Luth Engine

July 12, 2026 · View on GitHub

Long-tail wishlist of unscheduled work, grouped by area. Nothing here has a target version or scheduled slot — items move to ROADMAP.md when they become "next." Pulled here when planning a series so we don't lose the seed.

For what's done or planned next, see ROADMAP.md. For per-system architecture references, see arch/. For shipped history, see history/.

Physics maturity

Carried out of the jolt-physics v2.10.x series. Land when project data demands them.

Item	Notes
`feat/jolt-cooked-shapes`	On-disk persistence via `JPH::Shape::SaveBinaryState` sidecar/appended chunk keyed by `(modelUUID, meshIndex, shapeKind)`. `Physics::ShapeCache` key is already content-addressable, so the migration is a "load-from-blob" branch with no engine API changes. Trigger: startup time / RAM from in-memory shape rebuild becomes a complaint.
`jolt-character-tier-2`	`CharacterContactListener`, dynamic-body push, crouch, swim. Trigger: first project that needs character ↔ dynamic-body interaction.
Per-mesh `PhysicsBakeMode` override	Currently per-model. Add `ConvexHullPerMesh` / `MeshShapePerMesh` enum modes when authoring asks for it.
`JPH::MutableCompoundShape` composition	Multiple Colliders per entity / child-entity composition. Trigger: first model needing multi-part collision (vehicle chassis, articulated prop).
ConvexDecomposition (V-HACD)	Non-convex dynamic shapes that can't be a single ConvexHull. Trigger: first non-convex dynamic object that can't be hand-decomposed.
`PhysicsMaterial` cook parameters	Active-edge angle, double-sided, per-triangle user data. Pairs with on-disk cook landing.
`PhysicsMaterial`-per-shape	Currently per-body via `RigidBody.materialUUID`; per-shape is the natural next refinement (different friction on different sides of one body, character-material slot).
Heightfield / terrain shapes	`JPH::HeightFieldShape` for streamed terrain. Trigger: terrain system arrives.
Dynamic mesh updates	Skinned mesh as physics, deformable shapes. Trigger: cloth / soft-body system arrives.
`JPH::Shape` LOD / decimation	Distance-based collider LODs. Trigger: profiler flags it.

Gameplay enablement

Item	Depends on	Notes
Scripting (C# via Mono, or Lua)	`play-mode` ✅, `jolt-physics` ✅	Top priority of this category — deletes the `PlayerControllerSystem` stub. Big design surface (language pick, embedding, hot-reload, ECS bindings).
Prefab System	`play-mode` ✅, scripting (recommended)	Entity templates with override tracking.
Ragdoll	`jolt-physics` ✅, `animation-controller-v2`	Bone driven by Jolt body during ragdoll state.

Rendering (beyond planned-epic deps)

Item	Depends on	Notes
Full deferred shading	`rt-renderer` (slim G-buffer lands in Phase A.2)	If the forward path hits material-variant complexity limits; the slim G-buffer is already there feeding RT denoising, motion vectors, and SSR-replaced-by-RT-reflections. Full deferred would extend it with albedo + metallic/roughness + emissive + opaque shading model ID.
HZB Occlusion Culling	`compute-gpu-culling` ✅, `gpu-driven` series	Two-phase cull pipeline; depth pyramid as compute. Lands as part of the post-`rt-renderer` `gpu-driven` series (mesh shaders + GPU-driven culling).
Volume decals (`DecalVolume` component)	forward+ clustered (like lights); scene component like `FogVolume`	Box-projected decals as a scene component (transform + decal texture), gathered + cluster-binned like lights, applied in the forward PBR pass (the slim G-buffer has no albedo, so classic deferred-decal G-buffer writes don't fit). Key call: the raster==RT stance - raster-only is simple but decals miss RT reflections/GI; RT-correct needs decal projection at ray hits too (novel, the design crux). Deferred from surface-detail (#175, 2026-07-12); UV-space material decals shipped there instead as a complementary node. L/XL.

Removed because absorbed into rt-renderer: Volumetric Fog → Phase A.4 (Wronski full voxel volume), Global Illumination → Phase C (ReSTIR DI + GI), SSR → Phase D.1 (RT reflections supersede screen-space). Shadow frustum-union fit removed because Phase B.3 retires CSM entirely.

rt-renderer Phase A polish & follow-ups

Items deferred from Phase A history files (v3.0.0–v3.0.7). None on critical path; pick up opportunistically or fold into the next polish pass.

Item	Source	Trigger
Salvi K4 variance-clip TAA upgrade	image-quality v3.0.7	When RT specular pinpoints surface (Phase D RT reflections)
Mitchell-Netravali TAA post-sharpen pass	image-quality v3.0.7	After Phase A.5; close remaining TAA blur gap
TAA debug viz ShadeMode (`TaaHistory`)	image-quality v3.0.7	Diagnostic convenience
Blackman-Harris weight normalization (sum 0.9956 → 1.0)	image-quality v3.0.7	Imperceptible drift; opportunistic
AgX exposure-aware curve fit (Filament's expanded range)	image-quality v3.0.7	If users push `pp.exposure` past ±2 stops of unity
Slim G-buffer cutout coverage	slim-gbuffer v3.0.1	When cutout-heavy scene surfaces (foliage / glass cards)
Per-froxel light culling for fog	tracked as #134; volumetric-fog-polish v3.0.6	When a scene shows residual fog Z-banding
Volumetric per-pass timing in RenderPanel	volumetric-fog-polish v3.0.6	Convenience; FrameDebugger already shows pass timings
Sun-fog absorption via world-space LUT	tracked as #133; volumetric-fog-polish v3.0.6	When god rays through low fog with off-frustum density expose the artifact
IBL ambient multi-direction sampling (currently +Y only)	volumetric-fog-polish v3.0.6	Matters at fog scales rarely
RG `lastReader` / first-consumer cross-queue sync (foundational RG change)	vulkan-sync-hardening v3.0.11	Research-validated: NOT needed for correctness — Fix A's `ALL_COMMANDS` cross-queue waits cleared every cross-queue hazard. Full UE5-RDG first-consumer placement needs per-pass semaphores. Pick up when ReSTIR/SVGF surfaces concrete cross-queue-WAR cases + tests
N-buffer high-contention cross-queue persistent resources	vulkan-sync-hardening v3.0.11	Relaxes the first-view gA → prev-frame-compute cross-frame wait so frame N+1 graphics overlaps frame N compute. Trigger: profiler flags the stall, or ReSTIR/SVGF history buffers multiply contention
VKTexture creation transition → sync2 + cleaner initial layout	vulkan-sync-hardening v3.0.11	The 9 best-practices `UNDEFINED→read-only` warnings are VKTexture's one-time legacy `vkCmdPipelineBarrier` creation transitions — benign (GPU-AV-clean, written before read) but flagged. Migrate to `VkImageMemoryBarrier2` + transition to write/general first
Small-buffer VMA sub-allocation	vulkan-sync-hardening v3.0.11	The 6 best-practices `vkBindBufferMemory` dedicated-alloc hints (<1 MB); tagged-heap by design. VMA pool / sub-allocation if validation cleanliness is wanted
Memory Budget table refresh in `arch/rendering-pipeline.md`	this audit	Stale post-arc; Light UBO removed, GBuffer sizes wrong, volumetric/TAA/IBL/BoneMatrix-dual not listed

Animation maturity (post `animation-controller-v2`)

Item	Depends on	Notes
Animation v3 (DQS, IK, morph targets)	`animation-controller-v2`	Beyond state machine + blend trees.

Audio

Item	Depends on	Notes
3D Spatial Audio	`play-mode` ✅	Orthogonal; integrate when demo needs it.
Audio asset pipeline	—	`.wav` / `.ogg` importer + AssetManager dispatch.

Editor & Tools

Item	Depends on	Notes
Asset Streaming	`compute-gpu-culling` ✅	Async GPU upload via transfer queue.
Frame Debugger per-view capture	`game-panel` ✅	Capture currently scene-view-only; extend tracked RTs + archive slots to tag by view.
Scene-panel post-process toggle	—	Unity-style toggle to disable bloom/tonemap/vignette for lookdev.
Auto-wire `GPUTimerPool` into RenderGraph	—	Currently per-pass insertion is manual.

Profiling / Memory

Item	Depends on	Notes
Tracked STL allocators (`LH::Vector<T>`)	—	Closes the STL gap on `MemoryTracker` (currently covered only by Tracy hooks).
GPU heap stress test	`foundation-testing` ✅	`GPUTaggedPageAllocator` V6 overflow + backing-pool growth under sustained pressure. Deferred from foundation-testing because of Vulkan-device setup overhead. CPU side covers the architectural shape.