fix(systemd): use ManagedOOMMemoryPressureLimit=100% for oomd exemption

The previous ManagedOOMPreference=omit directive requires systemd >=250
(Ubuntu 24.04+). The deploy host runs Ubuntu 22.04 / systemd 249, where
omit is silently ignored and the alternate mechanism (=auto on the
ancestor user@ service) crashes user@<UID>.service with 219/CGROUP on
startup.

ManagedOOMMemoryPressureLimit=100% on this unit works on 249: when oomd
evaluates kill candidates inside the user slice, the worker's own PSI
is compared against 100% and never qualifies. The user slice still has
its distro-default 50% PSI threshold that triggers oomd; we just opt
the worker out of being chosen as the victim.

Drop the now-noop ManagedOOMPreference=omit; the host is not slated for
upgrade.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: akhilnarang

systemd/dumpyarabot-worker.service

@@ -21,7 +21,21 @@ KillSignal=SIGTERM
 # so systemd lets an in-flight dump drain on SIGTERM instead of SIGKILLing
 # mid-extraction.
 TimeoutStopSec=7260
-ManagedOOMPreference=omit
+# systemd-oomd PSI kill exemption for systemd 249 (Ubuntu 22.04, deployed
+# on `dumps`). The user slice's distro default
+# (/usr/lib/systemd/system/user@.service.d/10-oomd-user-service-defaults.conf)
+# sets ManagedOOMMemoryPressure=kill with a 50% PSI threshold, which causes
+# oomd to kill the heaviest child cgroup whenever slice-wide pressure
+# exceeds 50% for 20s — the worker, every time, because it's the heaviest.
+#
+# We raise THIS unit's own pressure-limit to 100% so that when oomd
+# evaluates kill candidates, the worker's PSI is compared against 100% and
+# never qualifies. Sibling cgroups (bot, redis) keep normal selection
+# behaviour but in practice are too small to ever be selected.
+#
+# OOMPolicy=continue ensures the worker doesn't get marked failed if a
+# CHILD process (aria2c, 7zz) gets OOM-killed by the kernel.
+ManagedOOMMemoryPressureLimit=100%
 OOMPolicy=continue
 
 [Install]