_q8blkdot_minrepro.nx

_q8blkdot_minrepro.nx -- the feature program for __q8blk_i16dot (search R0s-b, 2026-09-17): one Q8_0 block's dot, codes:*i8[32] . x:*i16[32] -> i64, the int8 codes sign-extended IN REGISTER (vpmovsxbw) so the block-native kernel never stores an i16 copy of the weights. nx_cc_equiv_gate builds this with the CHALLENGER only (a baseline that lacks the builtin cannot), runs it, and requires exit 0. Every arm is checked against an exact scalar reference read back from the same bytes: exit 1 = the 32-lane signed ramp disagrees with the scalar sum exit 2 = a planted +1 in one code lane does not move the dot by exactly that lane of x (a lane was skipped) exit 3 = a planted +1 in one x lane does not move the dot by exactly that lane's code (the other operand) exit 4 = the reference is zero (a vacuous fixture; the ramps are built so it cannot be) exit 5 = the widest legal block (every code 127, every x 32767) disagrees: 32 * 127 * 32767 = 133,153,088 exit 6 = the most negative code (-128) against the most negative x (-32768) disagrees: the sign extension is two's complement on both sides, 32 * 4,194,304 = 134,217,728 The contract the builtin holds the caller to: exactly 32 lanes, codes are int8 two's complement, x is i16. expect_exit: 0 license_tier: ORIGINAL