feat: GPU bakeoff — 3090 Ti vs V100 vs Strix Halo

Cross-host Gemma 4 throughput comparison across three architectures.
Harness at scripts/gpu-bakeoff/; writeup at
docs/reference/gpu-bakeoff-2026-04-20.md.

Key findings:
- RTX 3090 Ti wins decode decisively (128 tok/s on gemma4:26b MoE Q4,
  ~4.7× faster than gemma4:31b dense on the same card).
- AMD Strix Halo iGPU lands at ~42% of 3090 Ti decode on ~25% of the
  memory bandwidth — good SIMD utilization, especially for MoE.
- V100 numbers are DEGRADED: CT 167 ai-visualizer SDXL consumes 31/32
  GB of its VRAM, forcing Gemma 4 models 95% onto CPU. Isolated V100
  run requires SDXL eviction — left as follow-up.
- MoE vs dense is the dominant latency factor across all GPUs: ~4 B
  active params of gemma4:26b beats 31.3 B active of gemma4:31b by
  the same ratio (~4.7×) on every card tested.

Methodology: 1 warmup + 3 measurement runs per (host × model ×
prompt-length), Ollama's canonical timing fields, temp=0 greedy,
num_predict=256. All three Ollama servers accessed via HTTP (Strix
via Tailscale).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

README.md

@@ -18,6 +18,7 @@ Research corpus and implementation guidance for Google Gemma 4, based on product
 | `docs/openwebui-setup.md` | How to configure Gemma 4 inside OpenWebUI — per-setting reference, two ready-to-bake Workspace Model profiles (chat + extract), and a symptom→cause troubleshooting table mapped back to GOTCHAS.md. Assumes Ollama + OpenWebUI are already running. | When setting up or debugging a Gemma 4 model in OpenWebUI, or handing the front-end config to someone else |
 | `docs/reference/bakeoff-2026-04-18.md` | CLI-coding-agent bakeoff on 3090 Ti. **Rounds 1/2 misidentified the cause; Round 3 (the correct one): `think: false` silent-stops gemma4:26b at certain multi-turn states on 32K context.** 31B and Qwen3-Coder robust to the flag. Harness at `scripts/bakeoff/` | When deciding which model to back a CLI agent with, writing a custom agent payload, or debugging a silent tool-call halt |
 | `docs/reference/mort-bakeoff-2026-04-18.md` | mort-bot-specific `think=true` vs `think=false` bakeoff on mort's actual loop shape (gemma4:26b, num_ctx=8192). **Thinking does NOT accumulate in context on Ollama 0.20.4** — strips it from serialized history. Both settings behave identically on step counts, tool counts, wall clock. Harness at `scripts/mort-bakeoff/` | When deciding mort-bot's THINK env var, or when someone claims "think=true eats context" without pinning an Ollama version |
+| `docs/reference/gpu-bakeoff-2026-04-20.md` | Cross-GPU throughput bakeoff: steel141 RTX 3090 Ti vs pve197 V100 vs matt-strix (AMD Strix Halo). **3090 Ti wins decode decisively (128 tok/s on 26B MoE). Strix gets ~42% of that on ~25% of the bandwidth. V100 numbers are degraded because SDXL on CT 167 occupies 31/32 GB of its VRAM.** Also quantifies the MoE vs dense gap: 26B decodes ~4.7× faster than 31B on every card. Harness at `scripts/gpu-bakeoff/` | When choosing which host to run a Gemma 4 workload on, or deciding whether the V100 needs isolated for a given job |
 | `tooling/` | **Canonical upstream tooling** — real scripts, notebooks, model cards, and configs pulled from Google / HF / framework maintainers (147 files). Subdirs: `google-official/`, `huggingface/`, `inference-frameworks/`, `gemma-family/`, `fine-tuning/`. See `tooling/README.md` for index and findings that update the older `CORPUS_*` docs | When you need authoritative source material — model cards, chat templates, fine-tuning recipes, serving commands for vLLM / llama.cpp / MLX, or to scope a specialized sibling (ShieldGemma, EmbeddingGemma, etc.) |
 
 ## Source Projects