[antonreshetov]

Semantic search now uses only the built-in EmbeddingGemma model. Connecting an Ollama or LM Studio embedding server is removed before the 6.3 release.

Why:

• Ranking thresholds are calibrated for EmbeddingGemma; other models produce a different score distribution and miss results.
• A live check with LM Studio showed setup issues users would hit: chat models in the embedding model list, EmbeddingGemma 2 failing to load with LM Studio's default context, and meaning search silently stopping after LM Studio unloads an idle model.
• The feature is unreleased, so nothing needs migrating.

Changes:

• Preferences → Semantic Search: Source is Off or Built-in model; server, URL and embedding model fields are removed. A saved provider source reads as Off.
• Intel Macs: the settings show that semantic search is not available there, and the palette no longer offers Turn on semantic search. Search by words is unchanged.
• Main: removed the embedding server client, the semantic-models IPC channel and provider branches in the index; EmbeddingGemma task prompts are always applied.
• Locales: removed server and model keys, updated the description, the Intel message and the model start error (en_US, ru_RU).
• Docs: search and AI pages describe only the built-in model; added a screenshot of search by meaning.

Checks:

• Vitest: src/main/ai, src/renderer/composables, src/renderer/components/preferences.
• ESLint on changed files, tsc for main, vue-tsc shows no errors in changed files.
• VitePress docs build.
• Manual check in a dev build: settings, indexing with the built-in model, results by meaning in the palette.