[Tarzelf]

The PiPER simulation repeatedly allocated a force array and contact membership sets during each contact scan. This experimental patch reuses scratch storage within a scan while retaining every fresh force query and all 1 ms mechanics, safety and outcome checks.

Four fixed ABBA action-tape runs measured 15.70% lower pooled mean environment-step wall compute and 12.60% lower process CPU compute on the shared host. Ten separate diagnostics exactly match 200,010 recorded native snapshots, contact evidence and independently recomputed outcome state. The original 64 MiB reader refusal and supplemental streaming audit are both retained; no physics or timing reruns were used.

Validation: 11 native PiPER tests, six feedback tests, three contact freshness/threshold tests, four profile tests, ten comparison tests, packaging and guarded publication checks. This is one synthetic scenario and two timing pairs; no new task accuracy, physical qualification or real-time guarantee. No new wheel, stable-main change or production-runtime promotion is included. The prior learned held-out result remains 5/12.