feat(engine): add Python replay bindings, gallery pose-expansion, scene detection, embedding dumps

New C++ sources:
- kpn_bindings.cpp (sae_kpn): assembles the real face_tracker/identity_matcher/
  scene_tracker nodes inside a Python-driven KPN network via nanobind, for
  offline threshold-sweep replay against dumped embeddings (scripts/optimizer/).
- track_gallery.hpp: per-film gallery expansion — promotes a confidently-
  identified track's novel-pose reference views into an in-memory annex so
  later frames/tracks of that actor at similar poses are recognised, without
  touching the baked gallery.
- dump_embeddings.cpp: standalone exe that runs detect→embed only (no gallery,
  no matching) and dumps per-frame face embeddings + metadata to HDF5, so a
  parameter sweep can replay the expensive half once and vary tracking/matching
  config freely downstream.
- scene_detector.hpp / scene_detector_node.hpp: TransNetV2-based shot-boundary
  detection, opt-in alongside the always-on histogram cut detector.
- camera_position_change_detector_node.hpp, embedding_dump_node.hpp: supporting
  nodes for the above.

src/inference/scene_detector.hpp

@@ -0,0 +1,45 @@
+#pragma once
+#include <memory>
+#include <vector>
+
+#include <opencv2/core.hpp>
+
+// ── ISceneDetector ────────────────────────────────────────────────────────────
+// Backend-agnostic shot-boundary (scene-cut) detector interface, mirroring
+// IFaceDetector. The concrete implementation wraps TransNetV2 and is selected at
+// compile time by CMake (SAE_INFERENCE_BACKEND): exactly one of
+// backends/ort_backend.cpp or backends/trt_backend.cpp provides
+// make_scene_detector().
+//
+// TransNetV2 consumes a window of exactly kWindow consecutive frames, each
+// downscaled to kFrameW×kFrameH RGB, and predicts a per-frame boundary
+// probability. It is a DENSE model: the frames it sees must be consecutive at
+// (near) native frame rate — 1-FPS sampled frames give meaningless output.
+// Feeding is handled by the pipeline (dense decode + decimator); this interface
+// only exposes the fixed-size window inference.
+
+struct Config;
+
+struct ISceneDetector {
+    virtual ~ISceneDetector() = default;
+
+    // TransNetV2 fixed input contract (from the elya5/transnetv2 ONNX export):
+    //   input  "input" : float32 [1, 100, 27, 48, 3]  (RGB, channels-last, 0-255)
+    //   output "534"   : float32 [1, 100, 1]  single-frame boundary logits
+    //   output "535"   : float32 [1, 100, 1]  "many-hot" auxiliary head (unused)
+    static constexpr int kWindow  = 100;
+    static constexpr int kFrameH  = 27;
+    static constexpr int kFrameW  = 48;
+
+    // Run one window of exactly kWindow frames. Each frame must already be
+    // kFrameW×kFrameH, BGR, CV_8UC3 (the backend handles BGR→RGB). Returns
+    // kWindow boundary probabilities in [0,1] (sigmoid of the primary head),
+    // one per input frame, in order.
+    virtual std::vector<float> detect_window(
+        const std::vector<cv::Mat>& window) = 0;
+};
+
+// Construct the scene detector for the compiled-in backend. Reads
+// cfg.scene_model / cfg.scene_engine and cfg.trt. Only called when scene
+// detection is enabled (cfg.scene_detect).
+std::unique_ptr<ISceneDetector> make_scene_detector(const Config& cfg);