OP_ASSIGNMENTS.md

GGML op-enum assignments

Original GGML ops have no explicit integer assignments. The mainline ggml/include/ggml.h enum ggml_op enumerates positionally (no explicit values) and ends in GGML_OP_COUNT. New ops append immediately before GGML_OP_COUNT.

Per Phase 0.5 constraint C1.3, added ops are listed here so the intent and ordering are explicit and audit-able. The "slot" column is the position-index in enum ggml_op after each op is inserted (counting from 0). Mainline-side reordering will shift slot numbers — the contract is positional, not numeric.

Op	First landed	Phase	Purpose
GGML_OP_TURBO_WHT	feature/phase-1-turboquant-kv Step 3	Phase 1	TurboQuant Walsh-Hadamard Transform: O(d log d) rotation for KV-cache compression. Replaces dense ggml_mul_mat(128x128, ...). Sourced from TheTom 012faec26 (introducing commit).

Update protocol

When adding an original GGML_OP_*:

1. Append the op to enum ggml_op immediately before GGML_OP_COUNT in ggml/include/ggml.h.
2. Append a matching entry to GGML_OP_NAME[] and GGML_OP_SYMBOL[] in ggml/src/ggml.c (and bump both static_assert(GGML_OP_COUNT == N, ...) counters).
3. Wire CPU dispatch in ggml/src/ggml-cpu/ggml-cpu.c ggml_compute_forward and ggml_get_n_tasks and ggml_graph_plan.
4. Add CPU impl in ggml/src/ggml-cpu/ops.cpp + declaration in ops.h.
5. Add backend supports_op gates: ROCm (ggml-cuda.cu), Vulkan (ggml-vulkan.cpp), CPU (auto via dispatch).
6. Append a row to this table with the first-landed branch/phase and a one-sentence purpose.