@JDługosz, what you wrote, "both layer’s headers use mac addresses" is incorrect. MAC addresses are layer-2 addresses used by some layer-2 protocols, including ethernet. Some layer-2 protocols use 48-bit MAC addresses, some use 64-bit MAC addresses, and some do something else altogether. Layer-3 doesn't use MAC addresses. IP is one layer-3 protocol, and it uses IP addresses, not MAC addresses.

The layers are independent of each other, and IP can be used with any number of layer-2 protocols, including those which do not use MAC addresses. A MAC address is only significant (and seen) on the layer-2 LAN where the host with that MAC address is.

OK, one thing I had remembered wrong is that each host has an ARP and builds the L2 header for local network delivery. So, a desktop machine already knew the MAC of the gateway and didn't care that another device had a duplicate IP address.

That article doesn't address (pun intended) how arp is cached and when it knows something changed.

It also doesn't address how wifi necessarily acts as a router vs just a switch. Since the hosts use local addresses I'm supposing it's just like a switch but converts transport medias.

Now a tablet on wifi expects to get different networks all the time, so connecting to a new hotspot (even if it has the same ssid as one it used before) is actually a different device, so everything is flushed. The Asus tablet used arp to discover the mac of the gateway, and got an answer from the wifi hotspot's router, which is the interloper.

But that doesn't explain why another tablet was different. The gateway routing worked fine. So did this one use a cached arp even though the hotspot is different hardware? Doing http to .1, it initially got the one in the wifi router, so its arp must have gotten that! But then after a few minutes it started getting the gateway instead, so why did the arp entry change?

One difference, I speculate, may be that newer devices use ipv6 for addressing the gateway? But ipconfig (on win10) doesn't show anything other than ipv4 address for the default gateway and subnet mask.

Now the Gracepoint music player, in isolation, could be explained as a race to answer the arp request, given that the switch doesn't remember that but the reply has to come all the way from the device.

@JDługosz, actually, a Wi-Fi access point is not a router or a switch. It is a bridge, akin to a switch, but where a switch is typically a transparent bridge, a WAP is a translating bridge. Ethernet and Wi-Fi are two completely separate layer-1/2 protocols, and the WAP must translate between them. A router operates at layer-3, routing layer-3 packets, after the layer-2 frames are stripped off the packets.

A host will use ARP to resolve the layer-3 address to a layer-2 address when it is not in its ARP cache. Afterwards, it will put it in its ARP cache. It will remain there until it times out, or the host sees other traffic that doesn't conform to what is in its ARP cache. This is dependent on the OS version.

Unlike a host, a transparent switch will have an ARP table that relates a MAC address to the port on which it entered the switch. It has nothing to do with IP addresses, it only tells the switch to which port it should send traffic destined to that MAC address.

A WAP, having only two ports (wired and wireless), doesn't need to maintain an ARP table the way a transparent switch does, but some do.

A gateway, usually a router, is just another host on the network, so it has MAC addresses for interfaces which use a protocol with MAC addresses. A gateway will also use ARP and an ARP cache the way a host does: to resolve a layer-3 address to a layer-2 address.