2026.6.19 — MiniMax HTTP provider + door entry-gate

HACS update bundling everything since 2026.5.22 (the 2026.5.23, 2026.6.17 and 2026.6.18 changes ship here too — no separate releases were cut).

✨ Direct HTTP AI provider (MiniMax & any OpenAI-compatible API)

The LLM advisory can now call an OpenAI-compatible /chat/completions endpoint directly — no Home Assistant conversation agent required. Per-room Provider toggle (HA conversation agent or Direct HTTP) plus base URL / key / model (defaults https://api.minimax.io/v1, MiniMax-M3). Enables flat-rate / prompt-quota backends as a drop-in when a per-token agent (e.g. Gemini) hits a spending cap. Rooms sharing a backend still batch into one call; reasoning-model output (<think> blocks, ```json fences) is parsed correctly.

🚪 Door entry-gate (opt-in per room)

Mirror image of the door-closed lock-in: for a room with door contacts, presence may only mark it OCCUPIED if a door has opened since the room was last CLEAR. A closed door that never opened proves nobody entered → false PIR/mmWave triggers are rejected. Opt-in (require_door_entry, default off), fail-open at startup, never releases an already-occupied room.

🐛 Fixes

• LLM advisory froze after ~30 events (event-counter saturation) — now uses a monotonic counter.
• Restart/reconnect no longer fakes events — transitions out of unavailable/unknown (entity coming back online) are ignored, so a restored door reporting on is not logged as "door opened" (which had reset the lock-in / started a false score decay / would lift the entry-gate).
• Diagnostics, repair issues, config-entry unique_id, brand icon (from earlier unreleased versions).

Full details in CHANGELOG.md.