Tess

The tank shape explanation makes the fast last quarter make sense too, which I hadn't connected.

The thing that would worry me is a debug LED sitting on one stage rather than cycling. If your board has the four little lights, watch them: cycling through CPU, DRAM, VGA, BOOT and then going out is healthy. Parking on DRAM for two minutes is not.

Worth doing, but photograph it before you touch anything. Bent socket contacts are usually not covered, and a board vendor that sees fresh scratch marks from tweezers will say so.

For the record: run one full pass of memtest86 per stick, per slot, so you have four data points instead of two. Takes an evening but it's the difference between guessing and knowing. Also check Task Manager reports 32GB installed rather than 32GB usable, because a hardware reserve issue looks completely different.

Move it to the second x16 slot for a week even if it drops to x4. It costs you a few percent in games and it definitively answers whether it's the slot or the card.

Also worth knowing which stress test you ran. A small FFT prime run or a heavy AVX benchmark is a much harder load than anything you'll do in real life, and treating that number as your normal operating temperature will just make you miserable. Render for ten minutes and look at that instead.

Also stop using the drill on the way out of stubborn screws. First movement should be by hand with a proper fitting screwdriver and your whole weight over it. Drills spin before they turn, and that's exactly how a head rounds.

The two deployables point is the one that lands. Last internal tool I built this way I spent more days on deployment than on features.