_k_r2_001b2a_gate.nx

_k_r2_001b2a_gate.nx -- the K-R2-001b2a gate (virtio-MMIO DESCRIPTOR-DMA layer; the second slice of the K-R2-001b virtqueue epic, on top of b1's queue-config). Drives the full author->handshake->queue-config->descriptor-DMA chain with NO mocks: runs the REAL nx_virtio_hs_emit (the team AUTHORS the rv64 image + the TABLE-COMPUTED golden transcript from the descriptor-DMA spec), then RUNS the image on the SOVEREIGN rv64 emulator (rv64im_min_sim + the rv64im_min_virtio device now DMA-reading the descriptor on the QueueNotify kick and exposing the read field at the QueueDescPeek RO register -- the PRIMARY, gating lane: Nishi owns the runtime) and asserts the captured serial transcript CONTAINS the emitter's golden ("VIO ACK DRV FEAT OK VQ DESC\n" -- after the queue config the driver laid ONE legacy virtio descriptor into the QueuePFN-bound guest ring page, kicked QueueNotify so the device DMA-read it, then READ QueueDescPeek BACK and verified it == the descriptor addr-low word it wrote: the descriptor-DMA binding proof) AND the sovereign emu reports a clean SiFive-finisher halt. Then the ALIGNMENT lane: qemu-system-riscv64 -machine virt -global virtio-mmio.force-legacy=true with a REAL virtio-blk-device backing. The gate authors a SECOND image from a base-rewritten spec (the blk device lands at slot 0x10008000) and runs the SAME driver against the REAL legacy virtio-blk transport: it completes the handshake + queue config, lays the SAME descriptor into qemu guest RAM at the ring page, and kicks the real QueueNotify -- the real legacy transport genuinely walks the ring. The QueueDescPeek result register is a sovereign-model instrument the real device does not expose (it reads 0), so on qemu the descriptor-DMA read-back verify gracefully drops ONLY " DESC"; the gate asserts the qemu transcript contains the b1 prefix golden ("VIO ACK DRV FEAT OK VQ") -- proving the SAME driver lays the descriptor + kicks the real device -> lanes AGREE on every register the real transport can observe. Finally a TAMPER test: corrupt the QueueDescPeek-EXPECTED constant in the sovereign image (the driver loads its own QueueDescPeek read-back into t3 via lwu and compares against this li-loaded expected in t4; bumping the expected's immediate makes t3 != t4) -> the read-back verify branches PAST stage 6 straight to the finisher -> the transcript loses ONLY its " DESC" canary (the "VIO ACK DRV FEAT OK VQ" tail survives) -> the gate MUST go RED. Evidence -> knowledge/status/virtio_blk.log (DESCGATE row; the queue row's ||MARK= reads it). Sovereign orchestration (fork/dup3/execve/wait4). license_tier: ORIGINAL