feat(platform): opt-in local Ollama providers for embeddings + recall LLM

[jayesh-keychain]

Adds Ollama as a third real provider so the whole memory loop runs on-box with no API keys and nothing leaving the machine.

Why: with no key set the seam falls back to the fakes. That keeps CI offline (§2.5) but makes local recall untestable — the fake embedder is hash-based, so any query worded differently from the stored text scores ~0.0004 cosine and ask always answers "I don't have a memory about X". Ask-conversation threads (#265) are never persisted either, because the §05.6 miss gate returns before the persist call, so the feature reads as broken when it is merely unexercised.

What:

• OllamaEmbeddingProvider over Ollama's OpenAI-compatible /embeddings. Ollama cannot truncate to an arbitrary width, so we never send dimensions and instead VERIFY the returned width against EMBEDDING_DIM, failing loudly rather than writing a vector pgvector rejects later. mxbai-embed-large is the 1024-d model matching the schema.
• Recall LLM reuses OpenAiLlmProvider against Ollama's /chat/completions — all eight call sites and the strict-JSON citation parsing are shared, not reimplemented. The hardcoded URL becomes an injectable baseUrl; default unchanged.
• Model router gains an LLM_PROVIDER=ollama short-circuit forcing the local vendor for every class, mirroring the existing 'fake' switch.

Privacy fence (§2.2, M9): isThirdParty is false ONLY for a loopback base URL, and it is an explicit constructor argument defaulting to true. Moving OLLAMA_BASE_URL to a remote host re-arms the fence — a URL edit in .env can never silently disarm it.

Both providers are opt-in and NEVER inferred: Ollama is a daemon, not a key, so there is nothing to infer from, and a box running Ollama for something unrelated must not be hijacked. With no EMBEDDING_PROVIDER/LLM_PROVIDER set, the fakes still win.

• type-check, lint, build green
• 82 existing model-router + LLM unit tests pass unchanged
• End-to-end on a live stack with mxbai-embed-large + gemma3:4b: the semantic paraphrase "When does the quarterly meeting happen?" recalled a differently-worded memory ("The zorblatt quarterly review is scheduled for March 14th…") and answered "March 14th" with 1 citation — a query the fake embedder scores 0.0004 and always misses.
• Verified against the seeded demo corpus: "What is Priya allergic to?" → "peanuts" (conf 0.75, 1 citation).

No new tests: this adds provider implementations behind the existing seam, and the fakes remain forced in tests via vitest.config.ts.

🤖 Generated with Claude Code