Add documentation for collision detection and mesh classes

Co-authored-by: orange-cpp <59374393+orange-cpp@users.noreply.github.com>
2026-02-13 07:03:25 +00:00 · 2025-11-13 15:19:36 +00:00
parent d118e88f6b
commit 190a8bf91e
11 changed files with 2482 additions and 0 deletions
--- a/docs/collision/epa_algorithm.md
+++ b/docs/collision/epa_algorithm.md
@@ -0,0 +1,322 @@
+# `omath::collision::Epa` — Expanding Polytope Algorithm for penetration depth
+
+> Header: `omath/collision/epa_algorithm.hpp`
+> Namespace: `omath::collision`
+> Depends on: `Simplex<VertexType>`, collider types with `find_abs_furthest_vertex` method
+> Algorithm: **EPA** (Expanding Polytope Algorithm) for penetration depth and contact normal
+
+---
+
+## Overview
+
+The **EPA (Expanding Polytope Algorithm)** calculates the **penetration depth** and **separation normal** between two intersecting convex shapes. It is typically used as a follow-up to the GJK algorithm after a collision has been detected.
+
+EPA takes a 4-point simplex containing the origin (from GJK) and iteratively expands it to find the point on the Minkowski difference closest to the origin. This point gives both:
+* **Depth**: minimum translation distance to separate the shapes
+* **Normal**: direction of separation (pointing from shape B to shape A)
+
+`Epa` is a template class working with any collider type that implements the support function interface.
+
+---
+
+## `Epa::Result`
+
+```cpp
+struct Result final {
+    bool success{false};        // true if EPA converged
+    Vertex normal{};            // outward normal (from B to A)
+    float depth{0.0f};          // penetration depth
+    int iterations{0};          // number of iterations performed
+    int num_vertices{0};        // final polytope vertex count
+    int num_faces{0};           // final polytope face count
+};
+```
+
+### Fields
+
+* `success` — `true` if EPA successfully computed depth and normal; `false` if it failed to converge
+* `normal` — unit vector pointing from shape B toward shape A (separation direction)
+* `depth` — minimum distance to move shape A along `normal` to separate the shapes
+* `iterations` — actual iteration count (useful for performance tuning)
+* `num_vertices`, `num_faces` — final polytope size (for diagnostics)
+
+---
+
+## `Epa::Params`
+
+```cpp
+struct Params final {
+    int max_iterations{64};     // maximum iterations before giving up
+    float tolerance{1e-4f};     // absolute tolerance on distance growth
+};
+```
+
+### Fields
+
+* `max_iterations` — safety limit to prevent infinite loops (default 64)
+* `tolerance` — convergence threshold: stop when distance grows less than this (default 1e-4)
+
+---
+
+## `Epa` Template Class
+
+```cpp
+template<class ColliderType>
+class Epa final {
+public:
+    using Vertex = typename ColliderType::VertexType;
+    static_assert(EpaVector<Vertex>, "VertexType must satisfy EpaVector concept");
+
+    // Solve for penetration depth and normal
+    [[nodiscard]]
+    static Result solve(
+        const ColliderType& a,
+        const ColliderType& b,
+        const Simplex<Vertex>& simplex,
+        const Params params = {}
+    );
+};
+```
+
+### Precondition
+
+The `simplex` parameter must:
+* Have exactly 4 points (`simplex.size() == 4`)
+* Contain the origin (i.e., be a valid GJK result with `hit == true`)
+
+Violating this precondition leads to undefined behavior.
+
+---
+
+## Collider Requirements
+
+Any type used as `ColliderType` must provide:
+
+```cpp
+// Type alias for vertex type (typically Vector3<float>)
+using VertexType = /* ... */;
+
+// Find the farthest point in world space along the given direction
+[[nodiscard]]
+VertexType find_abs_furthest_vertex(const VertexType& direction) const;
+```
+
+---
+
+## Algorithm Details
+
+### Expanding Polytope
+
+EPA maintains a convex polytope (polyhedron) in Minkowski difference space `A - B`. Starting from the 4-point tetrahedron (simplex from GJK), it repeatedly:
+
+1. **Find closest face** to the origin
+2. **Support query** in the direction of the face normal
+3. **Expand polytope** by adding the new support point
+4. **Update faces** to maintain convexity
+
+The algorithm terminates when:
+* **Convergence**: the distance from origin to polytope stops growing (within tolerance)
+* **Max iterations**: safety limit reached
+* **Failure cases**: degenerate polytope or numerical issues
+
+### Minkowski Difference
+
+Like GJK, EPA operates in Minkowski difference space where `point = a - b` for points in shapes A and B. The closest point on this polytope to the origin gives the minimum separation.
+
+### Face Winding
+
+Faces are stored with outward-pointing normals. The algorithm uses a priority queue to efficiently find the face closest to the origin.
+
+---
+
+## Vertex Type Requirements
+
+The `VertexType` must satisfy the `EpaVector` concept:
+
+```cpp
+template<class V>
+concept EpaVector = requires(const V& a, const V& b, float s) {
+    { a - b } -> std::same_as<V>;
+    { a.cross(b) } -> std::same_as<V>;
+    { a.dot(b) } -> std::same_as<float>;
+    { -a } -> std::same_as<V>;
+    { a * s } -> std::same_as<V>;
+    { a / s } -> std::same_as<V>;
+};
+```
+
+`omath::Vector3<float>` satisfies this concept.
+
+---
+
+## Usage Examples
+
+### Basic EPA Usage
+
+```cpp
+using namespace omath::collision;
+using namespace omath::source_engine;
+
+// First, run GJK to detect collision
+MeshCollider<Mesh> collider_a(mesh_a);
+MeshCollider<Mesh> collider_b(mesh_b);
+
+auto gjk_result = GjkAlgorithm<MeshCollider<Mesh>>::check_collision(
+    collider_a,
+    collider_b
+);
+
+if (gjk_result.hit) {
+    // Collision detected, use EPA to get penetration info
+    auto epa_result = Epa<MeshCollider<Mesh>>::solve(
+        collider_a,
+        collider_b,
+        gjk_result.simplex
+    );
+    
+    if (epa_result.success) {
+        std::cout << "Penetration depth: " << epa_result.depth << "\n";
+        std::cout << "Separation normal: " 
+                  << "(" << epa_result.normal.x << ", "
+                  << epa_result.normal.y << ", "
+                  << epa_result.normal.z << ")\n";
+                  
+        // Apply separation: move A away from B
+        Vector3<float> correction = epa_result.normal * epa_result.depth;
+        mesh_a.set_origin(mesh_a.get_origin() + correction);
+    }
+}
+```
+
+### Custom Parameters
+
+```cpp
+// Use custom convergence settings
+Epa<Collider>::Params params;
+params.max_iterations = 128;    // Allow more iterations for complex shapes
+params.tolerance = 1e-5f;       // Tighter tolerance for more accuracy
+
+auto result = Epa<Collider>::solve(a, b, simplex, params);
+```
+
+### Physics Integration
+
+```cpp
+void resolve_collision(PhysicsBody& body_a, PhysicsBody& body_b) {
+    auto gjk_result = GjkAlgorithm<Collider>::check_collision(
+        body_a.collider, body_b.collider
+    );
+    
+    if (!gjk_result.hit)
+        return;  // No collision
+    
+    auto epa_result = Epa<Collider>::solve(
+        body_a.collider,
+        body_b.collider,
+        gjk_result.simplex
+    );
+    
+    if (epa_result.success) {
+        // Separate bodies
+        float mass_sum = body_a.mass + body_b.mass;
+        float ratio_a = body_b.mass / mass_sum;
+        float ratio_b = body_a.mass / mass_sum;
+        
+        body_a.position += epa_result.normal * (epa_result.depth * ratio_a);
+        body_b.position -= epa_result.normal * (epa_result.depth * ratio_b);
+        
+        // Apply collision response
+        apply_impulse(body_a, body_b, epa_result.normal);
+    }
+}
+```
+
+---
+
+## Performance Characteristics
+
+* **Time complexity**: O(k × f) where k is iterations and f is faces per iteration (typically f grows slowly)
+* **Space complexity**: O(n) where n is the number of polytope vertices (typically < 100)
+* **Typical iterations**: 4-20 for most collisions
+* **Worst case**: 64 iterations (configurable limit)
+
+### Performance Tips
+
+1. **Adjust max_iterations**: Balance accuracy vs. performance for your use case
+2. **Tolerance tuning**: Larger tolerance = faster convergence but less accurate
+3. **Shape complexity**: Simpler shapes (fewer faces) converge faster
+4. **Deep penetrations**: Require more iterations; consider broad-phase separation
+
+---
+
+## Limitations & Edge Cases
+
+* **Requires valid simplex**: Must be called with a 4-point simplex containing the origin (from successful GJK)
+* **Convex shapes only**: Like GJK, EPA only works with convex colliders
+* **Convergence failure**: Can fail to converge for degenerate or very thin shapes (check `result.success`)
+* **Numerical precision**: Extreme scale differences or very small shapes may cause issues
+* **Deep penetration**: Very deep intersections may require many iterations or fail to converge
+
+### Error Handling
+
+```cpp
+auto result = Epa<Collider>::solve(a, b, simplex);
+
+if (!result.success) {
+    // EPA failed to converge
+    // Fallback options:
+    // 1. Use a default separation (e.g., axis between centers)
+    // 2. Increase max_iterations and retry
+    // 3. Log a warning and skip this collision
+    std::cerr << "EPA failed after " << result.iterations << " iterations\n";
+}
+```
+
+---
+
+## Theory & Background
+
+### Why EPA after GJK?
+
+GJK determines **if** shapes intersect but doesn't compute penetration depth. EPA extends GJK's final simplex to find the exact depth and normal needed for:
+* **Collision response** — separating objects realistically
+* **Contact manifolds** — generating contact points for physics
+* **Constraint solving** — iterative physics solvers
+
+### Comparison with SAT
+
+| Feature | EPA | SAT (Separating Axis Theorem) |
+|---------|-----|-------------------------------|
+| Works with | Any convex shape | Polytopes (faces/edges) |
+| Penetration depth | Yes | Yes |
+| Complexity | Iterative | Per-axis projection |
+| Best for | General convex | Boxes, prisms |
+| Typical speed | Moderate | Fast (few axes) |
+
+EPA is more general; SAT is faster for axis-aligned shapes.
+
+---
+
+## Implementation Details
+
+The EPA implementation in OMath:
+* Uses a **priority queue** to efficiently find the closest face
+* Maintains face winding for consistent normals
+* Handles **edge cases**: degenerate faces, numerical instability
+* Prevents infinite loops with iteration limits
+* Returns detailed diagnostics (iteration count, polytope size)
+
+---
+
+## See Also
+
+- [GJK Algorithm Documentation](gjk_algorithm.md) - Collision detection (required before EPA)
+- [Simplex Documentation](simplex.md) - Input simplex structure
+- [MeshCollider Documentation](mesh_collider.md) - Mesh-based collider
+- [Mesh Documentation](../3d_primitives/mesh.md) - Mesh primitive
+- [Tutorials - Collision Detection](../tutorials.md#tutorial-4-collision-detection) - Complete collision tutorial
+- [API Overview](../api_overview.md) - High-level API reference
+
+---
+
+*Last updated: 13 Nov 2025*
--- a/docs/collision/gjk_algorithm.md
+++ b/docs/collision/gjk_algorithm.md
@@ -0,0 +1,216 @@
+# `omath::collision::GjkAlgorithm` — Gilbert-Johnson-Keerthi collision detection
+
+> Header: `omath/collision/gjk_algorithm.hpp`
+> Namespace: `omath::collision`
+> Depends on: `Simplex<VertexType>`, collider types with `find_abs_furthest_vertex` method
+> Algorithm: **GJK** (Gilbert-Johnson-Keerthi) for convex shape collision detection
+
+---
+
+## Overview
+
+The **GJK algorithm** determines whether two convex shapes intersect by iteratively constructing a simplex in Minkowski difference space. The algorithm is widely used in physics engines and collision detection systems due to its efficiency and robustness.
+
+`GjkAlgorithm` is a template class that works with any collider type implementing the required support function interface:
+
+* `find_abs_furthest_vertex(direction)` — returns the farthest point in the collider along the given direction.
+
+The algorithm returns a `GjkHitInfo` containing:
+* `hit` — boolean indicating whether the shapes intersect
+* `simplex` — a 4-point simplex containing the origin (valid only when `hit == true`)
+
+---
+
+## `GjkHitInfo`
+
+```cpp
+template<class VertexType>
+struct GjkHitInfo final {
+    bool hit{false};                    // true if collision detected
+    Simplex<VertexType> simplex;        // 4-point simplex (valid only if hit == true)
+};
+```
+
+The `simplex` field is only meaningful when `hit == true` and contains 4 points. This simplex can be passed to the EPA algorithm for penetration depth calculation.
+
+---
+
+## `GjkAlgorithm`
+
+```cpp
+template<class ColliderType>
+class GjkAlgorithm final {
+    using VertexType = typename ColliderType::VertexType;
+
+public:
+    // Find support vertex in Minkowski difference
+    [[nodiscard]]
+    static VertexType find_support_vertex(
+        const ColliderType& collider_a,
+        const ColliderType& collider_b,
+        const VertexType& direction
+    );
+
+    // Check if two convex shapes intersect
+    [[nodiscard]]
+    static GjkHitInfo<VertexType> check_collision(
+        const ColliderType& collider_a,
+        const ColliderType& collider_b
+    );
+};
+```
+
+---
+
+## Collider Requirements
+
+Any type used as `ColliderType` must provide:
+
+```cpp
+// Type alias for vertex type (typically Vector3<float>)
+using VertexType = /* ... */;
+
+// Find the farthest point in world space along the given direction
+[[nodiscard]]
+VertexType find_abs_furthest_vertex(const VertexType& direction) const;
+```
+
+Common collider types:
+* `MeshCollider<MeshType>` — for arbitrary triangle meshes
+* Custom colliders for spheres, boxes, capsules, etc.
+
+---
+
+## Algorithm Details
+
+### Minkowski Difference
+
+GJK operates in the **Minkowski difference** space `A - B`, where a point in this space represents the difference between points in shapes A and B. The shapes intersect if and only if the origin lies within this Minkowski difference.
+
+### Support Function
+
+The support function finds the point in the Minkowski difference farthest along a given direction:
+
+```cpp
+support(A, B, dir) = A.furthest(dir) - B.furthest(-dir)
+```
+
+This is computed by `find_support_vertex`.
+
+### Simplex Iteration
+
+The algorithm builds a simplex incrementally:
+1. Start with an initial direction (typically vector between shape centers)
+2. Add support vertices in directions that move the simplex toward the origin
+3. Simplify the simplex to keep only points closest to the origin
+4. Repeat until either:
+   * Origin is contained (collision detected, returns 4-point simplex)
+   * No progress can be made (no collision)
+
+Maximum 64 iterations are performed to prevent infinite loops in edge cases.
+
+---
+
+## Usage Examples
+
+### Basic Collision Check
+
+```cpp
+using namespace omath::collision;
+using namespace omath::source_engine;
+
+// Create mesh colliders
+Mesh mesh_a = /* ... */;
+Mesh mesh_b = /* ... */;
+
+MeshCollider collider_a(mesh_a);
+MeshCollider collider_b(mesh_b);
+
+// Check for collision
+auto result = GjkAlgorithm<MeshCollider<Mesh>>::check_collision(
+    collider_a,
+    collider_b
+);
+
+if (result.hit) {
+    std::cout << "Collision detected!\n";
+    // Can pass result.simplex to EPA for penetration depth
+}
+```
+
+### Combined with EPA
+
+```cpp
+auto gjk_result = GjkAlgorithm<Collider>::check_collision(a, b);
+
+if (gjk_result.hit) {
+    // Get penetration depth and normal using EPA
+    auto epa_result = Epa<Collider>::solve(
+        a, b, gjk_result.simplex
+    );
+    
+    if (epa_result.success) {
+        std::cout << "Penetration depth: " << epa_result.depth << "\n";
+        std::cout << "Separation normal: " << epa_result.normal << "\n";
+    }
+}
+```
+
+---
+
+## Performance Characteristics
+
+* **Time complexity**: O(k) where k is the number of iterations (typically < 20 for most cases)
+* **Space complexity**: O(1) — only stores a 4-point simplex
+* **Best case**: 4-8 iterations for well-separated objects
+* **Worst case**: 64 iterations (hard limit)
+* **Cache efficient**: operates on small fixed-size data structures
+
+### Optimization Tips
+
+1. **Initial direction**: Use vector between shape centers for faster convergence
+2. **Early exit**: GJK quickly rejects non-intersecting shapes
+3. **Warm starting**: Reuse previous simplex for continuous collision detection
+4. **Broad phase**: Use spatial partitioning before GJK (AABB trees, grids)
+
+---
+
+## Limitations & Edge Cases
+
+* **Convex shapes only**: GJK only works with convex colliders. For concave shapes, decompose into convex parts or use a mesh collider wrapper.
+* **Degenerate simplices**: The algorithm handles degenerate cases, but numerical precision can cause issues with very thin or flat shapes.
+* **Iteration limit**: Hard limit of 64 iterations prevents infinite loops but may miss collisions in extreme cases.
+* **Zero-length directions**: The simplex update logic guards against zero-length vectors, returning safe fallbacks.
+
+---
+
+## Vertex Type Requirements
+
+The `VertexType` must satisfy the `GjkVector` concept (defined in `simplex.hpp`):
+
+```cpp
+template<class V>
+concept GjkVector = requires(const V& a, const V& b) {
+    { -a } -> std::same_as<V>;
+    { a - b } -> std::same_as<V>;
+    { a.cross(b) } -> std::same_as<V>;
+    { a.point_to_same_direction(b) } -> std::same_as<bool>;
+};
+```
+
+`omath::Vector3<float>` satisfies this concept.
+
+---
+
+## See Also
+
+- [EPA Algorithm Documentation](epa_algorithm.md) - Penetration depth calculation
+- [Simplex Documentation](simplex.md) - Simplex data structure
+- [MeshCollider Documentation](mesh_collider.md) - Mesh-based collider
+- [Mesh Documentation](../3d_primitives/mesh.md) - Mesh primitive
+- [LineTracer Documentation](line_tracer.md) - Ray-triangle intersection
+- [Tutorials - Collision Detection](../tutorials.md#tutorial-4-collision-detection) - Complete collision tutorial
+
+---
+
+*Last updated: 13 Nov 2025*
--- a/docs/collision/mesh_collider.md
+++ b/docs/collision/mesh_collider.md
@@ -0,0 +1,371 @@
+# `omath::collision::MeshCollider` — Convex hull collider for meshes
+
+> Header: `omath/collision/mesh_collider.hpp`
+> Namespace: `omath::collision`
+> Depends on: `omath::primitives::Mesh`, `omath::Vector3<T>`
+> Purpose: wrap a mesh to provide collision detection support for GJK/EPA
+
+---
+
+## Overview
+
+`MeshCollider` wraps a `Mesh` object to provide the **support function** interface required by the GJK and EPA collision detection algorithms. The support function finds the vertex of the mesh farthest along a given direction, which is essential for constructing Minkowski difference simplices.
+
+**Important**: `MeshCollider` assumes the mesh represents a **convex hull**. For non-convex shapes, you must either:
+* Decompose into convex parts
+* Use the convex hull of the mesh
+* Use a different collision detection algorithm
+
+---
+
+## Template Declaration
+
+```cpp
+template<class MeshType>
+class MeshCollider;
+```
+
+### MeshType Requirements
+
+The `MeshType` must be an instantiation of `omath::primitives::Mesh` or provide:
+
+```cpp
+struct MeshType {
+    using NumericType = /* float, double, etc. */;
+    
+    std::vector<Vector3<NumericType>> m_vertex_buffer;
+    
+    // Transform vertex from local to world space
+    Vector3<NumericType> vertex_to_world_space(const Vector3<NumericType>&) const;
+};
+```
+
+Common types:
+* `omath::source_engine::Mesh`
+* `omath::unity_engine::Mesh`
+* `omath::unreal_engine::Mesh`
+* `omath::frostbite_engine::Mesh`
+* `omath::iw_engine::Mesh`
+* `omath::opengl_engine::Mesh`
+
+---
+
+## Type Aliases
+
+```cpp
+using NumericType = typename MeshType::NumericType;
+using VertexType = Vector3<NumericType>;
+```
+
+* `NumericType` — scalar type (typically `float`)
+* `VertexType` — 3D vector type for vertices
+
+---
+
+## Constructor
+
+```cpp
+explicit MeshCollider(MeshType mesh);
+```
+
+Creates a collider from a mesh. The mesh is **moved** into the collider, so pass by value:
+
+```cpp
+omath::source_engine::Mesh my_mesh = /* ... */;
+MeshCollider collider(std::move(my_mesh));
+```
+
+---
+
+## Methods
+
+### `find_furthest_vertex`
+
+```cpp
+[[nodiscard]]
+const VertexType& find_furthest_vertex(const VertexType& direction) const;
+```
+
+Finds the vertex in the mesh's **local space** that has the maximum dot product with `direction`.
+
+**Algorithm**: Linear search through all vertices (O(n) where n is vertex count).
+
+**Returns**: Const reference to the vertex in `m_vertex_buffer`.
+
+---
+
+### `find_abs_furthest_vertex`
+
+```cpp
+[[nodiscard]]
+VertexType find_abs_furthest_vertex(const VertexType& direction) const;
+```
+
+Finds the vertex farthest along `direction` and transforms it to **world space**. This is the primary method used by GJK/EPA.
+
+**Steps**:
+1. Find furthest vertex in local space using `find_furthest_vertex`
+2. Transform to world space using `mesh.vertex_to_world_space()`
+
+**Returns**: Vertex position in world coordinates.
+
+**Usage in GJK**:
+```cpp
+// GJK support function for Minkowski difference
+VertexType support = collider_a.find_abs_furthest_vertex(direction) 
+                   - collider_b.find_abs_furthest_vertex(-direction);
+```
+
+---
+
+## Usage Examples
+
+### Basic Collision Detection
+
+```cpp
+using namespace omath::collision;
+using namespace omath::source_engine;
+
+// Create meshes with vertex data
+std::vector<Vector3<float>> vbo_a = {
+    {-1, -1, -1}, {1, -1, -1}, {1, 1, -1}, {-1, 1, -1},
+    {-1, -1,  1}, {1, -1,  1}, {1, 1,  1}, {-1, 1,  1}
+};
+std::vector<Vector3<std::size_t>> vao_a = /* face indices */;
+
+Mesh mesh_a(vbo_a, vao_a);
+mesh_a.set_origin({0, 0, 0});
+
+Mesh mesh_b(vbo_b, vao_b);
+mesh_b.set_origin({5, 0, 0});  // Positioned away from mesh_a
+
+// Wrap in colliders
+MeshCollider<Mesh> collider_a(std::move(mesh_a));
+MeshCollider<Mesh> collider_b(std::move(mesh_b));
+
+// Run GJK
+auto result = GjkAlgorithm<MeshCollider<Mesh>>::check_collision(
+    collider_a, collider_b
+);
+
+if (result.hit) {
+    std::cout << "Collision detected!\n";
+}
+```
+
+### With EPA for Penetration Depth
+
+```cpp
+auto gjk_result = GjkAlgorithm<MeshCollider<Mesh>>::check_collision(
+    collider_a, collider_b
+);
+
+if (gjk_result.hit) {
+    auto epa_result = Epa<MeshCollider<Mesh>>::solve(
+        collider_a, collider_b, gjk_result.simplex
+    );
+    
+    if (epa_result.success) {
+        std::cout << "Penetration: " << epa_result.depth << " units\n";
+        std::cout << "Normal: " << epa_result.normal << "\n";
+    }
+}
+```
+
+### Custom Mesh Creation
+
+```cpp
+// Create a simple box mesh
+std::vector<Vector3<float>> box_vertices = {
+    {-0.5f, -0.5f, -0.5f}, { 0.5f, -0.5f, -0.5f},
+    { 0.5f,  0.5f, -0.5f}, {-0.5f,  0.5f, -0.5f},
+    {-0.5f, -0.5f,  0.5f}, { 0.5f, -0.5f,  0.5f},
+    { 0.5f,  0.5f,  0.5f}, {-0.5f,  0.5f,  0.5f}
+};
+
+std::vector<Vector3<std::size_t>> box_indices = {
+    {0, 1, 2}, {0, 2, 3},  // Front face
+    {4, 6, 5}, {4, 7, 6},  // Back face
+    {0, 4, 5}, {0, 5, 1},  // Bottom face
+    {2, 6, 7}, {2, 7, 3},  // Top face
+    {0, 3, 7}, {0, 7, 4},  // Left face
+    {1, 5, 6}, {1, 6, 2}   // Right face
+};
+
+using namespace omath::source_engine;
+Mesh box_mesh(box_vertices, box_indices);
+box_mesh.set_origin({10, 0, 0});
+box_mesh.set_scale({2, 2, 2});
+
+MeshCollider<Mesh> box_collider(std::move(box_mesh));
+```
+
+### Oriented Collision
+
+```cpp
+// Create rotated mesh
+Mesh mesh(vertices, indices);
+mesh.set_origin({5, 5, 5});
+mesh.set_scale({1, 1, 1});
+
+// Set rotation (engine-specific angles)
+ViewAngles rotation;
+rotation.pitch = PitchAngle::from_degrees(45.0f);
+rotation.yaw = YawAngle::from_degrees(30.0f);
+mesh.set_rotation(rotation);
+
+// Collider automatically handles transformation
+MeshCollider<Mesh> collider(std::move(mesh));
+
+// Support function returns world-space vertices
+auto support = collider.find_abs_furthest_vertex({0, 1, 0});
+```
+
+---
+
+## Performance Considerations
+
+### Linear Search
+
+`find_furthest_vertex` performs a **linear search** through all vertices:
+* **Time complexity**: O(n) per support query
+* **GJK iterations**: ~10-20 support queries per collision test
+* **Total cost**: O(k × n) where k is GJK iterations
+
+For meshes with many vertices (>1000), consider:
+* Using simpler proxy geometry (bounding box, convex hull with fewer vertices)
+* Pre-computing hierarchical structures
+* Using specialized collision shapes when possible
+
+### Caching Opportunities
+
+The implementation uses `std::ranges::max_element`, which is cache-friendly for contiguous vertex buffers. For optimal performance:
+* Store vertices contiguously in memory
+* Avoid pointer-based or scattered vertex storage
+* Consider SoA (Structure of Arrays) layout for SIMD optimization
+
+### World Space Transformation
+
+The `vertex_to_world_space` call involves matrix multiplication:
+* **Cost**: ~15-20 floating-point operations per vertex
+* **Optimization**: The mesh caches its transformation matrix
+* **Update cost**: Only recomputed when origin/rotation/scale changes
+
+---
+
+## Limitations & Edge Cases
+
+### Convex Hull Requirement
+
+**Critical**: GJK/EPA only work with **convex shapes**. If your mesh is concave:
+
+#### Option 1: Convex Decomposition
+```cpp
+// Decompose concave mesh into convex parts
+std::vector<Mesh> convex_parts = decompose_mesh(concave_mesh);
+
+for (const auto& part : convex_parts) {
+    MeshCollider collider(part);
+    // Test each part separately
+}
+```
+
+#### Option 2: Use Convex Hull
+```cpp
+// Compute convex hull of vertices
+auto hull_vertices = compute_convex_hull(mesh.m_vertex_buffer);
+Mesh hull_mesh(hull_vertices, hull_indices);
+MeshCollider collider(std::move(hull_mesh));
+```
+
+#### Option 3: Different Algorithm
+Use triangle-based collision (e.g., LineTracer) for true concave support.
+
+### Empty Mesh
+
+Behavior is undefined if `m_vertex_buffer` is empty. Always ensure:
+```cpp
+assert(!mesh.m_vertex_buffer.empty());
+MeshCollider collider(std::move(mesh));
+```
+
+### Degenerate Meshes
+
+* **Single vertex**: Treated as a point (degenerates to sphere collision)
+* **Two vertices**: Line segment (may cause GJK issues)
+* **Coplanar vertices**: Flat mesh; EPA may have convergence issues
+
+**Recommendation**: Use at least 4 non-coplanar vertices for robustness.
+
+---
+
+## Coordinate Systems
+
+`MeshCollider` supports different engine coordinate systems through the `MeshTrait`:
+
+| Engine | Up Axis | Handedness | Rotation Order |
+|--------|---------|------------|----------------|
+| Source Engine | Z | Right-handed | Pitch/Yaw/Roll |
+| Unity | Y | Left-handed | Pitch/Yaw/Roll |
+| Unreal | Z | Left-handed | Roll/Pitch/Yaw |
+| Frostbite | Y | Right-handed | Pitch/Yaw/Roll |
+| IW Engine | Z | Right-handed | Pitch/Yaw/Roll |
+| OpenGL | Y | Right-handed | Pitch/Yaw/Roll |
+
+The `vertex_to_world_space` method handles these differences transparently.
+
+---
+
+## Advanced Usage
+
+### Custom Support Function
+
+For specialized collision shapes, implement a custom collider:
+
+```cpp
+class SphereCollider {
+public:
+    using VertexType = Vector3<float>;
+    
+    Vector3<float> center;
+    float radius;
+    
+    VertexType find_abs_furthest_vertex(const VertexType& direction) const {
+        auto normalized = direction.normalized();
+        return center + normalized * radius;
+    }
+};
+
+// Use with GJK/EPA
+auto result = GjkAlgorithm<SphereCollider>::check_collision(sphere_a, sphere_b);
+```
+
+### Debugging Support Queries
+
+```cpp
+class DebugMeshCollider : public MeshCollider<Mesh> {
+public:
+    using MeshCollider::MeshCollider;
+    
+    VertexType find_abs_furthest_vertex(const VertexType& direction) const {
+        auto result = MeshCollider::find_abs_furthest_vertex(direction);
+        std::cout << "Support query: direction=" << direction 
+                  << " -> vertex=" << result << "\n";
+        return result;
+    }
+};
+```
+
+---
+
+## See Also
+
+- [GJK Algorithm Documentation](gjk_algorithm.md) - Uses `MeshCollider` for collision detection
+- [EPA Algorithm Documentation](epa_algorithm.md) - Uses `MeshCollider` for penetration depth
+- [Simplex Documentation](simplex.md) - Data structure used by GJK
+- [Mesh Documentation](../3d_primitives/mesh.md) - Underlying mesh primitive
+- [Tutorials - Collision Detection](../tutorials.md#tutorial-4-collision-detection) - Complete collision tutorial
+
+---
+
+*Last updated: 13 Nov 2025*
--- a/docs/collision/simplex.md
+++ b/docs/collision/simplex.md
@@ -0,0 +1,327 @@
+# `omath::collision::Simplex` — Fixed-capacity simplex for GJK/EPA
+
+> Header: `omath/collision/simplex.hpp`
+> Namespace: `omath::collision`
+> Depends on: `Vector3<float>` (or any type satisfying `GjkVector` concept)
+> Purpose: store and manipulate simplices in GJK and EPA algorithms
+
+---
+
+## Overview
+
+`Simplex` is a lightweight container for up to 4 points, used internally by the GJK and EPA collision detection algorithms. A simplex in this context is a geometric shape defined by 1 to 4 vertices:
+
+* **1 point** — a single vertex
+* **2 points** — a line segment
+* **3 points** — a triangle
+* **4 points** — a tetrahedron
+
+The GJK algorithm builds simplices incrementally to detect collisions, and EPA extends a 4-point simplex to compute penetration depth.
+
+---
+
+## Template & Concepts
+
+```cpp
+template<GjkVector VectorType = Vector3<float>>
+class Simplex final;
+```
+
+### `GjkVector` Concept
+
+The vertex type must satisfy:
+
+```cpp
+template<class V>
+concept GjkVector = requires(const V& a, const V& b) {
+    { -a } -> std::same_as<V>;
+    { a - b } -> std::same_as<V>;
+    { a.cross(b) } -> std::same_as<V>;
+    { a.point_to_same_direction(b) } -> std::same_as<bool>;
+};
+```
+
+`omath::Vector3<float>` satisfies this concept and is the default.
+
+---
+
+## Constructors & Assignment
+
+```cpp
+constexpr Simplex() = default;
+
+constexpr Simplex& operator=(std::initializer_list<VectorType> list) noexcept;
+```
+
+### Initialization
+
+```cpp
+// Empty simplex
+Simplex<Vector3<float>> s;
+
+// Initialize with points
+Simplex<Vector3<float>> s2;
+s2 = {v1, v2, v3};  // 3-point simplex (triangle)
+```
+
+**Constraint**: Maximum 4 points. Passing more triggers an assertion in debug builds.
+
+---
+
+## Core Methods
+
+### Adding Points
+
+```cpp
+constexpr void push_front(const VectorType& p) noexcept;
+```
+
+Inserts a point at the **front** (index 0), shifting existing points back. If the simplex is already at capacity (4 points), the last point is discarded.
+
+**Usage pattern in GJK**:
+```cpp
+simplex.push_front(new_support_point);
+// Now simplex[0] is the newest point
+```
+
+### Size & Capacity
+
+```cpp
+[[nodiscard]] constexpr std::size_t size() const noexcept;
+[[nodiscard]] static constexpr std::size_t capacity = 4;
+```
+
+* `size()` — current number of points (0-4)
+* `capacity` — maximum points (always 4)
+
+### Element Access
+
+```cpp
+[[nodiscard]] constexpr VectorType& operator[](std::size_t index) noexcept;
+[[nodiscard]] constexpr const VectorType& operator[](std::size_t index) const noexcept;
+```
+
+Access points by index. **No bounds checking** — index must be `< size()`.
+
+```cpp
+if (simplex.size() >= 2) {
+    auto edge = simplex[1] - simplex[0];
+}
+```
+
+### Iterators
+
+```cpp
+[[nodiscard]] constexpr auto begin() noexcept;
+[[nodiscard]] constexpr auto end() noexcept;
+[[nodiscard]] constexpr auto begin() const noexcept;
+[[nodiscard]] constexpr auto end() const noexcept;
+```
+
+Standard iterator support for range-based loops:
+
+```cpp
+for (const auto& vertex : simplex) {
+    std::cout << vertex << "\n";
+}
+```
+
+---
+
+## GJK-Specific Methods
+
+These methods implement the core logic for simplifying simplices in the GJK algorithm.
+
+### `contains_origin`
+
+```cpp
+[[nodiscard]] constexpr bool contains_origin() noexcept;
+```
+
+Determines if the origin lies within the current simplex. This is the **core GJK test**: if true, the shapes intersect.
+
+* For a **1-point** simplex, always returns `false` (can't contain origin)
+* For a **2-point** simplex (line), checks if origin projects onto the segment
+* For a **3-point** simplex (triangle), checks if origin projects onto the triangle
+* For a **4-point** simplex (tetrahedron), checks if origin is inside
+
+**Side effect**: Simplifies the simplex by removing points not needed to maintain proximity to the origin. After calling, `size()` may have decreased.
+
+**Return value**: 
+* `true` — origin is contained (collision detected)
+* `false` — origin not contained; simplex has been simplified toward origin
+
+### `next_direction`
+
+```cpp
+[[nodiscard]] constexpr VectorType next_direction() const noexcept;
+```
+
+Computes the next search direction for GJK. This is the direction from the simplex toward the origin, used to query the next support point.
+
+* Must be called **after** `contains_origin()` returns `false`
+* Behavior is **undefined** if called when `size() == 0` or when origin is already contained
+
+---
+
+## Usage Examples
+
+### GJK Iteration (Simplified)
+
+```cpp
+Simplex<Vector3<float>> simplex;
+Vector3<float> direction{1, 0, 0};  // Initial search direction
+
+for (int i = 0; i < 64; ++i) {
+    // Get support point in current direction
+    auto support = find_support_vertex(collider_a, collider_b, direction);
+    
+    // Check if we made progress
+    if (support.dot(direction) <= 0)
+        break;  // No collision possible
+    
+    simplex.push_front(support);
+    
+    // Check if simplex contains origin
+    if (simplex.contains_origin()) {
+        // Collision detected!
+        assert(simplex.size() == 4);
+        return GjkHitInfo{true, simplex};
+    }
+    
+    // Get next search direction
+    direction = simplex.next_direction();
+}
+
+// No collision
+return GjkHitInfo{false, {}};
+```
+
+### Manual Simplex Construction
+
+```cpp
+using Vec3 = Vector3<float>;
+
+Simplex<Vec3> simplex;
+simplex = {
+    Vec3{0.0f, 0.0f, 0.0f},
+    Vec3{1.0f, 0.0f, 0.0f},
+    Vec3{0.0f, 1.0f, 0.0f},
+    Vec3{0.0f, 0.0f, 1.0f}
+};
+
+assert(simplex.size() == 4);
+
+// Check if origin is inside this tetrahedron
+bool has_collision = simplex.contains_origin();
+```
+
+### Iterating Over Points
+
+```cpp
+void print_simplex(const Simplex<Vector3<float>>& s) {
+    std::cout << "Simplex with " << s.size() << " points:\n";
+    for (std::size_t i = 0; i < s.size(); ++i) {
+        const auto& p = s[i];
+        std::cout << "  [" << i << "] = (" 
+                  << p.x << ", " << p.y << ", " << p.z << ")\n";
+    }
+}
+```
+
+---
+
+## Implementation Details
+
+### Simplex Simplification
+
+The `contains_origin()` method implements different tests based on simplex size:
+
+#### Line Segment (2 points)
+
+Checks if origin projects onto segment `[A, B]`:
+* If yes, keeps both points
+* If no, keeps only the closer point
+
+#### Triangle (3 points)
+
+Tests the origin against the triangle plane and edges using cross products. Simplifies to:
+* The full triangle if origin projects onto its surface
+* An edge if origin is closest to that edge
+* A single vertex otherwise
+
+#### Tetrahedron (4 points)
+
+Tests origin against all four faces:
+* If origin is inside, returns `true` (collision)
+* If outside, reduces to the face/edge/vertex closest to origin
+
+### Direction Calculation
+
+The `next_direction()` method computes:
+* For **line**: perpendicular from line toward origin
+* For **triangle**: perpendicular from triangle toward origin
+* Implementation uses cross products and projections to avoid sqrt when possible
+
+---
+
+## Performance Characteristics
+
+* **Storage**: Fixed 4 × `sizeof(VectorType)` + size counter
+* **Push front**: O(n) where n is current size (max 4, so effectively O(1))
+* **Contains origin**: O(1) for each case (line, triangle, tetrahedron)
+* **Next direction**: O(1) — simple cross products and subtractions
+* **No heap allocations**: All storage is inline
+
+**constexpr**: All methods are `constexpr`, enabling compile-time usage where feasible.
+
+---
+
+## Edge Cases & Constraints
+
+### Degenerate Simplices
+
+* **Zero-length edges**: Can occur if support points coincide. The algorithm handles this by checking `point_to_same_direction` before divisions.
+* **Collinear points**: Triangle simplification detects and handles collinear cases by reducing to a line.
+* **Flat tetrahedron**: If the 4th point is coplanar with the first 3, the origin containment test may have reduced precision.
+
+### Assertions
+
+* **Capacity**: `operator=` asserts `list.size() <= 4` in debug builds
+* **Index bounds**: No bounds checking in release builds — ensure `index < size()`
+
+### Thread Safety
+
+* **Read-only**: Safe to read from multiple threads
+* **Modification**: Not thread-safe; synchronize writes externally
+
+---
+
+## Relationship to GJK & EPA
+
+### In GJK
+
+* Starts empty or with an initial point
+* Grows via `push_front` as support points are added
+* Shrinks via `contains_origin` as it's simplified
+* Once it reaches 4 points and contains origin, GJK succeeds
+
+### In EPA
+
+* Takes a 4-point simplex from GJK as input
+* Uses the tetrahedron as the initial polytope
+* Does not directly use the `Simplex` class for expansion (EPA maintains a more complex polytope structure)
+
+---
+
+## See Also
+
+- [GJK Algorithm Documentation](gjk_algorithm.md) - Uses `Simplex` for collision detection
+- [EPA Algorithm Documentation](epa_algorithm.md) - Takes 4-point `Simplex` as input
+- [MeshCollider Documentation](mesh_collider.md) - Provides support function for GJK/EPA
+- [Vector3 Documentation](../linear_algebra/vector3.md) - Default vertex type
+- [Tutorials - Collision Detection](../tutorials.md#tutorial-4-collision-detection) - Collision tutorial
+
+---
+
+*Last updated: 13 Nov 2025*