Okay, I've got myself out of immediate failure I think (fingers crossed).
I could not stop the array (/mnt/cache would not unmount despite nothing showing using it in lsof) so I disabled auto start and rebooted the server. I then managed to remove the faulty nvme drive from the cache pool, and then (using the small checkbox confirming I really wanted to do it) started the array again.
I'm now getting a number of 'reallocating block group' log messages in syslog and the cache files are accessable (I "think" this is btrfs doing a balance itself...). I was also able to start docker again. I'm now going to leave the cache to reallocate whatever else it may want while I go check the dates and warranties on the failed nvme (assuming it's the drive and not an issue with the motherboard slot).
I'd still be interested if anyone can help explain if I did the right thing or just took a risky option that fluked ending well.
A few forum posts I had read that I took largely some guesswork from (and I say guess as I still have no real understanding of balance or scrub, and only basic guesswork on the actual error I was having...)
hoppers99's post in [SOLVED (I think)] NVME drive failing in mirrored cache pool. BTRFS errors. was marked as the answer