--user-data-dir separates almost everything, and not the network

Everything a page can write and read back later, in one tree on disk.

Two profiles signed into the same site do not see each other's session. That is the whole job, and it is done.

Chromium keeps a body of state that is keyed by site rather than by profile directory, and it is not partitioned by the switch. The clearest examples are the ones that make the browser fast: how a host was reached last time, which protocol it negotiated, which alternative services it advertised, what the DNS layer already knows.

None of that is a cookie, none of it survives in a form a page can read directly, and none of it identifies a user by itself. But it is shared across the profiles running on one machine, and there is no flag that separates it. This is a property of where the state lives in the browser, not a configuration anybody has failed to switch on.

The practical consequence is narrow but real: two profiles on one machine are not network-isolated from each other in the way they are storage-isolated. If your threat model includes an adversary correlating connection-level behaviour rather than page level state, the profile directory is not the boundary you were hoping for.

Because the alternative is letting people believe that separate directories mean separate machines, and then discovering the gap at the point where it costs something.

There is a stronger version of this argument that is worth making explicitly. Most of what is sold in this category is described in terms of what is separated. Almost none of it is described in terms of what is not. That asymmetry is the whole marketing strategy of the category, and it works because the unseparated things are exactly the ones a customer cannot check.

The three that matter, in the order they will affect you:

1. The network stack is shared. Described above. Not fixable with a flag.
2. The graphics stack is shared. There is one card in the machine and every profile renders on it. A renderer string can be reported per profile, and a readback can be perturbed per profile on the way out, but neither of those is a second graphics stack, and the distinction between changing a label, shifting a result and doing the work somewhere else is the one to hold on to.
3. The font set is shared. The fonts installed on Windows are the fonts every profile sees. Virtualising them per profile is not something any tool that launches an unmodified browser does.

The right response is not despair, and it is not a different tool. It is knowing which boundary you have.

For keeping accounts from being linked through stored state - cookies, tokens, storage, cache, the things sites actually use to recognise a returning visitor - a profile directory per account is complete and correct. That is what it is for and it does it without qualification.

For everything below that layer, the boundary that actually separates two identities is a different machine, or a virtual machine, or hardware you do not share. That is a larger commitment and it is sometimes the right one. A profile manager is what you use when it is not.

The reason to write this down is that the distinction is invisible from inside any product in this category, including this one. You cannot check it by looking at the screen. You can only be told, and most people in this market will not tell you.