fix(segtrain): reduce input height to 800px on first run to avoid OOM

ketos segtrain has no batch-size flag (-B), so with the default 1800px
input height the intermediate CNN feature maps consume ~500 MB+ per
image, causing the kernel OOM-killer (exit -9) to terminate the process.

On first run (no existing blla.mlmodel), override the VGSL spec to use
800px height instead. Subsequent runs load the saved model with
--resize both, preserving incremental fine-tuning.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ocr-service/main.py

@@ -465,7 +465,15 @@ async def segtrain_model(
                 "-N", "10",
             ]
             if os.path.exists(blla_model_path):
-                cmd += ["-i", blla_model_path]
+                cmd += ["-i", blla_model_path, "--resize", "both"]
+            else:
+                # No pretrained model — train from scratch with reduced height (800px)
+                # to keep peak RAM under ~200 MB on CPU (default 1800px uses ~500 MB+)
+                cmd += [
+                    "-s",
+                    "[1,800,0,3 Cr7,7,64,2,2 Gn32 Cr3,3,128,2,2 Gn32 Cr3,3,128 Gn32 "
+                    "Cr3,3,256 Gn32 Cr3,3,256 Gn32 Lbx32 Lby32 Cr1,1,32 Gn32 Lby32 Lbx32]",
+                ]
             cmd += xml_files
 
             log.info("Running: %s", " ".join(cmd[:5]) + " ...")