feat(search): rank search results by semantic similarity

[bilipp]

What

Search now actually uses the on-device NLContextualEmbedding vectors that the background ContentIndexer already stores in Movie.embeddingData / Series.embeddingData. Until now those vectors were generated and persisted but never read — TextEmbedder.decode had no callers and search was pure substring matching.

How

SemanticSearchService (new actor, singleton like ContentIndexingService)

• Lazily creates and prepare()s a TextEmbedder, reused across searches. If the model can't load on the device, it caches that fact and returns nil so search degrades gracefully — no repeated OTA download attempts.
• Embeds the query, then on a background ModelContext fetches every indexed title (embeddingData != nil, propertiesToFetch to avoid faulting whole rows), decodes each vector, and cosine-ranks with Accelerate (vDSP).
• Returns ordered (id, score) lists above a 0.35 similarity threshold, capped per content type.

SearchView — now hybrid:

• The existing localizedStandardContains (lexical) pass still runs and always leads the results, so exact-title hits, live channels (never embedded), and not-yet-indexed titles always appear.
• Semantic-only matches are appended, deduped against the lexical set, and ordered by score across movies + series.
• Staleness-guarded after the await so debounced/cancelled queries don't publish stale results.

LumeApp — configures the service with the shared container at launch, alongside the indexer.

Behavior

• "Inception" → still top, via lexical match.
• "heist inside a dream" / "movie about robots falling in love" → now surfaces conceptually-related titles whose names don't contain those words.
• Same 300ms debounce; the heavy scan runs off the main thread.

Testing

• ✅ Builds clean on the iOS 26.5 simulator.
• ✅ Pre-commit gates (swiftformat + swiftlint) pass.
• ⚠️ Not runtime-verified for ranking quality — needs a library that's been through the background indexer (which downloads the embedding model OTA). The 0.35 threshold may want tuning against a real catalog.
• No unit tests added: ranking depends on the on-device ML model (non-deterministic to set up) and the vector math is private to the actor.

🤖 Generated with Claude Code