fix: resolve all critical runtime errors and bugs from audit

- Add COMPLETIONS_API_KEY to config.py (env var + auto-generated fallback)
- Fix perplexity auto-search: upstream sends logprobs=true, parse_llama_stream_chunk
  extracts per-token logprobs, all_logprobs populated during streaming
- Fix all /api/models endpoints to target LLAMA_SERVER_BASE (port 8081) not OLLAMA_BASE
- Fix RAG embedding endpoint URL from port 11434 (Ollama) to 8081 (llama-server)
- Correct misleading error messages: 'inference server' not 'Ollama'
- Remove raw_results leak from SSE event stream in /api/search
- Fix weather query extractor: pattern-match instead of unconditional suffix append
- Escape FTS5 operator keywords (AND/OR/NOT/NEAR) in memory search
- Move auth.py BODY_LIMIT_DEFAULT_BYTES imports to module level
- Change RAG injection log level from warning to info
- Fix all 8 test files after modular refactor (rewire imports from correct modules)
- Update AGENTS.md and README.md to reflect v1.8.0 changes

routers/completions.py

@@ -178,7 +178,7 @@ async def _stream_chat(payload: dict, model: str, conv_id: str, request: Request
                 async for line in resp.aiter_lines():
                     if not line.strip():
                         continue
-                    token, done, _ = parse_llama_stream_chunk(line)
+                    token, done, _, _ = parse_llama_stream_chunk(line)
                     if token:
                         full_response.append(token)
                         yield _build_openai_chunk(token, model, conv_id)
@@ -222,7 +222,7 @@ async def _blocking_chat(payload: dict, model: str, conv_id: str, request: Reque
                 async for line in resp.aiter_lines():
                     if not line.strip():
                         continue
-                    token, done, _ = parse_llama_stream_chunk(line)
+                    token, done, _, _ = parse_llama_stream_chunk(line)
                     if token:
                         full_response.append(token)
                     if done: