[endoftheline818]

Some vacuums show "NO MAP" even though a saved map is already selected. Selecting that same map again restores the image. This change performs that recovery automatically when the vacuum is idle or charging, without interrupting cleaning or a return to the dock.

The fix belongs in HomeTrait, as requested in home-assistant/core#186678:

• Load the known selected map before fetching its image during single-map discovery and normal refresh. Reuse the existing map-loading delay, and avoid loading a map twice during multi-map discovery.
• Defer discovery while the device is busy. Translate a refused reload into RoborockDeviceBusy and let refresh read the current map without repeating the rejected command.
• Keep "NO MAP" placeholders out of the saved map cache, preserve the last good map during a temporary empty response, and remove placeholders cached by older versions.

Fixes #985. Related: home-assistant/core#185133 and home-assistant/core#186678.

Validation:

• The added recovery and cache tests reproduce the failures before their fixes and pass afterward.
• Full Linux test suite on Python 3.11: 1,645 passed, 53 expected failures, 11 real-broker tests excluded by the default configuration; 99 snapshots passed.
• Full Linux test suite on Python 3.14: 1,645 passed, 53 expected failures, 11 real-broker tests excluded by the default configuration; 99 snapshots passed.
• All pre-commit checks passed, including Ruff, mypy and codespell.
• Wheel and source distribution builds passed.

Tests use simulated device responses. A physical Q5 Pro has not been tested with this change.

AI assistance: OpenAI Codex helped prepare and test this change under the contributor's direction.

[swastiktawate]

This is a thorough fix. The recovery path is properly guarded — idle/charging states only, never while cleaning or returning — and the fallback to recover_missing=False on a busy device means a locked action degrades gracefully instead of failing the whole refresh. Filtering placeholder 'NO MAP' images out of both the in-memory and persisted cache is the right call, and the guard matrix plus the exactly-once reload assertion in the tests show the edge cases were thought through.