Issue Prompt: memory-lancedb local embeddings support (LM Studio)

Hello!
I want to enable my local mac LM Studio embedings model, and asked LLM to solve it and prepare as a prompt, so you can generate the solution yourself too. It was tested to work and after here its AI written. Let me know what you think about such approach, I can participate more.

Title

memory-lancedb: support OpenAI-compatible local embeddings (baseUrl + dimensions) and fix vector-size mismatch

Problem

memory-lancedb is hard to use with local OpenAI-compatible embedding endpoints (for example LM Studio/Ollama proxies).

Before fix:

• Config only allowed OpenAI model enum values.
• No embedding.baseUrl for custom endpoints.
• No embedding.dimensions override for non-OpenAI models.
• Using SDK embedding calls against LM Studio could yield vector-size mismatch vs LanceDB schema in practice.

Reproduction

1. Configure plugins.slots.memory = "memory-lancedb".
2. Set embedding model to local nomic GGUF (text-embedding-nomic-embed-text-v1.5@q8_0) on LM Studio.
3. Run memory tools (memory_store then memory_recall).
4. Observe LanceDB query errors due vector dimension mismatch.

Expected behavior

• Plugin accepts any OpenAI-compatible model id.
• Plugin can target custom embedding endpoint via embedding.baseUrl.
• Plugin can pin vector size via embedding.dimensions.
• Stored/search vectors use consistent dimension and work end-to-end with local providers.

Proposed solution

1. Extend plugin config schema:
   • embedding.baseUrl?: string
   • embedding.dimensions?: integer
   • remove hardcoded embedding.model enum restriction
2. Use configured dimensions when initializing LanceDB vector schema.
3. Replace the openai SDK with raw HTTP fetch() to /embeddings for broader local-provider compatibility. Add AbortSignal.timeout(30_000) to prevent indefinite hangs when a local server is unresponsive. Remove the openai dependency from package.json.
4. Split embedding dimensions into two concerns:
   • requestDimensions: only sent in the API request body when the user explicitly sets embedding.dimensions in config. This avoids breaking local providers that reject unknown request fields.
   • expectedDimensions: always validated against the response vector length (derived from LanceDB schema). This catches dimension mismatches early instead of surfacing as cryptic LanceDB errors.
5. Support ${ENV_VAR} substitution in embedding.baseUrl (same pattern as embedding.apiKey).
6. Validate embedding.dimensions range (1..32768) and reject empty/whitespace baseUrl.
7. Update plugin manifest (openclaw.plugin.json) UI hints and JSON schema.
8. Add tests for:
   • custom model + baseUrl + dimensions
   • env var resolution for baseUrl
   • unknown model requiring dimensions
   • invalid dimensions (type, range: 0, fractional)
   • empty baseUrl rejection

Files touched (fork reference)

• extensions/memory-lancedb/config.ts
• extensions/memory-lancedb/index.ts
• extensions/memory-lancedb/openclaw.plugin.json
• extensions/memory-lancedb/index.test.ts
• extensions/memory-lancedb/package.json (remove openai dependency)
• pnpm-lock.yaml (lockfile cleanup)

Acceptance criteria

• memory-lancedb can run with LM Studio embeddings.
• memory_store, memory_recall, memory_forget work with local endpoint.
• Plugin config validates correctly for both OpenAI hosted and local OpenAI-compatible providers.
• Existing OpenAI usage remains backward compatible (no dimensions sent in request body for default models).

Validation performed (fork)

• Unit tests: pnpm vitest run extensions/memory-lancedb/index.test.ts — 15/15 passed
• Live smoke (LM Studio on Tailscale, text-embedding-nomic-embed-text-v1.5@q8_0, 768 dims):
  • memory_store -> created
  • memory_recall -> found stored memory
  • memory_forget -> deleted stored memory
  • Auto-recall hook: memory-lancedb: injecting 3 memories into context
  • Graceful degradation: app kept running when embedding server had no model loaded, auto-recovered once model was available
• Full suite: pnpm build && pnpm check && pnpm test — pass (8 pre-existing upstream failures only)

Risk notes

• Local providers may vary in /embeddings behavior; dimension enforcement surfaces misconfiguration early.
• For custom models, embedding.dimensions should be explicitly set.
• The openai SDK included automatic retry logic for transient errors (429, 500). The raw fetch replacement does not retry, but for a memory plugin making ~1 call per message, rate limiting is unlikely. Failures degrade gracefully via the existing try/catch in lifecycle hooks.