nx_placement.nx

nx_placement.nx -- multi-tier tensor placement + lazy materialization. L1.5 brick. The smallest demonstrable step toward the conductor cardinal (2026-05-19): tensors aren't FORCED into VRAM-equivalent storage at load time; they can live "where they already are" (mmap'd GGUF buffer) and only materialize per-use. Without this brick: nx_gguf_load_tensor allocates a full Q10 i64 buffer for every tensor at load. A real Llama-70B Q4_K file has ~40GB of weights; materializing all of them into i64 Q10 would require ~320GB of RAM. Infeasible on any consumer device. With this brick: tensors live as NxPlacedTensor with placement tag. Lazy (MMAP) tensors hold a (buf, offset, ggml_type) handle and only materialize a Q10 i64 NxTensor on first access -- callers that stream weights by layer can demote earlier layers, keeping working set bounded. This is Phase A of the multi-phase conductor arc: A) [this brick] Placement tag + lazy materialization B) Lazy GGUF loader (nx_gguf_load_tensor that returns lazy) C) Dispatcher (per-op hardware routing) D) Resource arbiter (multi-tier budget tracking) E) Async DMA primitives Bits-up composition: nx_tensor.nx -- NxTensor (the materialized form) nx_gguf.nx -- GGML type tags nx_gguf_load.nx -- dequant routing for materialization genealogy_id: numpy_memmap_pattern + mmap_lazy_load_canonical lineage_id: substrate_placement_v1_lazy_promote_demote