Fair, and I will give the probability away: a backdoor baked in at the fab is not the default threat out there, and most people will meet a wrench long before a management engine. But your own sentence hands the objection back. If the attack has to hit the offline and the online device together, then in this setup it is not two attacks. Your laptop is offline over the cable only; the board is the same, the same engine sits under the same OS. An air gap takes away the mouth and leaves the eyes.
Which is where "just check the destination address before broadcasting" needs a close look. Everything depends on what it gets checked against. Read the confirmation on the same machine that assembled the transaction and you compared the machine against itself. In a QR loop the second opinion is your phone, and your phone reads whatever the first machine decided to encode. The instinct is right, and a second witness has to be a second thing.
A padded fee is the other open door. Measurement from my own node a few minutes ago: 25,934 transactions in the mempool, a median fee of 26 sats. Ninetieth percentile 40, mean 0.2 sat/vB. Above 2,000 sats sit 29 of them, which is 0.11 percent. Four cross 10,000. An invoice pins an address and an amount; nothing pins the miner's cut. Today such a fee would be an outlier a check could catch. That is the good news, and it is a property of an empty mempool. Once fees run into the hundreds of thousands of sats, the same shave is rounding noise and the check stops meaning anything.
And yes, the wrench comes first, because it needs no firmware, only an address and a door. It also needs a witness. A bad batch needs none of the three. I do not rate it likely. I rate it the only one nobody has to be present for.
