fix: rojo died on SSH exit - launch via Task Scheduler, not Start-Process

Root cause: Windows OpenSSH kills the whole process tree when the SSH
session exits, so rojo started over SSH died seconds later. Proven:
1 proc inside the session, 0 from a fresh session.

Guard now calls 'schtasks /Run /TN RojoTWR' instead of Start-Process.
Verified unattended: killed rojo, scheduler restored it (Last Result: 0).

PROJECT-BRIEF.md

@@ -144,16 +144,30 @@ A scheduled task `RojoTWR` does exactly this — `schtasks /Run /TN "RojoTWR"`.
 (Launching via `start` over SSH does **not** work: the process dies when the SSH
 session exits. Use the task.)
 
-**Rojo does not stay up on its own.** It was observed dying repeatedly (Studio
+**Rojo does not stay up on its own.** The root cause was found on 2026-08-06:
-restarts, machine sleep, the one-shot task simply ending). A watchdog handles
+**Windows OpenSSH kills the entire process tree when the SSH session exits.**
-this now:
+Anything started over SSH with `Start-Process` — including Rojo — dies seconds
+later. This presents as "Rojo was running a minute ago and now the plugin won't
+connect," and it is *not* a plugin fault.
+
+Proof: `rojo procs: 1` inside the SSH session, `rojo procs: 0` from a new
+session moments later.
+
+The fix is to hand ownership to **Task Scheduler**, which is not part of the SSH
+process tree:
 
 ```
-%USERPROFILE%\rojo-guard.ps1          idempotent restart script
+schtasks /Run /TN "RojoTWR"           start rojo detached (survives SSH exit)
-schtasks /Query /TN "RojoTWR-Guard"   runs it every 3 minutes
+%USERPROFILE%\rojo-guard.ps1          idempotent watchdog
+schtasks /Query /TN "RojoTWR-Guard"   runs the watchdog every 3 minutes
 %USERPROFILE%\rojo-guard.log          only logs real events, silent when healthy
 ```
 
+Repo copy of the watchdog: `tools/win/rojo-guard.ps1`.
+
+> **Never start Rojo with `Start-Process`/`start` over SSH.** Always go through
+> `schtasks /Run /TN "RojoTWR"`. The guard itself follows this rule internally.
+
 The guard is deliberately conservative:
 - if `rojo` already owns 34872 it exits 0 and does nothing (safe mid-session —
   it will not disturb a live Studio connection)
@@ -162,8 +176,8 @@ The guard is deliberately conservative:
 - if `ServerPackages` is missing it logs that instead of letting Rojo silently
   exit on startup
 
-Verified by killing `rojo` and confirming the guard restored it with the port
+Verified end-to-end: killed `rojo`, waited, and the scheduler restored it
-owned by `rojo`.
+unattended (`Last Result: 0`, port owned by `rojo`).
 
 **Clicking Rojo → Connect in Studio is still manual and always will be.** The
 guard keeps the *server* alive; it cannot press the button. `Rojo_autoReconnect`

tools/win/rojo-guard.ps1

@@ -0,0 +1,38 @@
+# rojo-guard.ps1 -- keep Rojo alive for Tokyo Wangan Racing.
+# Idempotent: starts Rojo only if it is not already listening on 34872.
+# Safe to run every few minutes from Task Scheduler.
+
+$ErrorActionPreference = "SilentlyContinue"
+$repo = Join-Path $env:USERPROFILE "roblox\tokyo-wangan-racing"
+$exe  = Join-Path $env:USERPROFILE ".local\bin\rojo.exe"
+$port = 34872
+$log  = Join-Path $env:USERPROFILE "rojo-guard.log"
+
+function Log($m) { "$(Get-Date -f 'yyyy-MM-dd HH:mm:ss')  $m" | Add-Content $log }
+
+# Is something already LISTENING on the port?
+$listening = netstat -ano | Select-String ":$port\s.*LISTENING"
+
+if ($listening) {
+    $ownerPid = ($listening[0] -split "\s+")[-1]
+    $owner    = (Get-Process -Id $ownerPid).ProcessName
+    if ($owner -eq "rojo") { exit 0 }          # healthy, nothing to do
+    Log "PORT HIJACKED by '$owner' (pid $ownerPid) -- not starting. Kill it first."
+    exit 1                                      # sshd tunnel or similar: do NOT mask it
+}
+
+# Dependencies must exist or rojo exits instantly.
+if (-not (Test-Path (Join-Path $repo "ServerPackages"))) {
+    Log "ServerPackages MISSING -- rojo would exit on startup. Run 'wally install'."
+    exit 1
+}
+
+# Launch DETACHED via Task Scheduler.
+# Do NOT use Start-Process here: if this script is ever invoked over SSH,
+# Windows OpenSSH kills the whole process tree on session exit and rojo dies
+# seconds later. schtasks hands ownership to the scheduler, which survives.
+schtasks /Run /TN "RojoTWR" | Out-Null
+Start-Sleep -Seconds 4
+
+if (Get-Process rojo -EA SilentlyContinue) { Log "started rojo on $port" }
+else { Log "FAILED to start rojo" }