Clean up TODO.md

Author: llGaetanll

TODO.md

@@ -42,68 +42,6 @@
    - [x] Test cases: Lambertian diffuse, Metal reflection, Dielectric refraction (entering/exiting)
    - [x] Helps catch regressions like the normal-flip bug from instance transforms
 
-## Interactive Camera Controls (Live Mode)
-
-Add keyboard and mouse controls to move the camera in real-time during `live` mode.
-
-### Overview
-
-Currently the camera is constructed entirely on the GPU side from hardcoded parameters in each scene function. The host only passes frame dimensions, elapsed time, and cursor position via `ShaderConstants`. To enable interactive controls, camera state must live on the CPU and be passed to the shader each frame.
-
-### Phase 1: CPU-side input and logging (no shader changes)
-
-Get input handling working and validate the camera math before touching the shader.
-
-**Controls:**
-- WASD: horizontal movement (forward/back/strafe)
-- Space/C: up/down
-- Mouse: look direction (no cursor grab - just track movement while window focused)
-- Q: quit (in addition to Escape)
-
-**Implementation:**
-- [ ] Add camera state to app struct (position, yaw, pitch)
-- [ ] Track held keys (simple bools for W/A/S/D/Space/C)
-- [ ] Track mouse delta for look direction
-- [ ] Each frame: update position based on held keys and orientation, update orientation from mouse
-- [ ] Log computed camera params to stdout
-- [ ] Q key exits the loop
-
-### Phase 2: Wire up to shader
-
-Once CPU-side camera is working:
-
-- [x] Add camera fields to `ShaderConstants` (cam_pos, cam_dir, cam_fov, etc.)
-- [x] Modify shader to construct `Camera` from `ShaderConstants` instead of hardcoded values
-- [ ] Remove or bypass per-scene camera construction
-
-### Phase 3: Fix gimbal lock with quaternions
-
-The camera crashes when looking straight up or down because `vup` becomes parallel to the view direction, causing a zero-length cross product that produces NaN values. Tests in `crates/rtx-util/tests/camera_tests.rs` demonstrate this issue.
-
-**Stage 1: Add `vup` to data path**
-- [x] Add `vup: [f32; 3]` to `ShaderConstants`
-- [x] Thread `vup` through `two_spheres()` and `CameraParams`
-- [x] Host computes `vup` dynamically (switch reference vector when pitch exceeds ±45°)
-- [x] Update tests to pass with dynamic `vup`
-
-**Stage 2: Switch host to quaternions**
-- [x] Add `glam` crate dependency to host (for `Quat`)
-- [x] Replace `cam_yaw: f32` / `cam_pitch: f32` with `cam_orientation: Quat`
-- [x] Rewrite `update_camera()`:
-  - Apply yaw rotation in world space (around world Y axis)
-  - Apply pitch rotation in local space (around camera's right axis)
-  - Multiply rotations into orientation quaternion
-- [x] Extract `cam_dir` and `vup` from quaternion for shader:
-  - `cam_dir` = quaternion rotated `-Z` (forward)
-  - `vup` = quaternion rotated `+Y` (up)
-- [x] Remove pitch clamping (quaternions handle full rotation)
-- [x] Tune mouse sensitivity for natural feel
-
-### Future enhancements (not for first pass)
-
-- [ ] **Dynamic ray sampling**: Lower samples per pixel when camera/world is moving for faster feedback, then accumulate rays over time when stationary for higher quality. Requires tracking frame-to-frame camera changes and maintaining an accumulation buffer.
-- [ ] Screenshot with current camera position
-
 ## Technical Debt
 
 1. [ ] **Share types from `rtx-prim` with host**: The `rtx-prim` crate contains elementary types (`Vec3`, `Point3`, `Color`, etc.) that should be usable on both GPU and CPU. Currently `host/src/spline.rs` imports `Vec3` directly from `glam`, but ideally it would use the re-exports from `rtx-prim` to keep types consistent across the codebase. This requires making `rtx-prim` compilable for non-SPIR-V targets.

docs/tasks/benchmarking.md

@@ -8,29 +8,41 @@ Track rendering performance across commits to detect regressions.
 - `crates/host/src/camera_path.rs` - `CameraPath` and `CameraFrame` types (frame = position + look-at target)
 - `crates/host/src/bench_app.rs` - Benchmark application with render loop
 - `crates/host/src/cli.rs` - CLI definitions including `bench` subcommand
+- `benchmarks/*.toml` - Benchmark definition files
 
 ## Overview
 
-Run `cargo run --release -- bench` to execute a benchmark. The camera follows a predefined spline path through a scene, recording frame timing data. Results are saved to `bench-results/<git-sha>/<datetime>-<scene-name>.jsonl`.
+Run `cargo run --release -- bench` to execute all benchmarks. The camera follows a predefined spline path through a scene, recording frame timing data. Results are saved to `bench-results/<git-sha>/<datetime>-<benchmark-name>.jsonl`.
 
 ## Status
 
-### Completed
+### Core Infrastructure
+
 - [x] `CatmullRomSpline` - interpolates through control points
 - [x] `CameraPath` - combines position and look-at splines with duration
 - [x] `bench` CLI subcommand - runs benchmark with animated camera
 - [x] Benchmark exits after camera path completes
+
+### Output & Metadata
+
 - [x] Benchmark output file (JSONL format with metadata + per-frame records)
-- [x] `--scene` CLI argument
 - [x] Git SHA baked in at build time (7 characters)
 - [x] GPU info capture from wgpu adapter
 - [x] Frame timing (wall-clock)
 - [x] Datetime in output filename
 
-### Also Completed
+### Benchmark Definitions
+
 - [x] Benchmark definition files in `benchmarks/` directory (TOML format with scene + camera path)
+- [x] `--scene` CLI argument for single benchmark
 - [x] Output filename uses benchmark name (TOML filename) instead of scene name
-- [x] Run all benchmarks when no name specified (`cargo run --release -- bench`)
+- [x] Run all benchmarks when no name specified
+
+### Future Enhancements
+
+- [ ] Automated regression detection: compare results across commits
+- [ ] Warmup frames: discard first N frames to avoid startup costs
+- [ ] GPU timestamps: use wgpu timestamp queries for more accurate GPU timing
 
 ## Benchmark Definition Format
 
@@ -59,7 +71,7 @@ Both `position` and `look_at` require at least 4 control points for the Catmull-
 
 ## Output Format
 
-Single JSONL file per benchmark run: `bench-results/<git-sha>/<scene-name>.jsonl`
+Single JSONL file per benchmark run: `bench-results/<git-sha>/<datetime>-<benchmark-name>.jsonl`
 
 **Every line is a single JSON object, including the first line (metadata).** This allows streaming writes and easy parsing.
 
@@ -97,34 +109,3 @@ Single JSONL file per benchmark run: `bench-results/<git-sha>/<scene-name>.jsonl
 - `cam_pos` - camera position `[x, y, z]`
 - `cam_dir` - camera direction `[x, y, z]`
 - `cam_vup` - camera up vector `[x, y, z]`
-
-## Implementation Plan
-
-### Step 1: Add serde support
-- Add `serde` feature to `glam` dependency
-- Derive `Serialize` on `CameraPath`, `CatmullRomSpline`
-- Create metadata struct with GPU/CPU info
-
-### Step 2: Capture GPU info
-- Extract adapter info from wgpu (`adapter.get_info()`)
-- Store name, driver, backend
-
-### Step 3: Frame timing
-- Measure wall-clock time between frame start and `frame.present()`
-- Store in `Vec<FrameRecord>` during benchmark run
-- Note: This measures CPU submission + GPU present latency, not pure GPU render time. For more accurate GPU-only timing, wgpu timestamp queries could be added later (see Future Considerations).
-
-### Step 4: Git SHA
-- Run `git rev-parse HEAD` at startup
-- Create output directory `bench-results/<sha>/`
-
-### Step 5: Write output
-- After benchmark completes, write JSON file
-- Metadata on line 1, then one frame per line (JSONL)
-
-## Future Considerations
-
-- **Automated regression detection**: Compare results across commits
-- **Multiple runs**: Average multiple benchmark runs for stability
-- **Warmup frames**: Discard first N frames to avoid startup costs
-- **GPU timestamps**: Use wgpu timestamp queries for more accurate GPU timing

docs/tasks/live-mode.md

@@ -0,0 +1,62 @@
+# Interactive Camera Controls (Live Mode)
+
+Keyboard and mouse controls to move the camera in real-time during `live` mode.
+
+## Overview
+
+The camera state lives on the CPU and is passed to the shader each frame via `ShaderConstants`. The host tracks orientation using quaternions to avoid gimbal lock.
+
+## Relevant Files
+
+- `crates/host/src/live_app.rs` - `LiveApp` with camera state, input handling, `update_camera()`
+- `crates/shared/src/lib.rs` - `ShaderConstants` with `cam_pos`, `cam_dir`, `cam_vup`
+
+## Controls
+
+- WASD: horizontal movement (forward/back/strafe)
+- Space/C: up/down
+- Mouse: look direction
+- Q/Escape: quit
+
+## Status
+
+### Phase 1: CPU-side input and logging (no shader changes)
+
+- [x] Add camera state to app struct (position, yaw, pitch)
+- [x] Track held keys (simple bools for W/A/S/D/Space/C)
+- [x] Track mouse delta for look direction
+- [x] Each frame: update position based on held keys and orientation, update orientation from mouse
+- [x] Log computed camera params to stdout
+- [x] Q key exits the loop
+
+### Phase 2: Wire up to shader
+
+- [x] Add camera fields to `ShaderConstants` (cam_pos, cam_dir, cam_fov, etc.)
+- [x] Modify shader to construct `Camera` from `ShaderConstants` instead of hardcoded values
+- [x] Remove or bypass per-scene camera construction
+
+### Phase 3: Fix gimbal lock with quaternions
+
+The camera previously crashed when looking straight up or down because `vup` became parallel to the view direction. This was fixed with quaternion-based orientation.
+
+**Stage 1: Add `vup` to data path**
+- [x] Add `vup: [f32; 3]` to `ShaderConstants`
+- [x] Thread `vup` through `two_spheres()` and `CameraParams`
+- [x] Host computes `vup` dynamically
+- [x] Update tests to pass with dynamic `vup`
+
+**Stage 2: Switch host to quaternions**
+- [x] Add `glam` crate dependency to host (for `Quat`)
+- [x] Replace `cam_yaw: f32` / `cam_pitch: f32` with `cam_orientation: Quat`
+- [x] Rewrite `update_camera()`:
+  - Apply yaw rotation in world space (around world Y axis)
+  - Apply pitch rotation in local space (around camera's right axis)
+  - Multiply rotations into orientation quaternion
+- [x] Extract `cam_dir` and `vup` from quaternion for shader
+- [x] Remove pitch clamping (quaternions handle full rotation)
+- [x] Tune mouse sensitivity for natural feel
+
+## Future Enhancements
+
+- [ ] **Dynamic ray sampling**: Lower samples per pixel when camera/world is moving for faster feedback, then accumulate rays over time when stationary for higher quality. Requires tracking frame-to-frame camera changes and maintaining an accumulation buffer.
+- [ ] Screenshot with current camera position