Streaming stack is on-demand, not always-on

Measured rather than assumed: the compositor and Sunshine together hold
220 MiB on the 3060 — sway ~112, Sunshine ~108. That card also runs the
local LLM, and with an 11.3 GB model resident on a 12 GB card the free
headroom goes from 992 MiB to 772. Losing a fifth of what is left matters
for KV cache growth, so leaving it running was the wrong default.

sway-headless and sunshine no longer start at boot. `stream on` brings them
up, `stream status` reports state and VRAM, and sunshine-idle-stop.timer
checks every five minutes whether any Moonlight client still holds a
connection to 47984/47989/48010 — stopping the stack after 15 idle minutes
so a forgotten session does not sit on memory the inference stack wants.

Note for future edits: `systemctl --user disable` on a unit dotbot has
symlinked into ~/.config/systemd/user deletes the symlink itself, because
systemd treats it as a linked unit. Re-run ./install after disabling one.

hosts/mainframe/systemd/sway-headless.service

@@ -18,5 +18,6 @@ ExecStopPost=/usr/bin/systemctl --user stop sunshine.service
 Restart=on-failure
 RestartSec=5s
 
-[Install]
-WantedBy=default.target
+# Deliberately NOT WantedBy=default.target. The 3060 also holds the local LLM,
+# and this stack costs ~220 MiB of VRAM. Start it with `stream on`; the
+# idle-stop timer puts it away again.