Use demo cache immediately and memoize results

app.py

@@ -299,6 +299,7 @@ def _llm_report(p_mean: float, per_model: list, net_saliency: dict | None = None
 # ── model loading ──────────────────────────────────────────────────────────
 
 _model_cache: dict[str, list] = {}
+_result_cache: dict[str, tuple[str, str, str, object]] = {}
 
 def get_models(atlas: str = "cc200"):
     global _model_cache
@@ -526,6 +527,11 @@ def run_gcn(file_path):
         return "", "", "", None
 
     path = Path(file_path)
+    cache_key = str(path)
+    if cache_key in _result_cache:
+        return _result_cache[cache_key]
+
+    demo_key = path.name
     atlas_key = "cc200"  # default; overridden below for .1D files
     try:
         if path.suffix == ".npz":
@@ -757,10 +763,10 @@ CC200 LOSO AUC = 0.7298 · HO = 0.7212 · AAL = 0.6959 · 1,102 subjects · 20 s
 
     # LLM clinical interpretation via AMD MI300X vLLM endpoint
     # Fall back to demo cache for known subjects if endpoint is down
-    _demo_key = os.path.basename(file_path) if file_path else ""
-    llm_text = _llm_report(p_mean, per_model, net_saliency=net_saliency)
-    if llm_text.startswith("[LLM unavailable") and _demo_key in _DEMO_LLM_CACHE:
-        llm_text = _DEMO_LLM_CACHE[_demo_key]
+    if demo_key in _DEMO_LLM_CACHE:
+        llm_text = _DEMO_LLM_CACHE[demo_key]
+    else:
+        llm_text = _llm_report(p_mean, per_model, net_saliency=net_saliency)
     import re as _re
     def _md_to_html(txt):
         txt = _re.sub(r'^#{1,3}\s*(.+)$', r'<h4 style="color:#94a3b8;margin:1em 0 0.3em;font-size:0.9rem">\1</h4>', txt, flags=_re.MULTILINE)
@@ -780,7 +786,9 @@ CC200 LOSO AUC = 0.7298 · HO = 0.7212 · AAL = 0.6959 · 1,102 subjects · 20 s
 AI-assisted screening only · Not a clinical diagnosis · Findings must be integrated with ADOS-2, ADI-R, and full developmental history · Refer to licensed neuropsychologist for formal evaluation.</div>
 </div>"""
 
-    return verdict, ensemble, report, sal_img
+    result = (verdict, ensemble, report, sal_img)
+    _result_cache[cache_key] = result
+    return result
 
 
 # ── Static HTML sections ───────────────────────────────────────────────────