tune: increase default local llm timeouts for stability

transcribe.py

@@ -279,6 +279,9 @@ def post_correct_with_local_llm(text, args, llm_state, context_hint=""):
     if not args.post_correct_llm or not llm_state.get("available", True):
         return text
 
+    if args.post_correct_debug:
+        print(f"[POST-LLM][PRE]: {text}")
+
     prompt = (
         "You correct live English captions.\n"
         "Rules:\n"
@@ -306,13 +309,21 @@ def post_correct_with_local_llm(text, args, llm_state, context_hint=""):
         response.raise_for_status()
         raw = response.json().get("response", "").strip()
         cleaned = normalize_english_caption(raw)
+        if args.post_correct_debug:
+            print(f"[POST-LLM][RAW]: {raw}")
         if cleaned:
+            if args.post_correct_debug:
+                print(f"[POST-LLM][POST]: {cleaned}")
             return cleaned
+        if args.post_correct_debug:
+            print("[POST-LLM][POST]: <empty response, keeping pre text>")
     except Exception as exc:
         if not llm_state.get("warned", False):
             print(f"\n[SYSTEM]: Local post-correct LLM unavailable ({exc}). Falling back to rules-only.")
             llm_state["warned"] = True
         llm_state["available"] = False
+        if args.post_correct_debug:
+            print(f"[POST-LLM][ERROR]: {exc}")
     return text
 
 def maybe_merge_recent_caption(recent_en_lines, current_text, now_ts, window_sec):
@@ -522,9 +533,10 @@ def main():
     parser.add_argument("--post-correct-llm", action="store_true", help="Use local LLM for post-correction (fallback to rules on failure)")
     parser.add_argument("--post-correct-model", type=str, default="llama3.1:8b-instruct", help="Local Ollama model for post-correction")
     parser.add_argument("--post-correct-ollama-url", type=str, default="http://127.0.0.1:11434/api/generate", help="Ollama generate endpoint for local post-correction")
-    parser.add_argument("--post-correct-llm-timeout", type=float, default=3.0, help="Timeout seconds for local LLM post-correction (default: 3.0)")
+    parser.add_argument("--post-correct-llm-timeout", type=float, default=8.0, help="Timeout seconds for local LLM post-correction (default: 8.0)")
+    parser.add_argument("--post-correct-debug", action="store_true", help="Print pre/post LLM correction text for debugging")
     parser.add_argument("--llm-merge-decider", action="store_true", help="Use local LLM to validate/refine smart-correct line merges")
-    parser.add_argument("--llm-merge-timeout", type=float, default=0.7, help="Timeout seconds for LLM merge decision (default: 0.7)")
+    parser.add_argument("--llm-merge-timeout", type=float, default=1.5, help="Timeout seconds for LLM merge decision (default: 1.5)")
     parser.add_argument("--speaker-change-detect", action="store_true", help="Enable lightweight speaker-change detection to force caption line cuts")
     parser.add_argument("--speaker-sim-threshold", type=float, default=0.72, help="Cosine similarity threshold below which a speaker change is assumed (default: 0.72)")
     parser.add_argument("--speaker-min-buffer", type=float, default=1.6, help="Minimum buffered speech seconds before speaker-change checks (default: 1.6)")
@@ -542,6 +554,8 @@ def main():
         print("Post-correct mode: rules enabled.")
     if args.post_correct_llm:
         print(f"Post-correct LLM: {args.post_correct_model} via {args.post_correct_ollama_url}")
+    if args.post_correct_debug:
+        print("Post-correct debug logging enabled.")
     if args.llm_merge_decider and not args.post_correct_llm:
         print("[SYSTEM]: --llm-merge-decider requires --post-correct-llm. Falling back to heuristic merge.")
         args.llm_merge_decider = False