Regression and recipes

Regression record

Metric	Baseline	Candidate	Budget/result
peak process/platform memory
steady memory after traversal
texture/VT/Nanite pool pressure
ready latency / worst hitch ms
PSO misses/runtime compile hitches
visual quality/fallback result

Run repeated cold and warm samples. Preserve device profile, build, route, speed, camera, content, and cache preparation. Reject wins that shift work past the capture, lower required quality, or break cook.

Symptom recipes

Symptom	Bounded experiment
opening one menu loads a level's content	Reference Viewer from widget/class default; soften first boundary
textures stay blurry	build/inspect streaming data, then compare wanted versus budgeted mips
texture pool over budget	identify valid demand/forced/non-streaming owners before pool or mip change
VT tiles flicker	identify saturated format pool; A/B fixed pool size within VRAM budget
static Nanite view keeps streaming	inspect pool thrash; A/B pool or resident demand
repeated spawn/despawn grows RAM	marked Memory Insights retained-allocation query
first material use hitches	cold PSO validation; wait/precache or cover documented misses
teleport arrives into missing world	destination source plus complete readiness gate and timeout