v0.31.0

A full code-review pass over the whole integration: 46 confirmed findings fixed, each adversarially verified, plus four agent-introduced regressions caught and corrected during a diff-verification pass. mypy --strict clean (58 files), ruff clean, 342 tests (was 328).

⚠️ Behaviour changes to be aware of

• WaterStop diagnostic sensor states changed. motor_state now reports off / rotate_to_closing / rotate_to_open (was off / on) and external_power reports supply / no_supply (was supply / battery / unknown), matching the real Ajax motorState / extPower API enums. Automations or history templates keyed on the old on, battery or unknown states will need updating.
• SQS alarm-notification default flipped. An unset notification-filter option now defaults to "all" instead of "none". If you never set a filter and currently receive no alarm notifications, you will start receiving them after this upgrade (only an explicit "none" suppresses them).
• SSE / proxy mode now produces alarm history and persistent notifications. Motion, smoke/fire, flood and glass-break events — and a door opened while armed — now append to a space's recent-events history and raise a persistent notification, matching SQS behaviour. Proxy-mode users will start seeing these.
• Devices survive a partial poll. A device missing from a single non-empty 200 response is no longer deleted immediately; it must be absent for 3 consecutive polls first. A truncated proxy response no longer wipes devices (stale devices may linger up to ~3 polls longer).
• Hub "external power" binary sensor now reads the dedicated externallyPowered field instead of inferring from battery.state != DISCHARGED; the reported state may change for some hubs and now reports even when no battery data exists.
• Video Edge, Video Edge sensor and firmware-update entities go unavailable during failed coordinator updates instead of showing stale values.

Added

• Live discovery of spaces and groups. A space or group added in Ajax after setup now creates its alarm-panel entity at runtime via the new SIGNAL_NEW_SPACE / SIGNAL_NEW_GROUP dispatcher signals — no integration reload required (mirrors the existing new-device path).
• SSE / proxy-mode alarm history & notifications for motion, smoke/fire, flood, glass-break and door-while-armed (parity with the SQS transport). Door-while-armed records history without forcing the space to TRIGGERED, mirroring SQS.
• arm_failed event message added to EVENT_MESSAGES in fr / en / es / de / nl / sv / uk (used by the unsuccessful-arming event code).
• "Unknown" label for the Video Edge storage-status sensor across strings.json and all 7 languages, matching the code's fallback value.

Fixed

• Cross-hub event suppression (multi-hub). Replaced the coordinator's single global _skip_state_change_event boolean with a per-hub _skipped_state_change_hubs set, so a metadata refresh on one hub no longer drops the REST-side arm/disarm state-change event for every other hub processed in the same tick.
• Night-mode unavailability (#149 class). A transient async_get_hub failure inside the per-hub loop left is_new_space / real_space_id / space_name unbound, raising UnboundLocalError and flipping the whole refresh to failure (all entities unavailable). The except handler now initialises them so the space degrades gracefully (created as NONE, or kept) instead of crashing the tick.
• Optimistic arm/disarm reverted by a stale poll. Space-level state updates now honour has_pending_ha_action (the same ha_protected guard the group path already had), so a REST poll landing inside the ~10 s window no longer snaps the panel back to its previous state after a user command.
• SSE second-event misattribution. The SSE manager now peeks the pending-HA-action flag (has_pending_ha_action) instead of consuming it, so when Ajax emits two state events for one HA command (e.g. arm then armwithmalfunctions) both stay attributed to Home Assistant rather than the second firing a misleading "armed/disarmed by " notification.
• Optimistic LED / brightness reverted by polling. Siren led_indication and dimmer actualBrightnessCh1 are now only overwritten from the API when not inside their optimistic-update guard; a poll within the TTL window no longer reverts the user's change. The siren blink-while-armed switch maps its optimistic write to led_indication (the attribute it actually reads), and multi-gang LightSwitch tracks the optimistic guard per channel so one channel's poll can't overwrite another's in-flight change.
• Cache bypass missed most of the refresh. bypass_cache_next() set a one-shot boolean consumed by the first request of a cycle (the hub list), so device and space getters still served stale data. It now opens a short 2 s window covering the whole refresh cycle, and the space getter (which feeds cameras and smart locks) honours it too — entity state after an SSE/SQS event or user action is now fresh, not TTL-lagged.
• Retry-After HTTP-date crashed the back-off path. A bare int(Retry-After) raised ValueError when a server sent an RFC 7231/9110 HTTP-date on a 429, crashing the retry exactly when asked to back off. A new _parse_retry_after() handles both seconds and dates, clamps to non-negative, and falls back to 60 s.
• Panel state not restored on a failed command. A failed arm/disarm/night-mode on space and group panels now captures and restores the previous state synchronously (async_write_ha_state) before the debounced refresh, instead of showing the wrong optimistic state during the debounce window.
• Night-mode arm switch never worked on LightSwitch / Transmitter. Both read the camelCase nightModeArm, but the coordinator only stores the snake_case night_mode_arm alias — so the switch was never created on LightSwitch and always read False on Transmitter. Both now use night_mode_arm.
• Siren alarm-duration showed an invalid state. alarmDuration is reported in minutes (not seconds); an earlier /60 produced an off-list 0. The select now maps 1:1 and snaps any off-list hub value to the nearest of 1/2/3/5/10/15.
• FireProtect2 CO siren-trigger switch snapped back. The CO switch read either CO/CCO token but the write only removed CO, so turning it off left the other token and the read flipped it back on. It now uses one token (prefer CO, else CCO) for both directions.
• Video Edge channel lookup returned the wrong channel. _get_channel_by_id now does two passes (explicit-id match, then positional fallback) so out-of-order channel ids — e.g. after a camera is removed — no longer resolve to a neighbour by index.
• Video Edge enum sensors emitted raw API values. Record-mode, record-policy and storage-status now fall back to the in-options unknown instead of value.lower() for unmapped enums. The boot-time sensor is normalised to whole minutes so it stops jittering by seconds every poll and inflating recorder history.
• Per-room device counts went stale. room.device_ids is now rebuilt every poll (it was append-only), so a device moved between rooms or deleted no longer inflates the per-room device_count shown in panel attributes.
• Shared door-sensor polling stopped too early. The shared polling task now only stops when no space is disarmed or in night mode, instead of being cancelled when any one space armed — other disarmed spaces keep getting timely door updates.
• Auth errors now drive reauth. AjaxRestAuthError is re-raised (not swallowed) in the video-edge and smart-lock updaters, so token expiry counts toward the reauth threshold instead of being logged and ignored.
• Orphaned ONVIF clients. On a partial ONVIF init failure, already-connected clients are stopped via async_stop() before the reference is dropped, preventing orphaned poll tasks / PullPoint subscriptions.
• Colliding notification IDs. Arm/disarm persistent-notification ids now include the space and source name, so concurrent events no longer overwrite each other in the notifications panel.
• Arming-failure event misclassified. Removed the M_22_21 → TRIGGERED override; it now resolves to RECOVERED via the odd/even hex heuristic, so an arming failure is no longer surfaced as a one-shot triggered state.
• ONVIF detection zone name dropped on secondary events. The rule/zone name is now threaded into _parse_event so every emitted detection event (including secondary AI object-detection, motion, line-crossing and doorbell events) carries it, not just the single returned event.
• Diagnostics SQS connectivity is derived from the live receiver thread and queue URL (the old is_connected() call didn't exist and always read False).
• Diagnostic dumps redacted. ajax_nvr_recordings.json and ajax_smart_locks.json now pass through async_redact_data, and the diagnostics TO_REDACT set additionally covers address, ssid, gateway, netmask, dns and networkInterface.
• Flood-detector malfunctions sensor simplified to the normalised int count (the list/string branches were dead code).

Changed

• Force-arm services route through the coordinator. force_arm / force_arm_night now call coordinator.async_arm_space / async_arm_night_mode(force=True) (instead of the raw API), apply across all matched config entries, and report aggregated per-hub failures (raising invalid_target when nothing resolves) — multi-account users can drive them across every hub.
• Diagnostic services span all entries. get_raw_devices, force-metadata-refresh, NVR-recordings and smart-lock dump handlers now iterate every matched config entry instead of only the first (they simply run for each; no failure aggregation).
• Night-mode fast-poll filter default flipped so door/transmitter/wire-input sensors that don't report night_mode_arm are now included in night-mode fast polling (faster updates for those sensors in night mode).

Removed

• devices/dimmer.py dead handler methods (get_switches / get_numbers / get_selects, ~280 lines): dimmer entities are built from static definitions in the platform files gated by is_dimmer_device(), so these were never invoked.
• AjaxDevice.is_triggered and last_notification: dead code (trigger state is tracked via per-attribute flags + last_trigger_time in the SSE/SQS managers).
• Per-door-sensor fast-poll mechanism (_fast_poll_tasks, _async_fast_poll_door_sensor and its teardown): superseded by the shared continuous door-sensor polling task.

Tests

• +14 regression tests (328 → 342). New coverage for: Retry-After date parsing and the cache-bypass window (test_api_helpers.py); the night-mode runtime-import and hub-details UnboundLocalError paths (test_coordinator_spaces_nightmode.py); the transmitter / siren / WaterStop / Video Edge handler fixes (test_device_handlers.py); the Video Edge channel id-vs-index lookup (test_models_extras.py).
• New ENUM-options ⇄ translation parity guard (test_translations.py): scans devices/*.py and fails if any literal options list has a key without a matching translated state entry in strings.json — the exact class of bug (WaterStop / Video Edge) caught in this pass.