2026.3.2

[Hankanman]

Went on a Bug hunt 🐛

Bug Fixes

• Sync states blocking event loop — sync_states ran database queries on the event loop, causing UI lag. Refactored to run all DB operations in the executor and pre-compute entity-area mappings to avoid O(n×m) lookups. Fixes sync_states blocks the event loop with synchronous DB operations #388 in fix: Move sync_states DB operations to executor thread #403
• Entity evidence identity check — Used truthiness check (not entity.evidence) instead of identity check (entity.evidence is False), incorrectly treating None (unavailable) as inactive. Now correctly distinguishes unavailable from inactive entities. Fixes Unavailable entities incorrectly classified as inactive #389 in fix: Distinguish unavailable entities from inactive #404
• Orphan cleanup missing tables — _cleanup_area_orphans and delete_area_data did not clean up IntervalAggregates, CrossAreaStats, NumericSamples, or NumericAggregates, leaving stale data after area removal. All tables are now scoped and cleaned. Fixes delete_area_data leaves orphaned records in multiple tables #390 in fix: Clean up all related tables on area/entity deletion #405
• Motion sensor type guard on reload — load_data did not verify that restored motion/sleep sensor entity IDs still had the correct InputType, allowing misclassified sensors to carry stale probabilities after reconfiguration. Added type guard validation. Fixes DB load overwrites motion sensor probabilities with stale values #391 in fix: Prevent DB load from overwriting motion/sleep sensor probabilities #406
• Prune timestamp not persisted — set_last_prune_time opened a session but never committed, so the prune timestamp was lost on restart, causing redundant full-table prunes on every cycle. Consolidated into a single transaction with the prune deletes. Fixes prune_old_intervals never persists the last-prune timestamp #392 in fix: Persist last-prune timestamp after pruning intervals #407
• Duplicate person validation — Adding the same person entity to multiple people slots was silently accepted, causing undefined behavior. Now rejects duplicates with a clear error. Fixes No duplicate person entity validation in config flow #393 in fix: Add duplicate person entity validation in config flow #408
• Unavailable entities still decaying — Entities in unavailable state (value is None) were still included in the probability calculation when decay was active, inflating occupancy probability. Now skipped regardless of decay. Fixes Unavailable + decaying entity treated as having positive evidence #394 in fix: Skip unavailable entities in Bayesian calculation regardless of decay #409
• combine_priors unclamped early return — When only one area had data, combine_priors returned its raw prior without clamping, potentially producing values outside the 1–99% range. Now applies clamp_probability() on all return paths. Fixes combine_priors early return bypasses probability clamping #395 in fix: Clamp probability in combine_priors early return path #410
• Analysis exception scope too narrow — Analysis pipeline only caught HomeAssistantError | OSError | RuntimeError, letting unexpected exceptions (e.g., KeyError, TypeError) crash the coordinator. Broadened to catch Exception with full logging. Fixes Analysis pipeline _run_step catches only 3 exception types #396 in fix: Broaden exception handling in analysis pipeline steps #411
• "nan" string not filtered — Entities reporting the string "nan" as their state were treated as valid, producing NaN in probability calculations. Added "nan" to the invalid states filter. Fixes Lowercase "nan" state not filtered, silently treated as inactive #397 in fix: Filter lowercase nan in entity state property #412
• Lightweight config update crash — A single area failing during _async_lightweight_config_update would abort the entire update loop, leaving remaining areas with stale configuration. Each area is now wrapped in its own try/except. Fixes No exception handling in lightweight config update path #398 in fix: Add error handling to lightweight config update path #413
• WAL checkpoint errors silently suppressed — contextlib.suppress on WAL checkpoint hid potentially actionable errors. Replaced with explicit try/except and warning-level logging with exc_info=True. Fixes Silent WAL checkpoint failure can produce incomplete backups #399 in fix: Log warning on WAL checkpoint failure during backup #414
• Draft corruption on validation failure — Config flow updated _area_config_draft before validating, so a validation failure left the draft in a partially-modified state. Now validates on a candidate copy first. Fixes Config flow draft corrupted before validation in area behavior step #400 in fix: Validate area behavior input before updating draft #415
• CONF_AREAS type safety — CONF_AREAS was assumed to be a list[dict] but could contain non-dict items after manual config edits or corruption. Added per-item type validation with warning logging. Fixes _get_areas_from_config has no type safety on CONF_AREAS value #401 in fix: Add type safety check for CONF_AREAS value #416
• AnalysisStatus raw strings — Analysis status was tracked with raw string literals ("idle", "running", etc.), risking typo bugs and making refactoring fragile. Replaced with AnalysisStatus enum values throughout. Fixes create_from_config_spec uses raw strings instead of AnalysisStatus enum #402 in fix: Use AnalysisStatus enum instead of raw strings #417

A big thank you to those who have chosen to sponsor my work!
You can sponsor me here

Full Changelog: 2026.3.1...2026.3.2

---

This discussion was created from the release 2026.3.2.