Staring at the mempool for a while is pretty much like watching a subway station during peak hours: the same batch of transactions is lining up there, and you think you've already "sent it out," but in reality, you've just entered the waiting hall. During congestion, nodes first filter transactions by fee rate/priority; the packers (and a bunch of bots watching for front-running opportunities) will pull the more "valuable" ones to the front, and your transaction might stay stuck, or even be replaced by a later transaction with the same nonce. The more complex the routing, the more obvious this becomes: each additional hop increases the chance of slippage, changes the state, and ultimately results in a revert, leaving you with only a failed record and the pain of gas fees.

Recently, on testnets, the atmosphere of "earning points while waiting for mainnet token issuance" has caused the density of small, template-based transactions in the mempool to visibly thicken... Anyway, congestion isn't some mysterious phenomenon; it's basically queue rules plus someone cutting in line. My approach is simple and straightforward: avoid jumping queues when possible, don't set the fee rate too low, and if you're really stuck, increase the fee to replace the transaction in time. Otherwise, waiting for the state to change before re-entering is pretty pointless.