docs: define upstream-first fork policy

Author: sanchitmonga22

RUNANYWHERE.md

@@ -0,0 +1,31 @@
+# RunAnywhere llama.cpp maintenance notes
+
+This is a direct fork of `ggml-org/llama.cpp`. The RunAnywhere integration
+branch carries the DeepGrove Maple model architecture and ternary expert graph
+from `deepgrove-ai/llama.cpp`; it does not wholesale-track that fork.
+
+## Backend status
+
+- CPU: Maple Preview, PrismML Bonsai `Q1_0`, and canonical `Q2_0` are supported.
+- Metal: Bonsai `Q1_0` and canonical `Q2_0` are supported.
+- Maple must use CPU execution. Its expert path requires a ternary
+  `mul_mm_id` Metal pipeline that the current DeepGrove patch does not provide.
+- Prism's older `Q2_0` artifacts used a non-canonical 128-value block under the
+  same GGUF type ID. Do not change the global type to accommodate them. Use the
+  newer `Q2_0_g64` artifacts Prism publishes for canonical llama.cpp instead;
+  this preserves CUDA, Vulkan, Metal, SYCL, WebGPU, and conversion compatibility.
+
+## Rebase and release checklist
+
+1. Rebase the integration branch onto a pinned canonical llama.cpp revision.
+2. Re-apply the two small Maple commits and resolve only their architecture
+   registration/model graph surface.
+3. Build `llama-cli` on macOS with Metal and on Linux with the CPU backend.
+4. Run real inference with the pinned Maple Preview, Bonsai `Q1_0`, and Prism
+   `Q2_0_g64` GGUF fixtures recorded in the RunAnywhere SDK runtime plan.
+5. Keep Maple CPU-only until a real GPU expert pipeline passes the same model
+   smoke test.
+6. Tag the exact tested commit and pin the SDK to that immutable tag or SHA.
+
+The upstream remote for maintenance should always be
+`https://github.com/ggml-org/llama.cpp.git`.