) and "unparseable"
+    # split only for a readable artefact: the desktop reports Stale for both
+    # ("desktop_capability_probe_failed", managed.rs:521). The field is Option<bool>
+    # (managed.rs:43), so serde rejects a non-boolean and the WHOLE payload fails to
+    # deserialize -> Stale; bool() instead read the JSON string "false" as True and
+    # reported HEALTHY over a torn install. Only a literal JSON true counts.
+    install_ok: object = "absent"
+    try:
+        parsed = json.loads(caps_out)
+        if not isinstance(parsed, dict):
+            install_ok = "unparseable"
+        else:
+            v = parsed.get("studio_install_ok")
+            if v is None:
+                install_ok = "absent"
+            elif isinstance(v, bool):
+                install_ok = v
+            else:
+                install_ok = "non-boolean"
+    except json.JSONDecodeError:
+        install_ok = "unparseable"
+    say("capabilities.studio_install_ok", install_ok)
+
+    # The desktop's own conclusion: Ready only on rc 0 + a parsed payload + a true
+    # studio_install_ok. The predicate is `!= Some(true)` (managed.rs:445), so an ABSENT
+    # field is Stale exactly like a false one; a CLI too old to answer is already rejected
+    # one check earlier on desktop_manageability_version. Leaving "absent" undecided
+    # reported HEALTHY on every booting leg and skipped the repair assertion -- the very
+    # regression `unsloth_cli/commands/studio.py` sits in the path filter to catch.
+    caps_ready = caps_rc == 0 and install_ok is True
+    say("desktop_would_call_install_ok", caps_ready)
+
+    # ── the deeper probes the fix PRs add ────────────────────────────────────
+    # RECORDED, but NOT repair evidence: preflight runs only `-h` and
+    # `studio desktop-capabilities --json` (managed.rs:357, :445), so counting these would
+    # let a leg pass while the real app still reports ManagedReady over a torn install.
+    for label, args in (
+        ("verify_install", ["studio", "verify-install"]),
+        ("desktop_runtime_check", ["studio", "desktop-runtime-check"]),
+    ):
+        if not has_subcommand(binp, args):
+            say(label, "absent")
+            continue
+        r = run([binp, *args], timeout = 300)
+        (out / f"{label}.log").write_text(merged(r), encoding = "utf-8", errors = "replace")
+        say(label, "ok" if r[0] == 0 else "failed")
+
+    # The in-progress marker #7490 writes before spawning the installer. RECORDED ONLY:
+    # both drivers seed it and never clear it, so it is true on every leg by construction,
+    # and using it in the verdict would make FALSE_READY, the one outcome, unreachable.
+    home = Path(os.environ.get("UNSLOTH_STUDIO_HOME") or (Path.home() / ".unsloth" / "studio"))
+    say("install_in_progress_marker", (home / ".desktop-install-in-progress").exists())
+
+    # ── ground truth: does the backend actually boot? ────────────────────────
+    # Own the whole process tree: the CLI spawns uvicorn/python children that would hold
+    # the port and hang the next leg's probe. Same reason the driver kills the group.
+    popen_kw: dict = {}
+    if os.name == "posix":
+        popen_kw["start_new_session"] = True
+    else:
+        popen_kw["creationflags"] = getattr(subprocess, "CREATE_NEW_PROCESS_GROUP", 0)
+    # Straight to the artefact file, never a PIPE: nothing drains a pipe until after the
+    # polling loop, so a backend whose imports outrun the OS buffer (64 KiB on Linux and
+    # macOS, one page on Windows) blocks BEFORE binding the port and backend_ok, what this
+    # verdict pivots on, would be false for a perfectly good install.
+    blog_path = out / "backend.log"
+    blog_fh = blog_path.open("w", encoding = "utf-8", errors = "replace")
+    # An interrupted install can leave the console script in place with its venv
+    # interpreter gone. An unguarded spawn would raise, so no verdict.json is written and
+    # both workflows die on json.load. An unlaunchable CLI is a broken backend `-h` flags.
+    proc = None
+    try:
+        proc = subprocess.Popen(
+            [binp, "studio", "--api-only", "-H", "127.0.0.1", "-p", str(port)],
+            stdout = blog_fh,
+            stderr = subprocess.STDOUT,
+            text = True,
+            **popen_kw,
+        )
+    except OSError as e:
+        say("backend_spawn_error", f"{type(e).__name__}: {e}")
+    backend_ok = False
+    deadline = time.time() + a.boot_timeout
+    while proc is not None and time.time() < deadline:
+        if proc.poll() is not None:
+            break
+        for path in ("/api/health", "/healthz"):
+            try:
+                with urllib.request.urlopen(f"http://127.0.0.1:{port}{path}", timeout = 2) as r:
+                    if r.status == 200:
+                        backend_ok = True
+                        break
+            except (urllib.error.URLError, OSError, TimeoutError):
+                pass
+        if backend_ok:
+            break
+        time.sleep(1)
+
+    def reap() -> None:
+        if proc is None:
+            return
+        if os.name == "posix":
+            import signal
+
+            # start_new_session made this child its own group leader. Read the pgid
+            # BEFORE the reap: once waited on, os.getpgid() raises and escalation targets
+            # nothing.
+            try:
+                pgid = os.getpgid(proc.pid)
+            except OSError:
+                pgid = proc.pid
+            for sig in (signal.SIGTERM, signal.SIGKILL):
+                try:
+                    os.killpg(pgid, sig)
+                except OSError:
+                    pass
+                try:
+                    proc.wait(timeout = 10)
+                    break
+                except subprocess.TimeoutExpired:
+                    continue
+            # Unconditional, and to the GROUP -- the same escalation
+            # interrupt-install.sh:94 makes. The leader exits promptly on SIGTERM while a
+            # uvicorn worker does not, so returning once proc.wait() succeeded left that
+            # worker holding the port and venv while the repair reinstalled underneath.
+            # Signalling an empty group is a no-op.
+            try:
+                os.killpg(pgid, signal.SIGKILL)
+            except OSError:
+                pass
+        else:
+            # On win32 the CLI re-spawns the server as a CHILD and waits on it
+            # (unsloth_cli/commands/studio.py:1543), and CREATE_NEW_PROCESS_GROUP does not
+            # make terminate() reach descendants, so killing the wrapper alone leaves the
+            # venv locked against the repair. taskkill /T takes the tree.
+            run(["taskkill", "/F", "/T", "/PID", str(proc.pid)], timeout = 30)
+            try:
+                proc.wait(timeout = 10)
+            except subprocess.TimeoutExpired:
+                proc.terminate()
+                try:
+                    proc.wait(timeout = 10)
+                except subprocess.TimeoutExpired:
+                    proc.kill()
+
+    reap()
+    blog_fh.close()
+    blog = blog_path.read_text(encoding = "utf-8", errors = "replace")
+    say("backend_ok", backend_ok)
+
+    missing = ""
+    for line in blog.splitlines():
+        if "ModuleNotFoundError" in line:
+            missing = line.strip()
+    if missing:
+        say("backend_error", missing)
+
+    # ── verdict ──────────────────────────────────────────────────────────────
+    # A booting backend is not enough. The manifest is written LAST
+    # (install_python_stack.py:3255), so the data-designer leg boots while
+    # desktop-capabilities still says studio_install_ok=false and preflight reports Stale
+    # (managed.rs:445); calling that HEALTHY skipped the re-run step.
+    # `-h` gates it for the same reason: probe_managed_bin runs it FIRST and returns Stale
+    # "cli_unusable" without reaching the capability probe (managed.rs:465-478), so
+    # consulting cli_h_ok only in the repairable arm called a help-less CLI HEALTHY.
+    if backend_ok and caps_ready and facts.get("cli_h_ok"):
+        verdict = "HEALTHY"
+    elif not caps_ready or not facts.get("cli_h_ok"):
+        verdict = "REPAIRABLE"
+    else:
+        verdict = "FALSE_READY"
+
+    facts["verdict"] = verdict
+    (out / "verdict.json").write_text(json.dumps(facts, indent = 2), encoding = "utf-8")
+    print(f"[probe] VERDICT = {verdict}")
+
+    if verdict == "FALSE_READY":
+        print(
+            "::error::Interrupted install reports READY but the backend cannot boot"
+            f" ({missing or 'import failure'}). Preflight sees -h ok + desktop-capabilities"
+            " ok, so the app shows ManagedReady with can_auto_repair=false and the user"
+            " is stuck."
+        )
+        return 1
+    if verdict == "REPAIRABLE":
+        print("[probe] incomplete install is detectable -> the desktop app can auto-repair")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main(sys.argv[1:]))
diff --git a/.github/workflows/interrupted-install-ci.yml b/.github/workflows/interrupted-install-ci.yml
new file mode 100644
index 0000000000..ca0d6db1ff
--- /dev/null
+++ b/.github/workflows/interrupted-install-ci.yml
@@ -0,0 +1,401 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+# Proves an INTERRUPTED install can never masquerade as a healthy one.
+#
+# Reported failure: quitting the app mid-install kills the installer process group
+# (main.rs cleanup_child_processes -> install.rs:798-807) mid "studio deps", the step
+# installing studio/backend/requirements/studio.txt where structlog is declared. On
+# relaunch preflight probes `unsloth -h` and `studio desktop-capabilities --json`; both
+# succeed because the CLI's own deps (typer/click/rich) are core, so the app reports
+# ManagedReady with can_auto_repair=false while the backend dies on `import structlog`
+# and the user is stuck on "Server stopped unexpectedly".
+#
+# No CI job had ever interrupted an install. This one kills the installer at each phase
+# and asserts the result is genuinely healthy or explicitly repairable, never silently ready.
+
+name: Interrupted install recovery
+
+on:
+  pull_request:
+    paths:
+      - 'install.sh'
+      - 'install.ps1'
+      - 'studio/setup.sh'
+      - 'studio/setup.ps1'
+      - 'studio/install_python_stack.py'
+      - 'studio/src-tauri/src/install.rs'
+      - 'studio/src-tauri/src/preflight.rs'
+      - 'studio/src-tauri/src/preflight/**'
+      - 'unsloth_cli/commands/studio.py'
+      # Every leg installs the checkout with `--local`, so this file decides the console
+      # script and the core deps: `-h` and `desktop-capabilities` only survive a torn
+      # install because typer/click/rich are declared here, not in an extra. No other
+      # install workflow interrupts the installer.
+      - 'pyproject.toml'
+      # studio_install_ok and verify-install, the decisions the probe asserts on, live
+      # here, so an install_state() that accepts a missing manifest would merge unrun.
+      - 'unsloth_cli/_studio_deps.py'
+      - 'studio/install_manifest.py'
+      # The requirement files are the phases: studio.txt declares structlog, whose absence
+      # IS the reported false-ready bug, so moving a package between them changes what
+      # every interrupted state looks like.
+      - 'studio/backend/requirements/**'
+      # `interrupt*-install*` would match the .sh / .ps1 but NOT the underscored probe
+      # (`*` never matches `/`, and a literal `-install` follows), so list all three.
+      - '.github/scripts/interrupt-install.sh'
+      - '.github/scripts/interrupt-install.ps1'
+      - '.github/scripts/interrupted_install_probe.py'
+      - '.github/workflows/interrupted-install-ci.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+env:
+  UNSLOTH_STUDIO_DISABLE_PUBLIC_CHECK: '1'
+
+jobs:
+  # ── macOS + Linux: kill at each phase ─────────────────────────────────────
+  interrupt:
+    name: ${{ matrix.os }} kill@${{ matrix.label }}
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          # Only a marker: the driver signals the moment that line appears, and no leg gets
+          # to wait first. Every label prints BEFORE its work starts, so the kill is inside
+          # the phase at detection, and a delay only bets on how long the phase runs. The
+          # bet lost twice, both times turning a leg into a duplicate of the next one: a
+          # flat 3s wait moved 5 of the 12 legs of staging run 30419729244, and in
+          # 30426111484 it carried the macOS torch leg into "Installing Unsloth" because
+          # the PyTorch step, called minutes long here, finished in under three seconds.
+          #
+          # Every leg is a hard gate. There is no continue-on-error cell: a leg allowed to
+          # fail is a warning wearing a red icon, and this workflow's entire claim is that
+          # a killed install cannot report itself healthy.
+          #
+          # The exact reported case: killed during the sub-step that installs structlog.
+          - {os: macos-14, label: studio-deps, marker: 'studio deps'}
+          # Coarse phases, earliest to latest -- each leaves a different partial venv.
+          # No venv cell. "Creating virtual environment" ran 0.107s in staging run
+          # 30419729244 (03:31:07.371 -> 07.478 to "Installing PyTorch"), shorter than any
+          # poll that watches the log, so the kill landed in the NEXT phase every time it
+          # was tried (30423181897 and 30424366953 both). It was the torch leg with a
+          # different label. Lost with it: a venv caught half-written. That state is not
+          # reachable by interruption at this resolution, and it is the only thing lost --
+          # the torch leg lands at the top of the PyTorch step, so what it leaves behind is
+          # already a complete venv with nothing installed into it.
+          - {os: macos-14, label: torch,       marker: '\[TAURI:STEP\] Installing PyTorch'}
+          - {os: macos-14, label: unsloth,     marker: '\[TAURI:STEP\] Installing Unsloth'}
+          - {os: macos-14, label: setup,       marker: '\[TAURI:STEP\] Running Unsloth setup'}
+          # Other dependency-pass sub-steps around the named one. No pip-bootstrap cell for
+          # the same reason as venv: "1/10 pip bootstrap" is over before a poll can see it,
+          # so in both 30419729244 and 30424366953 the signal landed in "2/10 unsloth
+          # extras", which is the next cell down.
+          # No base-packages cell: --local sets skip_base, so install_python_stack returns
+          # before "base packages" ever prints -- that leg ran to completion, proving nothing.
+          - {os: macos-14, label: unsloth-extras, marker: 'unsloth extras'}
+          - {os: macos-14, label: data-designer,  marker: 'data designer deps'}
+          # Linux: same teardown path, different package manager and process semantics.
+          - {os: ubuntu-latest, label: studio-deps, marker: 'studio deps'}
+          - {os: ubuntu-latest, label: torch, marker: '\[TAURI:STEP\] Installing PyTorch'}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Linux system deps
+        if: runner.os == 'Linux'
+        run: |
+          sudo apt-get update -qq
+          sudo apt-get install -y -qq --no-install-recommends cmake git build-essential libcurl4-openssl-dev
+
+      - name: Install, interrupted at "${{ matrix.label }}"
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          KILL_AT_SECONDS: '1500'
+        run: |
+          # --local is load-bearing. Without it install.sh:3996 resolves
+          # `unsloth>=2026.7.5` from PyPI, so the venv gets the PUBLISHED CLI, every
+          # `verify-install` / `desktop-runtime-check` probe reports "absent" whatever the
+          # branch does, and the lane cannot observe the fix it tests. --local overlays the
+          # checkout editable (install.sh:3990) before the dep pass, so a kill at "studio
+          # deps" leaves the branch's CLI installed.
+          bash .github/scripts/interrupt-install.sh \
+            '${{ matrix.marker }}' logs/install.log -- --tauri --local
+
+      - name: The kill must have landed where it was aimed
+        run: |
+          . logs/interrupt.env
+          echo "reason=$interrupt_reason killed=$interrupt_killed exit=$installer_exit"
+          if [ "$interrupt_reason" != "marker-hit" ]; then
+            echo "::error::installer never reached '${{ matrix.marker }}' (reason=$interrupt_reason)."
+            echo "::error::This leg proves nothing. Without this check it passes via the"
+            echo "::error::NO_CLI 'safe' path, which is how a --tauri/UNSLOTH_STUDIO_HOME"
+            echo "::error::conflict once made all 11 POSIX legs vacuously green."
+            tail -30 logs/install.log || true
+            exit 1
+          fi
+          # reason alone is not proof: the installer can finish between the driver's
+          # post-delay liveness check and the signal, leaving reason=marker-hit over a
+          # COMPLETED install that probes HEALTHY and skips the re-run assertion below. The
+          # exit status separates them: SIGTERM takes install.sh's trap to 143
+          # (install.sh:716) and SIGKILL to 137, while only a completed install exits 0.
+          if [ "$installer_exit" = "0" ]; then
+            echo "::error::installer exited 0 -- it COMPLETED inside the kill window, so"
+            echo "::error::nothing was interrupted and this leg asserts nothing."
+            tail -30 logs/install.log || true
+            exit 1
+          fi
+          # ...and it must have died from OUR signal rather than on its own. The installer
+          # can also fail naturally in the same window -- a dependency error exits 1 -- and
+          # that leg would test a broken installer while claiming to test an interrupted
+          # one. Recording whether `kill` returned 0 does not separate them: the leader is
+          # still an unreaped member of its own group, so signalling the group succeeds
+          # even when every process in it is already a zombie. The exit status does: every
+          # POSIX leg of staging run 30424366953 reported 143.
+          case "$installer_exit" in
+            143|137) ;;
+            *)
+              echo "::error::installer exited $installer_exit, which is neither SIGTERM"
+              echo "::error::(143, install.sh's trap at install.sh:716) nor SIGKILL (137)."
+              echo "::error::It died on its own, so this leg interrupted nothing."
+              tail -30 logs/install.log || true
+              exit 1
+              ;;
+          esac
+          # ...and the signal must land in the phase this leg is NAMED for. Warning-only
+          # left staging run 30419729244 fully green with the venv leg's install.log
+          # byte-identical to the torch leg's, and both "studio deps" legs killed past
+          # structlog (their own probe artefacts report backend_ok=true), so the flagship
+          # cell never reproduced the bug it is named after.
+          if [ "$interrupt_phase_mismatch" = "true" ]; then
+            echo "::error::the kill landed in a LATER phase than '${{ matrix.marker }}', so this"
+            echo "::error::leg duplicates whichever leg owns that phase and its label lies."
+            tr '\r' '\n' < logs/install.log | grep -aE '^\[TAURI:STEP\]|\[[=-]+\] *[0-9]+/[0-9]+' || true
+            exit 1
+          fi
+
+      - name: What state is the install in?
+        id: probe
+        run: |
+          # --tauri refuses a custom UNSLOTH_STUDIO_HOME, so it installs here.
+          BIN="$HOME/.unsloth/studio/unsloth_studio/bin/unsloth"
+          [ -x "$BIN" ] || BIN="$HOME/.unsloth/studio/bin/unsloth"
+          if [ ! -x "$BIN" ]; then
+            # No CLI at all is SAFE: preflight reports NotInstalled, the app reinstalls.
+            echo "verdict=NO_CLI" >> "$GITHUB_OUTPUT"
+            echo "[probe] no unsloth CLI installed -> preflight reports NotInstalled (safe)"
+            exit 0
+          fi
+          rc=0
+          python3 .github/scripts/interrupted_install_probe.py "$BIN" --out probe || rc=$?
+          v="$(python3 -c "import json;print(json.load(open('probe/verdict.json'))['verdict'])")"
+          echo "verdict=$v" >> "$GITHUB_OUTPUT"
+          exit "$rc"
+
+      - name: A re-run must repair, not short-circuit
+        # NO_CLI included: a kill at torch lands before "Installing Unsloth"
+        # (install.sh:2125 / :3667 / :3961), so those legs always take NO_CLI and skipping
+        # the re-run left three of them asserting only that a marker appeared. The bug's second half is `install.sh` seeing a "current" version and
+        # no-opping over a broken venv. HEALTHY also needs the install reported complete,
+        # so the data-designer leg (killed before the manifest is written last,
+        # install_python_stack.py:3255) arrives here instead of skipping that assertion.
+        if: always() && steps.probe.outputs.verdict != 'HEALTHY'
+        run: |
+          set -o pipefail
+          rc=0
+          bash install.sh --tauri --local < /dev/null 2>&1 | tee logs/repair.log || rc=$?
+          echo "repair exit: $rc"
+          BIN="$HOME/.unsloth/studio/unsloth_studio/bin/unsloth"
+          [ -x "$BIN" ] || BIN="$HOME/.unsloth/studio/bin/unsloth"
+          # No verdict.json when the bin is missing, so check here or json.load crashes.
+          if [ ! -x "$BIN" ]; then
+            echo "::error::after a full re-run there is still no unsloth CLI at $BIN"
+            tail -30 logs/repair.log || true
+            exit 1
+          fi
+          python3 .github/scripts/interrupted_install_probe.py "$BIN" --out probe-after || true
+          v="$(python3 -c "import json;print(json.load(open('probe-after/verdict.json'))['verdict'])")"
+          # A booting backend IS the repair, whatever the log narrated: judging by log
+          # text failed a leg whose venv was fine, matching only the frontend's "up to date".
+          if [ "$v" = "HEALTHY" ]; then
+            echo "re-run repaired the install (verdict=HEALTHY)"
+            exit 0
+          fi
+          echo "::error::after a full re-run the backend still does not boot (verdict=$v)"
+          if grep -qiE "(venv|dependenc|python stack)[^|]*(up to date|already current)" logs/repair.log; then
+            echo "::error::and the re-run treated the venv as current instead of repairing it"
+          fi
+          exit 1
+
+      - name: Upload logs
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: interrupted-${{ matrix.os }}-${{ matrix.label }}
+          path: |
+            logs/
+            probe/
+            probe-after/
+          retention-days: 7
+          if-no-files-found: warn
+
+  # ── Windows: no process groups, so the kill path differs ──────────────────
+  interrupt-windows:
+    name: windows kill@${{ matrix.label }}
+    # No UNSLOTH_STUDIO_HOME here. The app scrubs it (install.rs:202, :356) and
+    # install.ps1:189-215 rejects a custom root under --tauri, so a workspace-scoped root
+    # forced these legs down the shell-install path: UNSLOTH_TAURI_MODE=0, frontend build
+    # on, different root resolution, no bundled-file overlay. Worse, "Installing PyTorch"
+    # is only printed by Write-TauriLog (install.ps1:2440), so the torch leg's marker could
+    # never appear. The runner is ephemeral, so the default root is safe to install into.
+    runs-on: windows-latest
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          # install.ps1:121 parses `--no-torch`; `-SkipTorch` matches no case there and is
+          # silently dropped. The torch leg must NOT skip torch or its marker never appears.
+          - {label: studio-deps, marker: 'studio deps', installArgs: '--tauri --no-torch --local'}
+          - {label: torch, marker: 'Installing PyTorch', installArgs: '--tauri --local'}
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Install, interrupted at "${{ matrix.label }}"
+        shell: pwsh
+        run: |
+          pwsh -NoProfile -File .github/scripts/interrupt-install.ps1 `
+            -Marker '${{ matrix.marker }}' -LogPath logs/install.log `
+            -InstallArgs '${{ matrix.installArgs }}' -KillAtSeconds 1500
+
+      - name: The kill must have landed where it was aimed
+        shell: pwsh
+        run: |
+          $vals = @{}
+          foreach ($line in (Get-Content logs/interrupt.env)) {
+            $kv = $line -split '=', 2
+            if ($kv.Count -eq 2) { $vals[$kv[0]] = $kv[1] }
+          }
+          Write-Host "reason=$($vals['interrupt_reason']) killed=$($vals['interrupt_killed']) root_killed=$($vals['interrupt_root_killed']) exit=$($vals['installer_exit'])"
+          if ($vals['interrupt_reason'] -ne 'marker-hit') {
+            Write-Host "::error::installer never reached '${{ matrix.marker }}' (reason=$($vals['interrupt_reason']))."
+            Write-Host '::error::This leg proves nothing: without this check it passes via the'
+            Write-Host '::error::probe NO_CLI safe path, exactly as the POSIX legs once did.'
+            Get-Content logs/install.log -Tail 30 -ErrorAction SilentlyContinue
+            exit 1
+          }
+          # Same guard as the POSIX leg: the installer can finish between the driver's
+          # post-delay HasExited check and Stop-Tree, leaving reason=marker-hit over a
+          # COMPLETED install that probes HEALTHY and skips the re-run assertion.
+          # Stop-Process -Force is non-zero, so only a completed install reports 0.
+          # 'running' too, not just 0: the driver reports that when the installer outlived
+          # Stop-Tree and WaitForExit, and a live installer writing into the venv under the
+          # probe is not an interrupted install either. Only a real non-zero code counts.
+          if ($vals['installer_exit'] -eq '0' -or $vals['installer_exit'] -notmatch '^-?\d+$') {
+            Write-Host "::error::installer exit=$($vals['installer_exit']) -- it completed or never"
+            Write-Host '::error::died inside the kill window, so this leg asserts nothing.'
+            Get-Content logs/install.log -Tail 30 -ErrorAction SilentlyContinue
+            exit 1
+          }
+          # ...and the signal has to have been DELIVERED. A non-zero code is weaker proof
+          # here than on POSIX, where only a signal produces 143/137: install.ps1 failing
+          # on its own also exits non-zero, so a natural failure landing between the
+          # driver's last HasExited check and Stop-Tree would otherwise read as a kill.
+          # Stop-Process throws on a process that is already gone, so the driver records
+          # false exactly when it found nothing left to interrupt.
+          if ($vals['interrupt_root_killed'] -ne 'true') {
+            Write-Host '::error::the driver never terminated the installer -- it was already'
+            Write-Host '::error::gone when Stop-Tree reached it, so it failed on its own and'
+            Write-Host '::error::this leg interrupted nothing.'
+            Get-Content logs/install.log -Tail 30 -ErrorAction SilentlyContinue
+            exit 1
+          }
+          # Same landing check as the POSIX leg: a phase already over when the poll saw the
+          # marker means the kill hit a LATER phase, so the leg duplicates another one.
+          if ($vals['interrupt_phase_mismatch'] -eq 'true') {
+            Write-Host "::error::the kill landed in a LATER phase than '${{ matrix.marker }}', so this"
+            Write-Host '::error::leg duplicates whichever leg owns that phase and its label lies.'
+            ((Get-Content logs/install.log -Raw) -replace "`r", "`n") -split "`n" |
+              Where-Object { $_ -match '^\[TAURI:STEP\]' -or $_ -match '\[[=-]+\]\s*\d+/\d+\s' }
+            exit 1
+          }
+
+      - name: What state is the install in?
+        id: probe
+        shell: pwsh
+        run: |
+          # --tauri refuses a custom root, so this is where install.ps1:254-262 puts it.
+          $bin = Join-Path $env:USERPROFILE '.unsloth\studio\unsloth_studio\Scripts\unsloth.exe'
+          if (-not (Test-Path $bin)) {
+            "verdict=NO_CLI" | Out-File -FilePath $env:GITHUB_OUTPUT -Append -Encoding utf8
+            Write-Host '[probe] no unsloth CLI -> preflight reports NotInstalled (safe)'
+            exit 0
+          }
+          # The SAME probe the other platforms run. The bespoke inline version it replaced
+          # checked only `-h` and `desktop-capabilities`, so it could not observe
+          # studio_install_ok / verify-install / desktop-runtime-check and would have
+          # failed the very PRs that add them.
+          python .github/scripts/interrupted_install_probe.py $bin --out probe
+          $rc = $LASTEXITCODE
+          $v = (Get-Content probe/verdict.json -Raw | ConvertFrom-Json).verdict
+          "verdict=$v" | Out-File -FilePath $env:GITHUB_OUTPUT -Append -Encoding utf8
+          exit $rc
+
+      - name: A re-run must repair, not short-circuit
+        # Same assertion the POSIX legs make, NO_CLI included: without it a Windows leg
+        # proves only that the break was DETECTED, never that install.ps1's version fast
+        # path does not short-circuit over it, the half of the bug that strands the user.
+        if: always() && steps.probe.outputs.verdict != 'HEALTHY'
+        shell: pwsh
+        run: |
+          # powershell.exe with install.rs:325-339's flags, matching the interrupted run:
+          # the desktop repairs under Windows PowerShell 5.1, so a repair that only works
+          # under pwsh 7 would pass here and still strand the user.
+          powershell.exe -NoLogo -NoProfile -NonInteractive -WindowStyle Hidden `
+            -ExecutionPolicy Bypass -File install.ps1 ${{ matrix.installArgs }} *>&1 |
+            Tee-Object -FilePath logs/repair.log
+          $bin = Join-Path $env:USERPROFILE '.unsloth\studio\unsloth_studio\Scripts\unsloth.exe'
+          if (-not (Test-Path $bin)) {
+            Write-Host "::error::after a full re-run there is still no unsloth CLI at $bin"
+            Get-Content logs/repair.log -Tail 30 -ErrorAction SilentlyContinue
+            exit 1
+          }
+          python .github/scripts/interrupted_install_probe.py $bin --out probe-after
+          $v = (Get-Content probe-after/verdict.json -Raw | ConvertFrom-Json).verdict
+          # A booting backend IS the repair, whatever the log narrated: judging by log text
+          # failed a POSIX leg whose venv was fine, matching only the frontend's "up to date".
+          if ($v -eq 'HEALTHY') {
+            Write-Host 're-run repaired the install (verdict=HEALTHY)'
+            exit 0
+          }
+          Write-Host "::error::after a full re-run the backend still does not boot (verdict=$v)"
+          $log = Get-Content logs/repair.log -Raw -ErrorAction SilentlyContinue
+          if ($log -match '(?i)(venv|dependenc|python stack)[^|]*(up to date|already current)') {
+            Write-Host '::error::and the re-run treated the venv as current instead of repairing it'
+          }
+          exit 1
+
+      - name: Upload logs
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: interrupted-windows-${{ matrix.label }}
+          path: |
+            logs/
+            probe/
+            probe-after/
+          retention-days: 7
+          if-no-files-found: warn