Graphics Developer - Photorealistic Volumetric Nebulae (Vulkan / HLSL / C#)

We're modernizing the renderer of a custom C#/.NET game engine (an open-source space sim, LibreLancer-based), Vulkan-first, HDR/PBR forward pipeline. We need a graphics engineer to own the volumetric rendering system and bring our nebulae to photorealistic quality. Every system in our game has volumetric nebulae, gas clouds the player flies through and around. Today they don't look good enough. We want good-grade, physically-plausible volumetrics: real depth and parallax on fly-by, soft multiple-scattering light, light shafts, temporally stable, across all nebula types. Current state (honest): there's an existing froxel-based volumetric pipeline, density/light/integrate/composite compute passes, temporal-reprojection scaffolding, a near cascade, and a working offline OpenVDB import path. It is incomplete, currently feature-flagged off, prone to a GPU device-loss crash, and was never art-tuned. You decide what to keep, finish, or rebuild, we care about the result, not preserving code. You'll own: A stable, photorealistic volumetric nebula renderer (froxel near-volume and/or raymarched far-volume, your call) Physically-based volumetric lighting (phase functions, multiple-scattering approximation, light shafts) Temporal stability (reprojection / temporal upsampling) with no ghosting or boiling A performance budget (target ≤ ~2.5 ms at 4K internal on RTX 5090-class), with quality tiers for weaker GPUs Fixing the existing device-loss crash in the froxel path Required: Proven real-time volumetric rendering (clouds, fog, smoke, or nebulae) — please link examples Strong Vulkan, especially compute pipelines, synchronization, and descriptor management HLSL shader authoring with the SPIR-V / DXC toolchain C# / .NET comfort (engine and render graph are C#; shaders are HLSL) GPU profiling/optimization (RenderDoc/Nsight) and experience debugging DEVICE_LOST / TDR Nice to have: OpenVDB, EmberGen/Houdini volumetric authoring, atmospheric scattering, temporal upsampling/TAA, prior work on a custom (non-Unity/Unreal) engine. The codebase is public, you can read the existing volumetric module before we talk. To apply: link one volumetric-rendering example you've built, and in 3–4 sentences describe how you'd approach photoreal nebulae here (froxel, raymarch, or hybrid and why).