Fix FOV/vision system: hard black mask, real occlusion, subsystem wiring

[henkhooft]

Summary

Gets the client FOV/fog-of-war system (ported earlier from RE-SS3D#1491 onto the URP/tilemap rework) actually working: it was rendering nothing at all, and once fixed, was leaking vision through walls and only lightly shadowing hidden areas instead of hiding them.

This branch was also brought up to date with develop (merging in the intervening atmospherics, selection, and tile-system work) before these fixes were made.

PR checklist

• The game builds properly without errors.
• No unrelated changes are present.
• No "trash" files are committed.
• Relevant code is documented.
• Update the related design or architecture doc in Documents/, or create one if needed.

Testing

1. Enter Play mode as host or client (vision is client-side; it no-ops on a dedicated IsServerOnly server).
2. Spawn/possess a character - vision binds to the local player via the existing LocalPlayerObjectChanged event.
3. Areas outside the vision cone/range should be fully opaque black, not a partial shadow.
4. Walking around, walls should correctly hide what's behind them (including another wall behind a nearer one), while doors opening should unblock vision without any extra action.
5. VisionSubSystem's obstacleMask Inspector field controls which layers raycasts can hit; hits are then filtered down to walls/doors only (see IsVisionOccluder) so tall furniture doesn't incorrectly occlude and open airlock trigger volumes don't darken FOV.

Changes

• Vertical UV flip in depth reconstruction (Vision.hlsl): VisionClipToWorld rebuilt world position from the fullscreen triangle's texture UV without correcting for UNITY_UV_STARTS_AT_TOP, so on D3D/Vulkan/Metal the reconstructed position was mirrored vertically before being projected onto the XZ polar map - desyncing the vision mask from the actual geometry under each pixel. Same class of bug already fixed once in AtmosCommon.hlsl's AtmosUnproject; mirrored the same guard here.
• VisionSubSystem was never instantiated anywhere - it existed as a script and the URP renderer feature/shaders were fully wired into the pipeline asset, but nothing created the component (SubSystems.Get<T>() is a FindObjectOfType lookup, it doesn't auto-create). Added a RuntimeInitializeOnLoadMethod self-bootstrap mirroring ScreenEffectsSubSystem's existing pattern; a scene instance was also added directly to Game.unity.
• Hard black mask instead of a soft fog: the old composite blended unseen pixels toward black at a tunable _VisionFogStrength (0.9) and ran a 25-direction x 5-sample AO-style blur per pixel (also softening the visible side of the boundary). Replaced with a straight binary composite - fully opaque black outside the mask, source color inside, nothing in between - and dropped the now-dead blur/fog-strength knobs.
• Real occlusion instead of the CPU tile grid: the original VisionGridCaster/VisionOcclusionProvider walked an abstract per-tile boolean grid with no notion of real wall shape, so a wall behind another wall (or anything whose silhouette didn't match the grid abstraction) leaked through as visible. Tried a GPU depth-cubemap capture from the player's position first (mirroring AtmosCommon.hlsl's AtmosIsSampleOccluded fix for the same class of bug in atmos), but landed on a RaycastCommand/Physics.Raycast batch from Entity.ViewPoint instead - each ray iteratively skips furniture/props (TryFindNearestOccluder) until it hits the nearest actual wall/door, capped at 64 skips so a dense prop pile can't exhaust a fixed hit buffer and falsely report a clear line of sight through a wall. Windows stay see-through, trigger volumes (e.g. airlock open triggers) are ignored, and a new Walls physics layer plus similar-depth neighbor dilation address angular-resolution stripe artifacts.
• Documented the final approach in Documents/architecture/systems/rendering.md.

Related issues/PRs

Continues the vision-system port from RE-SS3D#1491.

---

Generated by Claude Code