fix: bypass LiteLLM for Ollama embeddings to resolve 400 Bad Request (#1425)

[GratefulDave]

Fixes #1425 — persistent `httpx.HTTPStatusError: 400 Bad Request` on Ollama `/api/embed` during memory similarity search.

Root Cause

Two distinct causes, both fixed:

1. LiteLLM's Ollama handler sends a malformed request — leaks the `ollama/` prefix into the model name field and forwards unsupported kwargs (e.g. `encoding_format: null`) that Ollama 0.18.x+ rejects with 400.

2. `None` values in the embedding input array — when a `None` ends up in the texts list (e.g. from a failed upstream LLM call or race condition), it serialises to JSON `null`, which Ollama rejects with `{"error": "invalid input type"}` → 400.

Changes

`models.py`

• `_ollama_embed()` — new helper on `LiteLLMEmbeddingWrapper` that calls Ollama's `/api/embed` directly via `httpx` (already a transitive dependency), bypassing LiteLLM entirely
  • Strips the `ollama/` prefix from the model name
  • Sanitises input: converts `None` → `""` and any non-str → `str()` before sending, with a `logging.warning` that records which inputs were coerced
  • Retries only on transient errors (429, 503) with exponential backoff; raises immediately on 400 (retrying a bad payload is pointless)
  • Logs the HTTP status, Ollama response body, and first 100 chars of each input on non-200 responses
• `embed_query` / `embed_documents` — route through `_ollama_embed()` when `provider == "ollama"`
• `_is_ollama()` — helper predicate

`plugins/_memory/helpers/memory.py`

• `search_similarity_threshold` — wraps `asearch` in `try/except` so any embedding failure returns `[]` (no memories this turn) instead of propagating and crashing the agent's monologue loop

Test Plan

• Start agent-zero with `ollama/nomic-embed-text` as the embedding model
• Trigger a memory operation — confirm no 400 errors
• Verify memory similarity search returns results
• Verify `embed_documents` path works (store and retrieve a memory)
• Verify agent continues normally when Ollama is temporarily unreachable (returns empty memories, does not crash)

Tested locally on Docker with `nomic-embed-text` (768-dim), Ollama 0.18.3, `host.docker.internal:11434`.