Bug Description

After switching memorySearch.provider from openai to local (node-llama-cpp with nomic-embed-text-v1.5), the indexer detects the provider change and sets Dirty: yes, but never actually reindexes. No error or progress appears in gateway logs. The SQLite metadata remains unchanged (old OpenAI 1536-dim embeddings), and memory_search returns empty results.

Steps to Reproduce

1. Start with working OpenAI embedding (text-embedding-3-small, 1536 dims, 541 chunks indexed)
2. Apply config:

{
  "agents": {
    "defaults": {
      "memorySearch": {
        "provider": "local",
        "fallback": "none",
        "local": {
          "modelPath": "/path/to/nomic-embed-text-v1.5.Q8_0.gguf"
        }
      }
    }
  }
}

3. Restart gateway
4. Run openclaw memory status
5. Call memory_search to trigger sync
6. Wait several minutes, repeat

Expected Behavior

• Indexer detects provider/model change (OpenAI 1536 → local 768)
• SQLite is reset and all chunks are re-embedded with the local model
• memory_search returns results using local embeddings

Actual Behavior

• openclaw memory status shows:
  • Provider: local (requested: local) ✅
  • Dirty: yes (never clears)
  • Vector dims: 1536 (old OpenAI dims, never changes to 768)
  • Indexed: 63/63 files · 541 chunks (old data)
• memory_search returns empty results with provider: "local"
• Gateway logs contain zero mentions of embedding, reindex, llama, or local model loading
• No errors, no warnings — completely silent failure

Additional Context

• node-llama-cpp works correctly when tested directly:

$ node --input-type=module -e "
import { getLlama } from 'node-llama-cpp';
const llama = await getLlama();
const model = await llama.loadModel({ modelPath: '...' });
const ctx = await model.createEmbeddingContext();
const vec = await ctx.getEmbeddingFor('test');
console.log('dims:', vec.vector.length); // 768 ✅
"

• npx node-llama-cpp inspect gpu shows Metal available, Apple M4 Pro
• Also tried: pnpm rebuild node-llama-cpp, deleting SQLite and forcing rebuild (0/52 forever), multiple gateway restarts
• Tested over 2 days with same result
• Reverted to OpenAI embedding (works immediately)

Environment

• OpenClaw version: 2026.2.13
• Node: v22.22.0
• OS: macOS 25.2.0 (arm64)
• Hardware: Mac mini M4 Pro, 64GB RAM
• node-llama-cpp: 3.15.1 (prebuilt binaries, Metal available)
• Model: nomic-embed-text-v1.5.Q8_0.gguf (Q8_0, 139MB)
• Memory files: 63 files, ~1MB total (largest 442KB)