"The engine ships render quality profiles that cap your scene — and post-processing is opt-in per camera"

• A scene's authored graphics settings don't render as authored: fewer shadow cascades, lower shadow resolution, no ambient occlusion, no bloom, no tonemapping — even though every value was set correctly.
• A camera created in code renders flat, ungraded, un-occluded, non-blooming lighting no matter how carefully the lights are placed.
• A performance number measured on one machine doesn't reproduce on another, because the two machines are at different engine quality levels.

Two separate engine facts combine here.

1. The engine ships render quality profiles that cap your scene. core/cfg/quality_profiles.json defines four groups — TextureQuality, PostProcessQuality, VolumetricFogQuality, ShadowQuality — each with Low/Medium/High, applied through Sandbox.Engine.Settings.RenderQualityProfiles.SetGroupConVars(group, level). These are a ceiling, not a starting point:

• DirectionalLight.ShadowCascadeCount is a request, not a setting. Its own XML doc says "User settings will set a maximum," and ShadowQuality Low pins r.shadows.csm.maxcascades 2. A scene authoring 4 cascades renders 2 for that player. The same shape applies to r.shadows.maxresolution (512/1024/2048), r.shadows.csm.maxresolution (1024/2048/4096) and r.shadows.quality (1/2/3).
• So any measured cascade or shadow cost is the cost at the measuring machine's engine setting, not a portable constant.

2. Post-processing is opt-in per camera — the convars tune passes that don't exist until you add the component. r_ao_quality, r_dof_quality, r_motionblur_quality and friends are real and settable, and setting them alone does nothing. The GTAO pass only runs when a Sandbox.AmbientOcclusion component sits on the camera. A scene whose cameras are created in code with AddComponent<CameraComponent>() and never given Tonemapping / Bloom / AmbientOcclusion renders with none of them.

The full opt-in component set is Sandbox.{AmbientOcclusion, Bloom, Blur, ChromaticAberration, ColorGrading, DepthOfField, FilmGrain, MotionBlur, Pixelate, ScreenSpaceReflections, Sharpen, Tonemapping, Vignette}, all in Sandbox.Engine.xml, with gtao_cs.shader, postprocess_bloom.shader, screen_space_reflections_cs.shader and tonemapping/ all shipping compiled in addons/base/Assets/shaders/.

Screen-space reflections follow the same opt-in rule — and this is a common trap. A non-zero r_ssr_downsample_ratio (it reads 2 on a stock machine) does not mean SSR is running. SSR needs a Sandbox.ScreenSpaceReflections component on the camera; the convar only tunes a pass that already exists. Measured across r_ssr_downsample_ratio 0, 2, and full-resolution 1 with a world rebuild per setting, frames came back identical within noise and with no mirrored geometry at any setting, because no SSR component was attached. The general rule: a non-zero engine convar is necessary but not sufficient — the component is what decides whether the pass runs.

• Add the post-process components you want, per camera. If a camera is built in code, attach Tonemapping, Bloom, AmbientOcclusion, ScreenSpaceReflections, etc. explicitly. The convars only shape passes that already exist.
• Treat shadow cascade/resolution settings as requests that the player's quality tier can cap. Don't benchmark shadow cost as a portable constant; measure it at a known engine setting and note that setting.
• Restore any quality convars you write. These are client-global console state, not scene tuning. A game preset that writes them and does not restore them leaves the next performance measurement grading a configuration nobody pinned.

Post-processing in this engine is a per-camera opt-in stack, and the quality profiles are a global ceiling applied on top of whatever the scene requested. Once you understand both, the "correct settings, wrong picture" mystery resolves: the picture is flat because no post-process component is attached, and the shadows are coarser than authored because the player's quality tier capped them. Attaching the components makes the passes exist; measuring at a pinned quality tier makes the numbers portable.