2. Execute-stage redirects take a configurable latency

Context. A branch misprediction or other execute-time redirect used to change the fetch PC in the cycle it executed, and fetch ran at the new PC in the same tick. gem5's O3 takes iewToCommitDelay + commitToFetchDelay (2 cycles) to get a squash from execute to fetch; MinorCPU takes 1.

Decision. Execute returns a Redirect (core/pipeline/squash.rs). The engine files a PendingSquash for it at the result's completion cycle, and takes it pipeline.redirect_latency cycles later: default 2 for O3 and 1 for in-order. Until then commit retires nothing the squash will remove. O3 keeps issuing wrong-path work meanwhile, as gem5's IEW does; in-order issues nothing the squash will remove, as Minor drops those instructions at once. Predictor repair happens when the squash is taken. Memory-order and coherence violations squash through the same path.

Consequences. The misprediction penalty includes the time the squash takes to reach fetch. tests/unit/core/pipeline/timing.rs pins the exact redirect cost on both backends.