Symptom recipes

Symptom	Prove first	First bounded experiment
slow only at high resolution	pixel-scaled GPU passes	lower screen percentage; compare pass deltas
slow in dense city/forest	Render thread, Nanite/material/VSM/visibility	isolate primitive/section, overdraw, caster invalidation
shadow spike while foliage moves	VSM uncached pages/casters	disable distant WPO/shadows on one foliage class
camera cuts hitch	visibility/VSM/streaming/upscaler history	marked trace plus cache/streaming views
many lights stay fast but noisy	MegaLights sample competition	reduce competing radii/lights or raise samples in controlled A/B
reflections dominate	Lumen reflection pass	compare reflection scalability separately from GI
translucent combat effect tanks GPU	translucency/overdraw event	cut overlap/coverage for one effect
render thread slow, GPU has headroom	primitive/section/submission count	instance/cull one high-count category
GPU memory grows by location	resource snapshot and streaming residency	identify largest new owners and correct residency/LOD
low tier changes nothing	device profile/group not applied or restart needed	display active CVars/profile and verify packaged build
dynamic resolution bottoms out	sustained GPU cost exceeds minimum-resolution budget or CPU bound	inspect GPU at minimum percentage; optimize dominant pass

For every recipe, restore the toggle, implement the production-quality change through owned settings, and capture the same worst case again.