#947 highway-postgres OOM-killed inside its 4GiB cgroup (crash recovery, in-flight tasks failed)

EARS: The database container shall not be OOM-killed under the engine's maximum connection budget plus a full parallel test run; worst-case memory (shared_buffers + max_connections x work_mem x expected sort/hash nodes) shall fit the container limit with headroom >= 20%. FOUND 2026-10-16 01:24:09 UTC: dmesg 'Memory cgroup out of memory: Killed process 3362620 (postgres) anon-rss 66MB shmem-rss 480MB'; docker inspect highway-postgres OOMKilled=true, limit 4GiB; Postgres crash recovery; highway_internal tasks failed 'server closed the connection unexpectedly' and retried. Coincided with a full tests/unit run by an agent (parallel DB connections). Host swap 100% used (12221/12287 MB) across co-hosted stacks. Note #836 bounded the fleet connection budget; verify budget x per-conn memory vs 4GiB. Moving to pg-nano (#925) changes the limits; still must hold on this host meanwhile. Verify through a load test that reproduces then no longer reproduces.