A simple explanation is that the nature of a cellular network requires that all base stations have a relatively large amount of shared state (which can get as fine as TDMA synchronization) and the ability to exchange information with each other in realtime. WiFi generally doesn't have this requirement, especially now that roaming extensions have mostly obsoleted AP virtualization (which is what pricey enterprise WiFi systems with a dedicated controller used to add).

An even simpler explanation is that the architecture of the cellular network is both old and comes out of the telecom industry, which both mean that it has many layers and strict requirements for QoS, traffic engineering, etc.

In a little more detail (given that this is not a topic I'm an expert in and it can get confusing): most of the magic in 5G happens in a component called the RAN or radio access network. The RAN is basically everything between your phone and the "core network" that provides actual services like telephony and IP access (which is going to be in a data center). The RAN can be fairly complex as in newer technologies it involves things like making intelligent heuristic decisions about which base station a given device should be communicating with.

Historically base stations were expected to be largely independent and handle basically everything between the phone and the existing ISDN network, which just sort of dated back to how analog radiotelephones had worked. This required a lot of equipment to provide the entire RAN on-site. The new direction has been to absolutely minimize what is located in the field, both for size and power savings but also to simplify management since there's less field equipment to upgrade and maintain. This means that a typical 5G gNB, the actual radio station, basically does nothing but encode/decode to/from binary, which it then sends to a "virtual RAN" or vRAN running in a data center somewhere. All of the actual protocol implementation, access control, traffic engineering, etc. is done in the vRAN. This adds a lot of flexibility since the vRAN can be maintained and iterated on more easily and can flexibly allocate resources between sites. It also simplifies field installation because the site only needs connectivity back to the vRAN, which is a little simpler to arrange (via VPN, fiber, metro ethernet, ISDN, whatever) than getting the site connectivity into an actual mobile exchange, and to IP capabilities, etc.

Or in other words, your WiFi AP does contain the business logic of IP switching and 802.11 session management. But that's relatively simple and done for a relatively small number of clients compared to a cellular base station, and WiFi has (mostly) always been designed with the idea of minimal to zero requirements for inter-AP communication beyond existing IP capabilities.

Not only is there some additional network logic to help handoffs go smoothly, there's a lot more complicated business logic going on with cellular APs than home WiFi APs. On top of that, you're going to want to try and control potentially thousands of APs and have them orchestrate and you may not want to rely on RF links to propagate changes.

A home WiFi AP is doing a pretty straightforward relation of SSID to some Ethernet-like network. Anything talking on the Ethernet network addressing a client goes out the radio, things coming from the client to the Ethernet network goes in. Even with this, it can make sense to have the management of these systems centralized, most enterprise systems do this already.

Cellular networking gets a lot more complicated with various APNs and other advanced networking concepts. There is a lot more going on than just doing a handshake, pulling an IP address, and starting to talk networking.

My understanding of 5G and modern cell phone protocols is there isn't really an idea of "handoff". With CDMA multiple towers in an area all get the same signals from your phone and are all transmitting on the same frequencies as well. It is kind of a "mesh" like thing.

it's sort of like asking whether for self driving cars whether it is prudent to have cars that strictly know to act based on what they sense (in a reactive way), or if it is better to collect maps of sensed environments and figure out how to proactively figure out the best optimal approaches in given situations conditioned against various environments and push out updates.

the answer is that _both_ are prudent. There is enough cooperation and coordination challenges that both offline and online approaches are important.