jordanpaulee/local-agent-lab: RX 9070 XT ROCm vs Vulkan backend experiment (Ollama, qwen3:8b Q4_K_M)

A single-machine, method-stated experiment on an AMD Radeon RX 9070 XT (16GB, gfx1201) under Windows 11 driving stock Ollama 0.35.1 via /api/generate. Ships the harness (bench.mjs), the raw rows (data/) and the Ollama log excerpts, so the decode figures are checkable, not just asserted. Reports MEDIAN of 3 with [min-max] ranges. It is a community measurement by an individual researcher, not a vendor or independent-lab result - filed under measured_by 'third_party' because schemas/benchmark.schema.json has no 'community' member. Scope limit stated by the author: one model, one quant, one machine, n=3. Decode is an Ollama/eval_rate figure, not a llama.cpp llama-bench figure, so do NOT mix it with llama-bench rows in the same column.