Migrate runtime persistence to SQLite and add hot-bot load testing

[pmbstyle]

What changed

• Migrates runtime persistence from MariaDB to an embedded SQLite database, including schema setup, migration tooling, server scripts, and operational documentation.
• Keeps persistence writes buffered and adds SQLite-aware queue and checkpoint handling for character state.
• Adds an isolated hot-bot load runner that provisions real bot characters, drives normal bot AI paths in shards, persists character state, and records AI, event-loop, database, and memory metrics.
• Adds cache invalidation for real-player visibility so bot behavior updates immediately on player login or logout.

Why

The server should run without a separately installed database service while retaining observable, durable persistence. The load runner establishes a repeatable baseline for scaling the active bot population and catches database or scheduling regressions before they reach a live world.

Validation

• npm test
• Syntax check across 669 JavaScript files.
• Isolated 10-minute soak with 500 hot bots: 1.1 million buffered persistence writes, zero database failures, 59 ms AI-tick p95, and 19 ms event-loop lag p95.

Current operating guidance

The hot AI and persistence path sustained 500 bots in isolation. Until the separate cold-population scheduler is further optimized under mixed live-world load, use 300 hot bots as the conservative production target and treat 400-500 as controlled capacity tiers.