nx_drv_proto_emit.nx

nx_drv_proto_emit.nx -- DRIVER-PROTOCOL-FROM-SPEC (X-DRV-W1), the GENERIC driver emitter. The thesis (driver-as-shape-composition): a driver SPEC carries the device register map (base + offsets) AND -- crucially -- the PROTOCOL STEP-SEQUENCE itself as DATA: an op-list (the STATE_MACHINE shape). This organ WALKS that op-list and SYNTHESIZES a bare-metal rv64 driver image whose control flow IS the spec's sequence -- so a DIFFERENT protocol sequence is a DIFFERENT spec -> a DIFFERENT driver, with ZERO emitter changes. That is the emitter-of- emitters keystone (X-AUT-006c/e) COMPOSED into the device-driver layer (X-DRV-W0 HWMAP gives the bind side; this gives the protocol side). CONTRAST with nx_virtio_hs_emit: that emitter parameterises the virtio register offsets/values from a spec but HARD-CODES the protocol sequence (identity->ACK->FEAT->VQ->DESC->USED->...). Here the sequence is DATA, so virtio-blk is merely ONE spec instance (the last-mile/benchmark transport per the nishi-ecosystem-only law); a virtio-net handshake, or a future Nishi-native device protocol, is simply a different op-list. The gate proves this by emitting TWO distinct op-lists from the ONE binary and running BOTH against the real devices on the sovereign emu. SPEC FORMAT (line-based, '#'=comment): out <path> where to write the flat image (and <path>.gold) base <hex> device MMIO base (0x1000-aligned -> lui t2) op verify <regoff> <const> LW reg ; li const ; bne -> fail (signed read-back) op verifyu <regoff> <const> LWU reg ; li const ; bne -> fail (unsigned/address read-back) op write <regoff> <val> li val ; SW reg (also: notify = a write) op notify <regoff> <val> alias for write (device kick) op read <regoff> LW reg (exercise a RO register, value discarded) op checkbit <regoff> <mask> LW reg ; li mask ; and ; beq x0 -> fail (bit-stuck proof) op setbase <addr> li32u t5 = addr (membase for the following memstore ops) op memstore <memoff> <f3> <val> li val ; S[f3] val, memoff(t5) (f3: 1=SH 2=SW; ring/desc lay) op emit <literal-to-eol> write the literal bytes over the UART (the transcript token) The image always ends with the SiFive finisher (clean halt). Any verify/checkbit that fails BRANCHES PAST the rest of the op-list to the finisher, so the transcript loses its tail -- that is the tamper handle the gate exploits. nx_drv_proto_emit <specpath> -> the spec's `out` (flat rv64 image) + <out>.gold (golden transcript = the concatenation of the emit literals, in op order). VERDICT log -> knowledge/status/driver_spec.log. Sovereign: syscalls only, no gcc/.sh. license_tier: ORIGINAL