bug(vision): find_elements_by_voice_description imports a class that does not exist and always returns [] #18270

Defect

autobot-backend/computer_vision/screen_analyzer.py find_elements_by_voice_description runs from npu_semantic_search import NPUSearchEngine. That module defines NPUSemanticSearch and get_npu_search_engine(), not NPUSearchEngine. The ImportError is caught by except Exception and the method returns [], so voice-guided element detection never runs. It also builds a fresh engine without initialize().

Once it is wired, cross_modal_search will raise EmbeddingUnavailableError for the image collection until #18217 records that collection's space. The caller has to surface that instead of returning [].

Acceptance criteria

• It uses get_npu_search_engine().
• A refusal reaches the caller, not [].
• A test fails on the current import.

Refs #18185, #18217.