feat(embeddings): LlamaCppEmbeddingProvider behind MYCELIUM_LLM_PROVIDER — #178 part 1

[Dewinator]

Summary

• Lands the embedding side of Spike 2's recommendation (PR spike(native-llm): node-llama-cpp validates as in-process LLM (#178) #182, docs/native-llm-spike.md) as a real provider — the centerpiece of the native-app track (epic: native standalone app (no Docker) — macOS / Windows / Linux #176, Welle 1).
• Sets MYCELIUM_LLM_PROVIDER=llama-cpp to swap the in-process node-llama-cpp provider in for Ollama. Default stays ollama, so existing installs are unaffected.
• Identical EmbeddingProvider contract → all consumers (MemoryService, import_markdown, REM digest pre-stage) work unchanged.

What changes

• LlamaCppEmbeddingProvider (new) — lazy-init, dynamic import of node-llama-cpp, OS-native models dir (~/Library/Application Support/mycelium/models on macOS, %APPDATA%\mycelium\models on Windows, $XDG_DATA_HOME/mycelium/models on Linux), 768d default to match VECTOR(768) schema, dim-mismatch guard, PGlite-style queue serializer (matches PR feat(native): PGlite adapter foundation — issue #184 part 1 #185).
• createEmbeddingProvider() — env-driven factory: ollama (default) | llama-cpp (alias llamacpp). Unknown values throw at startup.
• Packaging — node-llama-cpp@^3.4.0 added as optionalDependencies so niche/unsupported platforms can still install the rest of mycelium. Dynamic import inside the provider yields a clear "install node-llama-cpp or switch to ollama" error if the runtime is absent.
• Tests — 7 new unit tests (selection, dim, path defaults, alias, error path) plus an opt-in real-GGUF e2e test gated on MYCELIUM_TEST_LLAMA_CPP=1 (downloads ~84 MB on first run, skipped in CI).
• Docs — docs/native-llm-spike.md adds an "Implementation 1" section documenting the integration contract.

Why default is not flipped

The spike (docs/native-llm-spike.md §3) flagged a tokenizer-config warning on the nomic-embed-text GGUF: cross-path embedding parity with Ollama is not yet validated, and divergence below ~0.95 cosine would invalidate stored vectors on existing installs. Default-flip waits until a CI gate verifies parity. Fresh native-app installs (#176 sub-task 4, Tauri shell) start empty, so they can opt in safely.

Out of scope

• Cosine-similarity cross-validation against Ollama (deferred CI gate)
• SHA-256 verification of GGUF downloads (Pillar 6 follow-up)
• Chat provider + REM-digest re-route (separate sub-task of feat(app): Spike 2 — node-llama-cpp embedding + chat bridge (drop Ollama dependency) #178)

Test plan

• npm run build — clean
• npm test — 946 pass, 1 skipped (gated e2e), 0 fail
• Local opt-in: MYCELIUM_TEST_LLAMA_CPP=1 npm test -- --test-name-pattern 'real GGUF' (verifies first-run download, 768d, queued concurrent embeds)
• Manual: MYCELIUM_LLM_PROVIDER=llama-cpp npm run dev, remember() a memory, recall() it back

🤖 Generated with Claude Code