fix: cherry-pick improvements from feature/semantic_caching

- semantic-cache.ts: use upsert instead of add to avoid duplicate-ID errors
- semantic-cache.ts: add crypto.randomUUID() fallback for constrained runtimes
- ollama-client.ts: JSON.stringify non-string Ollama errors to avoid [object Object]
- tests: update semantic-cache mocks and assertions for upsert

tests/semantic-cache.test.ts

@@ -10,6 +10,7 @@ jest.mock('chromadb', () => ({
       getOrCreateCollection: jest.fn().mockResolvedValue({
         query: jest.fn(),
         add: jest.fn(),
+        upsert: jest.fn(),
         reset: jest.fn(),
       }),
       deleteCollection: jest.fn(),
@@ -135,13 +136,13 @@ describe('SemanticCacheService', () => {
 
       await disabledCacheService.setCache('test query', 'test response');
 
-      expect(mockCollection.add).not.toHaveBeenCalled();
+      expect(mockCollection.upsert).not.toHaveBeenCalled();
     });
 
     it('should add content to cache', async () => {
       await cacheService.setCache('test query', 'test response');
 
-      expect(mockCollection.add).toHaveBeenCalled();
+      expect(mockCollection.upsert).toHaveBeenCalled();
     });
   });