Vibecoded this after seeing models do amazing things but still drift on simple recursive steps; tracks exact match, answer accuracy, prefix correctness. Feedback welcome.
Linux page cache + PHP-FPM OPcache already keep hot PHP in RAM (no per-request disk hits after warm-up), and if your dataset fits in memory you size innodb_buffer_pool you don’t move the whole MariaDB datadir to tmpfs and throw away durability.