I reached a similar conclusion. Dynamic memory is seductive for the task of indefinite scaling, but a practical system always encounters bottlenecks that aren't along the memory management axis, and in the meantime your code is much harder to verify.
NB: Historically, game engines have tended towards object pooling at runtime without any dynamic allocations. In that case there is a defined limit to what a scene will accommodate, and the object counts often simply reflect the other performance bottlenecks involved.
GC is still nice, but mostly in the sense of stitching together the most dynamic elements of the system. You don't want to have to trace a ton of stuff, and that also leads in the direction of flattening the data and making it more manual and static as the type system allows.
NB: Historically, game engines have tended towards object pooling at runtime without any dynamic allocations. In that case there is a defined limit to what a scene will accommodate, and the object counts often simply reflect the other performance bottlenecks involved.
GC is still nice, but mostly in the sense of stitching together the most dynamic elements of the system. You don't want to have to trace a ton of stuff, and that also leads in the direction of flattening the data and making it more manual and static as the type system allows.