harness: --direct-tcp, to exercise the direct path and the pod's TLS

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CofY4TJrAY7rf1WYJnPtdL

tools/runpod_model_test.py

@@ -111,7 +111,7 @@ _OCR_ENGINES = ("paddle", "doctr", "surya")
 
 def _pin_to_runpod(model: str, engine: str, vram: float, usd: float,
                    image: str = "", add_as: str = "",
-                   capability: str = "") -> bool:
+                   capability: str = "", direct_tcp: bool = False) -> bool:
     """Give this ONE model a runpod block. True when the catalogue changed."""
     data = _load_models()
     section, index, entry = _find(data, model)
@@ -147,6 +147,11 @@ def _pin_to_runpod(model: str, engine: str, vram: float, usd: float,
         # that had already been published under that very tag. Production still
         # runs `:latest`; this override exists so a result means something.
         entry["runpod"]["image"] = image
+    if direct_tcp:
+        # The pod's public ip:port instead of the Cloudflare-fronted proxy —
+        # with TLS the pod brings itself on a coderai image. Exercised here
+        # because nothing else does until a request runs past 100 s.
+        entry["runpod"]["direct_tcp"] = True
     data[section][index] = entry
     _save_models(data)
     os.utime(MODELS, None)           # the front watches this mtime to push a reload
@@ -216,6 +221,9 @@ def main() -> int:
                          "VERSIONED tag when testing: a machine may serve a "
                          "cached ':latest' and the result would not say which "
                          "build it exercised.")
+    ap.add_argument("--direct-tcp", action="store_true",
+                    help="reach the pod at its public ip:port (TLS on coderai "
+                         "images) instead of RunPod's Cloudflare-fronted proxy")
     args = ap.parse_args()
 
     backup = MODELS + ".testrun-backup"
@@ -238,7 +246,8 @@ def main() -> int:
         for model in args.models:
             print(f"================ {model} ================", flush=True)
             if not _pin_to_runpod(model, args.engine, args.vram, args.usd,
-                                  args.image, args.add_as, args.capability):
+                                  args.image, args.add_as, args.capability,
+                                  direct_tcp=args.direct_tcp):
                 results.append((model, {"error": "not in models.json"}))
                 continue
             problem = _wait_for_engine(model)