feat: upgrade to SigLIP2 ViT-B/16 for semantic search

Replace OpenCLIP ViT-B/32 (512-d, ~78% recall) with SigLIP2 ViT-B/16
(768-d, ~84% recall) as the default embedding model for significantly
better image-text retrieval quality.

- New SigLIP2Embedder class with 384px input and SigLIP normalization
- ONNX export pipeline for SigLIP2 visual + textual encoders
- Migration 0010: resize embeddings.vector from 512 to 768 dimensions
- Config-driven model selection: "siglip2_vitb16" (default) or
  "openclip_vitb32" (legacy) — both models can coexist
- Content classifier follows the configured embedder family
- Existing embeddings cleared on migration; vision backfill regenerates

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/app/services/vision/export_models.py

@@ -120,6 +120,76 @@ def export_openclip(models_dir: Path):
         logger.info("Textual encoder exported (%.1f MB)", size_mb)
 
 
+def export_siglip2(models_dir: Path):
+    """Export SigLIP2 ViT-B/16 to two ONNX files (visual + textual)."""
+    import torch
+    import open_clip
+
+    out_dir = models_dir / "embed_siglip2"
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    visual_path = out_dir / "visual.onnx"
+    textual_path = out_dir / "textual.onnx"
+
+    if visual_path.exists() and textual_path.exists():
+        logger.info("SigLIP2 ONNX files already exist, skipping export")
+        return
+
+    logger.info("Loading SigLIP2 ViT-B-16-SigLIP2 webli...")
+    model, _, preprocess = open_clip.create_model_and_transforms(
+        "ViT-B-16-SigLIP2", pretrained="webli"
+    )
+    model.eval()
+
+    export_kwargs = dict(opset_version=14, dynamo=False)
+
+    # ── Visual encoder ────────────────────────────────────────────────
+    if not visual_path.exists():
+        logger.info("Exporting SigLIP2 visual encoder → %s", visual_path)
+        dummy_image = torch.randn(1, 3, 384, 384)
+        torch.onnx.export(
+            model.visual,
+            dummy_image,
+            str(visual_path),
+            input_names=["image"],
+            output_names=["embedding"],
+            dynamic_axes={"image": {0: "batch"}},
+            **export_kwargs,
+        )
+        size_mb = visual_path.stat().st_size / 1e6
+        logger.info("SigLIP2 visual encoder exported (%.1f MB)", size_mb)
+
+    # ── Textual encoder ───────────────────────────────────────────────
+    if not textual_path.exists():
+        logger.info("Exporting SigLIP2 textual encoder → %s", textual_path)
+        tokenizer = open_clip.get_tokenizer("ViT-B-16-SigLIP2")
+        dummy_text = tokenizer(["a photo"]).to(torch.int64)
+
+        class SigLIP2TextEncoder(torch.nn.Module):
+            """Wrap the SigLIP2 text transformer for ONNX export."""
+            def __init__(self, clip_model):
+                super().__init__()
+                self.text = clip_model.text
+
+            def forward(self, text):
+                return self.text(text)
+
+        text_enc = SigLIP2TextEncoder(model)
+        text_enc.eval()
+
+        torch.onnx.export(
+            text_enc,
+            dummy_text,
+            str(textual_path),
+            input_names=["text"],
+            output_names=["embedding"],
+            dynamic_axes={"text": {0: "batch"}},
+            **export_kwargs,
+        )
+        size_mb = textual_path.stat().st_size / 1e6
+        logger.info("SigLIP2 textual encoder exported (%.1f MB)", size_mb)
+
+
 def export_yolov8n(models_dir: Path):
     """Export YOLOv8n to ONNX."""
     out_dir = models_dir / "detect"
@@ -174,6 +244,7 @@ def main():
     logger.info("Exporting models to %s", models_dir)
 
     export_openclip(models_dir)
+    export_siglip2(models_dir)
     export_yolov8n(models_dir)
 
     logger.info("Done. Run bootstrap_models.py next to download YuNet + SFace.")