_k_r2_001c3_gate.nx

_k_r2_001c3_gate.nx -- the K-R2-001c3 gate (virtio-NET FRAME LOOPBACK; the THIRD and final slice of the K-R2-001c net virtqueue epic). It is the K-R2-001b3 sector-data round-trip gate applied to the SECOND virtio device, virtio-net (DeviceID=1) @ base 0x10002000. NO mocks: runs the REAL nx_virtio_hs_emit (the SAME shared emitter the blk b-series + the net c1/c2 use -- data-only on base/device-id/banner/queue-config/descriptor/avail-used/frame, so it stays byte-for-byte backward-compatible), then RUNS the emitted image on the SOVEREIGN rv64 emulator (rv64im_min_sim + the second rv64im_min_virtio net instance @ 0x10002000 -- the PRIMARY, gating lane: Nishi owns the runtime). After the c2 queue-config the net driver lays a frame (a virtio-net WIRE_TLV header word: flags/gso_type/hdr_len packed into the first 32-bit word) + payload into a tx descriptor's data buffer in guest RAM, publishes the descriptor on the AVAIL ring, kicks QueueNotify -> the net QueueNotify now triggers nx_virtio_notify_dma on s.virtio_net (the device-instance-generic DMA walk: desc-read + avail/used + the data round-trip), so the device FOLLOWS desc[head].addr into guest RAM, DMA-READS the frame's first word, and latches it into QueueSectPeek. The driver reads QueueSectPeek BACK (lwu) and verifies it == the sent frame word -> emits the trailing " FRAME" canary. The gate asserts the captured serial transcript CONTAINS the emitter's golden ("VNET ACK DRV FEAT OK VQ DESC USED STAT FRAME\n") AND the sovereign emu reports a clean SiFive-finisher halt. The blk device @ 0x10001000 is left BYTE-UNTOUCHED (its own gates VIRTGATE..SECTGATE still pass, re-proven by re-run). Then the ALIGNMENT lane: qemu-system-riscv64 -machine virt -global virtio-mmio.force-legacy=true with a REAL virtio-blk-device (first) + a REAL virtio-net-device (second). qemu fills its virtio-mmio slots in REVERSE, so the first device (blk) lands at the top slot 0x10008000 and the SECOND device (net) lands at the next slot DOWN, 0x10007000. The gate authors a SECOND net image from a base-rewritten spec (base = the qemu net reverse-slot 0x10007000) and runs THAT on qemu; the SAME loopback driver lays the SAME tx descriptor + the SAME avail ring + the SAME frame word into qemu guest RAM and kicks the real QueueNotify (the real transport walks the ring). The sovereign-only QueueDescPeek/ QueueUsedIdxPeek/QueueStatPeek/QueueSectPeek RO peek registers (0x060/0x064/0x068/0x06C) are a SIM modelling device -- they read 0 on the real qemu virtio-net transport -- so the descriptor- DMA + used + status + FRAME read-back verifies drop " DESC"/" USED"/" STAT"/" FRAME" there; the gate asserts the PREFIX golden ("VNET ACK DRV FEAT OK VQ") appears, proving the driver runs IDENTICALLY against the REAL virtio-net transport through the handshake + queue-config + tx- descriptor lay + avail publish + QueueNotify kick (lanes AGREE on the real-hardware-driveable behaviour; the frame round-trip proof itself is the sovereign PRIMARY lane, the same alignment semantics the blk b3 SECTGATE established). Finally a TAMPER test: corrupt the FRAME-EXPECTED constant in the sovereign image (the driver loads its own QueueSectPeek frame read-back into t3 via `lwu t3,0x06C(t2)` and compares against the li-loaded expected in t4; the expected li is the lui/addi following the unique SectPeek read-back load). Bumping the addi's high immediate byte makes the read-back verify branch PAST