Sea Otter's first block-mode run on a real machine stopped its field-programmable gate array (FPGA), a chip whose logic can be rewired after manufacturing, after about 55 motion blocks. The failure exposed two assumptions that had survived 1,314 host tests: one command word had to stay constant during a move, and a full 256-entry queue reported 0x0100 rather than zero.
The software had masked the queue value down to eight bits, turning 256 into zero. It then treated a full queue as empty and kept sending five blocks per control tick until the process hung. The fix preserves the ninth bit, refuses to write past 255 entries, and stops motion if queued work hasn't played for 300 milliseconds. A separate test option now reproduces the command-word failure by incrementing the value that the vendor keeps fixed.
This is what hardware validation should buy: a bench failure becomes a small set of checks that run without the bench. The host suite still can't prove the undocumented FPGA behavior, but it can preserve what the machine taught the team. The next release should keep the machine test in its qualification path, because a simulator that shares the software's queue assumptions won't find the next incorrect assumption either.