Resilient LLM runner: model fallback chain, provider-death blacklist, cooldown, ANSI-repair

- llm.py: single retry/fallback/cooldown authority used by analyze/verify/pipeline
- provider [404] blacklists model + switches; rate limit escalates then cooldown
- unparsable model output (bad JSON/ANSI noise) skips to next model
- verify.py: DS_VERIFY_TIMEOUT, true rate-limit handling, honest FAILED reporting
- pipeline_all.py: honors model, reports verify status accurately
- aggregate.py: skip epNNNN_corrected.json strays (was duplicating ep 304)
- gitignore data outputs; drop stray analysis/analyst/analysts dirs

aggregate.py

@@ -9,20 +9,23 @@ Contents:
 """
 import json
 import os
+import re
 from datetime import datetime, timezone
 
 from config import DATA_DIR
 from analyze import ANALYSIS_DIR, MASTER_JSON
 from feed import load_episodes
 
+_EP_FILE_RE = re.compile(r"^ep\d{4}\.json$")
+
 
 def load_analyses() -> list[dict]:
     out = []
     if not os.path.isdir(ANALYSIS_DIR):
         return out
     for f in sorted(os.listdir(ANALYSIS_DIR)):
-        if not f.endswith(".json"):
-            continue
+        if not _EP_FILE_RE.match(f):
+            continue  # ignore stray files like epNNNN_corrected.json
         try:
             with open(os.path.join(ANALYSIS_DIR, f), encoding="utf-8") as fh:
                 out.append(json.load(fh))