Initial commit: scene-actor-extraction pipeline

Source (KPN++ pipeline nodes, ArcFace embedders, SCRFD/YuNet detectors,
gallery builder), build scripts, and eval artifacts.

- external/KPN as a git submodule (gitea.tourolle.paris/dtourolle/KPN)
- ONNX models tracked via Git LFS (models/*.onnx)
- generated outputs, TensorRT engines, reference repos, and media ignored

src/gallery/gallery_builder.hpp

@@ -0,0 +1,36 @@
+#pragma once
+#include "types.hpp"
+#include <string>
+
+// Build an ActorGallery from a directory tree:
+//
+//   gallery_root/
+//     nm0000093_Brad_Pitt/
+//       img1.jpg
+//       img2.jpg
+//       ...
+//     nm0000129_Cate_Blanchett/
+//       ...
+//
+// Each subdirectory name is parsed as "<imdb_id>_<Name_With_Underscores>".
+// For every image:
+//   1. Detect face with SCRFD-500MF (expect exactly one; warn and skip if 0 or >1).
+//   2. Align with ArcFace 5-point transform → 112×112 crop.
+//   3. Embed with ArcFace ONNX → 512-dim L2-normalised embedding.
+// All embeddings are stored (best-of-N match at query time).
+//
+// Returns a gallery ready to pass to save_gallery() / IdentityMatcherFunc.
+
+struct BuildConfig {
+    std::string gallery_root;   // directory tree described above
+    std::string detector_model;
+    std::string arcface_model;
+    float       detector_conf{0.5f};
+    float       detector_nms{0.4f};
+    int         max_side{500};  // downscale source images to this max dimension
+                                // before detection — TMDB portraits are ~2k px,
+                                // SCRFD trains on smaller faces and detection
+                                // confidence drops on huge inputs. 0 = disabled.
+};
+
+ActorGallery build_gallery(const BuildConfig& cfg);