fix(embed): fall back to embeddings probe when models route is missing

[spacemolt-molty]

What

Available() for OpenAI-compatible servers now falls back to a real embeddings round-trip when the GET <endpoint>/models route is missing (404/405/501).

Why

Availability for OpenAI-compatible providers is currently decided solely by GET /models. Some compatible providers serve embeddings but no models route at all — e.g. Voyage AI returns 404 there while /v1/embeddings works fine. Against such endpoints mnemon reports embedding_available: false and silently runs without semantic recall, even though the embedding endpoint itself is fully functional.

The OpenAI compatibility surface mnemon actually depends on is POST /embeddings, so availability should reflect that route when the conventional discovery route doesn't exist.

Behavior after the change:

• models route 200 → available (unchanged)
• models route 404/405/501 → probe POST /embeddings once; available iff it returns a usable embedding
• auth failures (401), quota (429), server errors (5xx) on either probe → unavailable, so the probe still catches misconfigured keys
• Ollama protocol path unchanged

Implementation notes: the embeddings round-trip is shared with Embed() via embedWithContext/decodeEmbedResponse, so the probe and the real call cannot drift apart.

Verified end-to-end against the Voyage AI production API (voyage-3.5) and against a local Ollama instance.

Checklist

• Deterministic tests pass (make test)
• Relevant E2E/process/Docker boundaries pass (make test-integration, when affected)
• New/changed behavior is covered by tests
• Documentation updated (USAGE.md, DESIGN.md, or README) if applicable
• User-facing release-note impact described in this PR, if applicable