An additional point to make is about assumptions.

Consider that blockchain analysis pretty much started with what Meiklejohn et. al. called "Heuristic 1", which I guess should really be attributed to Satoshi:

"All inputs have the same owner."

The "first cut" of blockchain analysis is just that: clustering ownership based on that assumption. That's mostly what walletexplorer.com was/is all about, and although I don't pretend to know what other additional analysis it does, that's clearly the backbone of it.

Now, coinjoin breaks it of course, and for this reason I always say to people: "I use sendpayment for retail payments specifically to break automated clustering analysis". People who have got as far as the kind of thinking in this thread may counter: "well, but that's nearly useless since the Taker/Maker role behaviour distinction makes it almost always easy to isolate your output".

But, again, assumptions: if you're going to have automated, large scale wallet analysis, are you comfortable to start putting in a whole raft of assumptions about:

1. what is and is not a joinmarket transaction
2. that a person taking Taker role in TX1 is not taking maker role in TX2
3. the right way to handle cases where certain outputs have not yet been spent

etc etc. It is messy, and it gets messier the more different ways of operation exist; other coinjoin implementations than joinmarket, by-chance false positives, deliberate false positives, mixed role bots etc etc.

Something that Greg Maxwell has always been at pains to point out during discussions: there is no fixed "fact" about a coinjoin pattern, once heuristic 1 is dropped; even what looks like it couldn't possibly be a coinjoin, e.g. : inputs(1,3), outputs(3,1) could actually be a payment of 2 bitcoins from party A to party B, or two payments of 3btc A->C, 1btc B->D. Neither of these are likely today, but ... are they? How do we know?

Joinmarket's motivating philosophy is a bit "worse is better" (see gwern's article on bitcoin is worse is better) - that usage is what counts, because the set of near-watertight assumptions splinters and, in the most optimistic future, eventually collapses.