Tx: https://mempool.space/tx/7393096d97bfee8660f4100ffd61874d62f9a65de9fb6acf740c4c386990ef73
Witness consists of:
1 signature:
809edb01f5931cc992763731cda9e983d7e2030a0863352530907490ef2a289721358c386d0b23d82fe78aab1e2f7f3bcf9ae7409bb771c98e7222dc136209f9
repeated 998 times.
followed by the script, consisting of:
20 (length), then pubkey: 1ea539fd851574f6802e6cc0cda3b2bd60afcfca9cd72d9279c5dc8c2054f6b6
then OP_CHECKSIG
then the same pubkey: 20 1ea539fd851574f6802e6cc0cda3b2bd60afcfca9cd72d9279c5dc8c2054f6b6 OP_CHECKSIGADD
repeated a further 998 times (so 999 times in total for that pubkey)
then the script finishes with:
02e603a2
which unpacks to:
998 (02: how many bytes, e603 is little endian, 03 x 256 + 230 = 998)
and OP_GREATERTHANOREQUAL
the remaining section of the witness is the control block:
c11dae61a4a8f841952be3a511502d4f56e889ffa0685aa0098773ea2d4309f624
(minor point: no actual merkle root or proof needed in case there is only one script; in BIP341 it says we need 33 + x where x can be 0 as the length of the control block; this is basically: one version byte and 32 bytes of the internal key P)
So in summary, the transaction has used the same public key 999 times and required a signature to exist for at least 998 of 'them'. The fact that the signatures are all the same looks weird, but it's a result of this artificial choice.
Why use GREATERTHANOREQUAL? The idea there is, it means we can use any of the 999 possible sets of 998 keys, and it satisfies the same single script.
The alternative is to build a merkle tree with 998 leaves, each of which allows satisfaction of the script by a different subset of 998. This would mean the control block would need a merkle proof of depth ~ 10 of course, which is going to be a few hundred bytes more.
The tradeoff there is interesting: if you use GREATERTHANOREQUAL you can minimize the size of the control block; if you use all combos as leaves, you can in principle hide 1 of the 999 keys, if it never signs. Obviously this would be more interesting if you had a threshold like 1 of 999, because then, using GREATERTHANOREQUAL instead of EQUAL would unnecessarily have to reveal all of the keys.
The main point stands, whatever the details: you can make very large multisig thresholds with taproot, even discounting Musig or FROST etc.