Add seed support to NeuralNetwork for reproducible weight initialisation

[Copilot]

Introduces NRE::NeuralNetwork — a seeded feedforward neural network enabling reproducible procedural generation across runs.

New files

• engine/neural/NeuralNetwork.h/.cpp — Fully-connected network with He-initialised weights via a seeded std::mt19937. Key API:
  • NeuralNetwork(uint32_t seed = 0) — seed 0 falls back to std::random_device
  • SetSeed(uint32_t) / GetSeed() — change seed and reinitialise all layer weights in-place
  • AddLayer(int in, int out) / Forward(vector<float>) — layer construction and ReLU forward pass
• examples/neural_demo/main.cpp — Demonstrates same-seed reproducibility and seed isolation
• CMakeLists.txt — Builds nre_neural static lib and neural_demo executable
• .gitignore — Excludes CMake build dirs and compiled artefacts

Usage

NRE::NeuralNetwork nn(42);
nn.AddLayer(4, 8);
nn.AddLayer(8, 4);
auto out1 = nn.Forward({0.5f, 0.3f, 0.8f, 0.1f});

nn.SetSeed(42);          // reinitialises weights — out2 == out1
auto out2 = nn.Forward({0.5f, 0.3f, 0.8f, 0.1f});

Original prompt

Reference: af11a6e add seeds to neural network

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.