Fix: clear benchmark cache, add binary verification and debug logging

app.py

@@ -438,23 +438,34 @@ def _ensure_llama_cpp():
     # Use a known stable release
     release = "b5440"
     url = f"https://github.com/ggml-org/llama.cpp/releases/download/{release}/llama-{release}-bin-ubuntu-x64.zip"
+    _log(f"Benchmark: URL: {url}")
     zip_path = os.path.join(bin_dir, "llama.zip")
     urllib.request.urlretrieve(url, zip_path)
+    _log(f"Benchmark: downloaded {os.path.getsize(zip_path)} bytes")
 
     _log("Benchmark: extracting llama.cpp...")
     with zipfile.ZipFile(zip_path) as zf:
+        _log(f"Benchmark: zip contents: {zf.namelist()}")
         for name in zf.namelist():
-            if name.endswith("llama-cli"):
-                # Extract to bin_dir
+            if name.endswith("llama-cli") or name.endswith("llama-cli.exe"):
                 data = zf.read(name)
                 with open(llama_cli, "wb") as f:
                     f.write(data)
+                _log(f"Benchmark: extracted {name} ({len(data)} bytes)")
                 break
     os.chmod(llama_cli, 0o755)
     os.unlink(zip_path)
 
+    # Verify binary works
+    _log("Benchmark: verifying llama-cli binary...")
+    try:
+        check = subprocess.run([llama_cli, "--help"], capture_output=True, text=True, timeout=10)
+        _log(f"Benchmark: llama-cli --help exit code: {check.returncode}")
+        _log(f"Benchmark: llama-cli --help output: {check.stdout[:200]}")
+    except Exception as e:
+        _log(f"Benchmark: llama-cli verification failed: {e}")
+
     _LLAMA_CPP_BIN = llama_cli
-    _log(f"Benchmark: llama.cpp ready at {llama_cli}")
     return llama_cli
 
 
@@ -589,10 +600,12 @@ def _patched_create_app(blocks, **kwargs):
             return JSONResponse({"error": str(e)}, status_code=502)
 
     @fa_app.get("/respite/benchmark")
-    def _benchmark():
+    def _benchmark(force: bool = False):
         """Run CPU LLM benchmark. Takes 5-15 minutes. Returns cached results on repeat calls."""
         try:
-            import time
+            global _BENCHMARK_RESULTS
+            if force:
+                _BENCHMARK_RESULTS = []
             results = _run_benchmark()
             return JSONResponse({
                 "status": "complete",