nx_osbench_micro.nx

nx_osbench_micro.nx -- OSBENCH: NishiOS native-contract vs Linux host primitives, measured in the SAME process on the SAME silicon through the nx_pabi seam (backends PABI_NISHIOS / PABI_LINUX). Rung F853 (lane osbench). Operator altitude ruling 2026-07-20: native-contract-on-same-metal -- NOT emulated-guest-vs-native-host, so these rows are honestly MEASURED, not DIFFERENT-ALTITUDE. ROLE: this organ MEASURES and asserts NOTHING. It emits tab rows to knowledge/status/osbench_measured.dat; nx_osbench_judge (F855) computes every verdict from the numbers. No verdict is decided here. LIAR-KILL / anti-stale-binary: every run records the measured artifact's own byte-length and a wrapping polynomial digest (artifact identity). This is a CHANGE DETECTOR, not a cryptographic hash -- named honestly as fnv-class 'adig', not sha256. It exists because debts seq207/124/236/242 are one family: you cannot otherwise prove the binary you measured is the binary you built, and a speed verdict on byte-identical binaries measures nothing (the asm-diff-guard lesson). FAIRNESS: both backends CREATE-NEW every iteration (names carry the round index), so neither side gets a cheaper overwrite path. The NishiOS block-VFS is additive-only (vb_create REFUSES overwrite) and gets a FRESH 256KB image per round; VB_BS*VB_NB = 1024 blocks bounds N. DECLARED ENVELOPE (printed in output -- silent capping is forbidden): N=200 ops/round, ROUNDS=3, payload 64B, single-threaded, single-node, warm cache. COVERS: create+write, read, resolve. DOES NOT COVER: context-switch, IPC, thread-spawn, syscall-trap cost -- those need NishiOS running AS an OS and are gated on the bare-metal node (rung F856). They are emitted as UNMEASURED rows, never as zeros. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0