nx_v128_probe.nx

nx_v128_probe.nx -- F618 DIFFERENTIAL PROBE, POST-FLIP FORM (2026-07-21; no main; every fn exported by the wat backend, imported natively by nx_v128_gate). Since the FLIP, the SHIPPING vm_* kernels (nx_vmotion.nx) route T==16 rows through v128_sad16 -- scalar natively, INTERCEPTED to wasm-SIMD by the shared emitter. The scalar REFERENCE therefore lives HERE as pvr_* copies of the pre-flip vm_* bodies (pure per-pixel loops, no v128_sad16 call, so the intercept can never touch them). pv_sweep_scalar / pv_time_scalar -> pvr_* (reference, scalar in EVERY lane) pv_sweep_simd / pv_time_simd -> vm_* (shipping path: scalar native, SIMD in wasm) The sweeps compute an ORDER-SENSITIVE rolling checksum over best-SAD + MV per macroblock (full center-first search, cap-aborts live) plus an UNALIGNED raw-SAD pass and a capped-SAD pass. scalar-chk == simd-chk == native-chk is the bit-exactness claim, mechanically. i64 multiply wraps mod 2^64 identically in native x86, the sovereign VM, and V8 -- the checksum is portable-exact. NOTE (inherited, both arms identically): sys_mmap is a 0-stub in the wasm lane, so the searchers' 32-byte b scratch sits at linear address 0 (the shipped codec's own behavior; sequential calls re-init it). Frames must live well above -- callers use base >= 0x20000. license_tier: ORIGINAL