README.md

@@ -43,6 +43,31 @@ shaded through an intersection shader with an any-hit cut-out.
 > [examples/VulkanTriangle/README.md](examples/VulkanTriangle/README.md)
 > for the full investigation. WebGPU RT is unaffected.
 
+> **Native RT limitation — dynamic `descriptor_heap` indexing in hit
+> shaders:** on the same NVIDIA driver, indexing a `descriptor_heap`
+> array with a **runtime (non-constant)** index inside a ray-tracing
+> **hit** shader also device-losts (`VK_ERROR_DEVICE_LOST`), for plain
+> SSBO **and** sampled-image descriptors. A **constant / spec-constant**
+> index is fine (that's why [Sponza](examples/Sponza/README.md)'s
+> closest-hit reads `albedo[albedoSlot]` through a spec constant), and
+> the identical dynamic pattern works in fragment shaders (the UI
+> renderer indexes `uiTextures[]` by per-item runtime slots) — so this
+> is **RT-stage-specific**, not a general heap problem. Unlike the
+> AS-read fault above this **cannot** be worked around transparently:
+> sampled images have no device-address escape hatch the way an
+> acceleration structure does (`OpConvertUToAccelerationStructureKHR`).
+> The recommended pattern for bindless per-mesh geometry/material is to
+> **bind one resource and index *within* it dynamically** rather than
+> selecting a descriptor dynamically: pack geometry into a single SSBO
+> (or reach it via `buffer_reference`) at a spec-constant slot and index
+> by element offset, and put materials in one `texture2DArray` indexed
+> by layer. Dynamic addressing *inside* a bound resource is ordinary
+> memory/layer addressing and is unaffected; only dynamic selection of a
+> *descriptor* faults. This is exactly what the WebGPU path already does
+> (bucketed texture arrays + a single buffer). Full investigation and
+> GLSL in [examples/Sponza/README.md](examples/Sponza/README.md) (issue
+> #23). WebGPU RT is unaffected.
+
 ## What's in here
 
 - **Window** — Wayland, Win32, and DOM backends, swapchain ring / canvas

examples/Sponza/README.md

@@ -14,6 +14,117 @@ it via ray tracing on both Vulkan (native) and WebGPU (wasm). Same
   at the barycentric attribs as UVs — proof-of-binding, not visually
   accurate. Per-vertex UV interpolation is the next step.
 
+The closest-hit reads its texture through a **spec constant**
+(`albedo[albedoSlot]`), not a runtime index. That is deliberate — see
+below.
+
+## Native RT limitation: dynamic `descriptor_heap` indexing in hit shaders
+
+On NVIDIA driver `610.43.02` (Vulkan 1.4), indexing a
+`layout(descriptor_heap)` array with a **runtime (non-constant)** index
+inside a ray-tracing **hit** shader aborts the device with
+`VK_ERROR_DEVICE_LOST` (an instruction-pointer / `READ_INVALID`
+device-fault) with validation off. GPU-Assisted Validation masks it —
+the scene runs fine under GPU-AV — which is why a validated run doesn't
+catch it. It is a **driver-side fault**, the same family as the
+descriptor-heap AS-read fault (#7 / #15) and the RT recursion / compute
+TLAS-push issues (#21 / #22), but here for plain **SSBO and
+sampled-image** descriptors read with a non-constant heap index
+(issue #23).
+
+### What was isolated (NVIDIA RTX 4090, driver `610.43.02`)
+
+Driving a native bindless RT scene headlessly and bisecting the
+closest-hit:
+
+- A closest-hit that reads only `lightHeap[lightSlot]` where `lightSlot`
+  is a **spec constant** survives indefinitely. ✅ (This example's
+  `albedo[albedoSlot]` is exactly this case.)
+- Reading `indexHeap[assetIndexStart + gl_InstanceCustomIndexEXT]` /
+  `vertexHeap[...]` — a heap index offset by a **runtime** value —
+  device-losts on the first geometry hit. ❌
+- Reading a **texture** dynamically,
+  `textureHeap[assetColorStart + gl_InstanceCustomIndexEXT]`, also
+  device-losts. ❌ So it is SSBO *and* sampled-image descriptors.
+- `nonuniformEXT()` on the dynamic index does **not** help.
+- The identical dynamic-heap-index pattern works fine in **fragment**
+  shaders (the UI renderer indexes `uiTextures[]` / `ui*Heap[]` by
+  per-item runtime slots), so this is **RT-stage-specific**, not a
+  general `descriptor_heap` problem.
+- Reading a spec-constant-indexed SSBO in **raygen** works; only the
+  *dynamic* index in the hit stage faults.
+
+### Why there is no transparent engine workaround
+
+The AS-read fault (#15) is worked around transparently because an
+acceleration structure can be reached two ways: through a descriptor, or
+through its **device address** via
+`OpConvertUToAccelerationStructureKHR` (which reads no descriptor). There
+is exactly one TLAS, so the engine rewrites the heap AS read into an
+address load and feeds the address in as push data.
+
+Neither half of that applies here:
+
+- **Sampled images have no device-address path.** A texture *must* be
+  reached through a descriptor; there is no `OpConvertUToImage`. A
+  dynamic heap texture index cannot be rewritten into anything that
+  avoids dynamic descriptor selection.
+- **There are many buffers, dynamically selected.** SSBOs *can* be
+  reached by address (`buffer_reference` / `OpConvertUToPtr`), but a
+  per-mesh array selected by `gl_InstanceCustomIndexEXT` would need the
+  engine to maintain and bind an address-table buffer and a SPIR-V
+  rewrite far larger than the single-TLAS AS case — and it would still
+  leave the texture half broken.
+
+So the engine cannot paper over this the way it does the AS read. The
+fix is on the **consumer** side: avoid dynamically selecting a
+*descriptor* in a hit shader.
+
+### Recommended pattern
+
+The fault is dynamic selection of a **descriptor**. Indexing *within* a
+single bound resource — an element offset into one SSBO, a layer into
+one array texture — is ordinary memory / layer addressing and is
+**unaffected**. So bind one resource and index inside it, rather than
+indexing the heap:
+
+- **Geometry** — pack all meshes' vertices/indices into a single SSBO
+  bound at a **spec-constant** slot and index it by a runtime element
+  offset, or reach each mesh's buffer via `buffer_reference` (a device
+  address loaded from one bound table). Either way the *descriptor* is
+  constant; only the offset/address is dynamic.
+
+  ```glsl
+  // ❌ faults in a hit shader on NVIDIA: dynamic descriptor selection
+  layout(descriptor_heap) buffer Verts { Vertex v[]; } vertexHeap[];
+  Vertex vtx = vertexHeap[assetVertexStart + gl_InstanceCustomIndexEXT].v[i];
+
+  // ✅ one descriptor (spec-constant slot), dynamic element offset
+  layout(constant_id = 0) const uint16_t vertexSlot = 0us;
+  layout(descriptor_heap) buffer Verts { Vertex v[]; } vertexHeap[];
+  uint base = assetVertexStart[gl_InstanceCustomIndexEXT]; // from a bound SSBO
+  Vertex vtx = vertexHeap[vertexSlot].v[base + i];
+  ```
+
+- **Materials / textures** — put them in one `texture2DArray` (or a small
+  number of arrays bucketed by format/size) bound at a spec-constant
+  slot and index by **layer**:
+
+  ```glsl
+  // ✅ one array texture (spec-constant slot), dynamic layer index
+  layout(constant_id = 1) const uint16_t materialArraySlot = 0us;
+  layout(descriptor_heap) uniform sampler2DArray materials[];
+  uint layer = materialLayer[gl_InstanceCustomIndexEXT]; // from a bound SSBO
+  vec3 albedo = texture(materials[materialArraySlot], vec3(uv, layer)).rgb;
+  ```
+
+This is precisely what the WebGPU path already does — bucketed texture
+arrays plus a single geometry buffer — so it is a proven, cross-backend
+pattern, and it sidesteps the NVIDIA RT fault on the native path.
+
+Remove this section once a fixed NVIDIA driver ships and dynamic
+`descriptor_heap` indexing in hit shaders stops faulting.
+
 ## Asset fetch
 
 `project.cpp` calls `Crafter::GitFetch(...)` on