nx_conv2d_mt_gate.nx

nx_conv2d_mt_gate.nx -- adversarial gate for the multi-threaded conv2d forward (nx_conv2d_forward_mt / nx_conv2d_forward_pool) against the serial oracle nx_conv2d_forward. The MT path is an INDEPENDENT implementation of the same math (organ policy: the serial path was left untouched), so full-buffer bit-exact comparison here is a real two-implementation oracle, not a self-comparison. Checks (10): 1 serial oracle runs OK on the small shape 2..6 MT bit-exact vs serial for nworkers in {1, 2, 3, 0=auto, 1000=clamped-past-row-count}; output poisoned before each run so unwritten rows cannot pass 7 no-bias variant bit-exact (nworkers=3) 8..9 caller-owned pool reused across two calls (second call has different input contents) -- proves the completed-counter DELTA wait, both bit-exact 10 speedup on a big shape (1x64x128x128, C_out=64), auto workers: bit-exact AND wall-time speedup >= floor Small shape uses prime-ish dims (H=17, W=13, C_out=5) so any stride or band-boundary bug lands off-pattern and fails the compare. genealogy_id: substrate_conv2d_v1_direct + nx_conv_speedup_banding lineage_id: conv2d_mt_gate_v1