P141 makes P140’s repair budget adaptive. Demand-fetch fills get two repair words because the line is proven to be on the instruction path. Prefetch-only fills stay at one word.
The shell workload passes and reaches P141-FILE-OK.
This is a real policy win. The shell window drops from P140’s 65,035,481 cycles to 64,289,267 cycles. That beats P140 by 746,214 cycles and P138 by 348,494 cycles. It is only 67,625 cycles slower than P134.
Repair traffic stays near P140: 32,151,196 repair fills versus 32,047,006. The extra 2,739,556 second-word grants improve first-use and repeat-use fetch hits enough to matter.
Next: give the second word to the right prefetch lines too, but keep the full-line repair stream off.