P142 gives frontend-consuming prefetch fills the same two-word repair budget P141 gave demand-fetch fills. Auxiliary/background prefetch fills stay at one word.
The shell workload passes and reaches P142-FILE-OK.
The result is better than expected: the shell window drops to 63,926,691 cycles. That beats P141 by 362,576 cycles and beats P134 by 294,951 cycles.
This is not free. Background repair fills rise from P141’s 32,151,196 to 42,853,866, and the usefulness ratio falls. But absolute first/repeat fetch hits rise enough to reduce I-cache misses and win the workload.
Next: split the broad prefetch bucket. The policy handed out 48.85M prefetch second-word grants, so there is likely room to keep the speedup while trimming waste.