docs: Phase 4 performance design

Benchmark suite (testing.B), parallel slots via go-inference,
flash attention manual comparison.

Co-Authored-By: Virgil <virgil@lethean.io>

docs/plans/2026-02-19-phase4-performance-design.md

@@ -0,0 +1,29 @@
+# Phase 4: Performance Design
+
+Approved 19 Feb 2026.
+
+## 1. Benchmark Suite (`rocm_benchmark_test.go`)
+
+Go benchmark tests (`testing.B`) with `b.ReportMetric()`, build-tagged `//go:build rocm`.
+
+Three benchmarks per model:
+
+- **BenchmarkDecode_{Model}** — Generate 128 tokens, report tok/s
+- **BenchmarkTTFT_{Model}** — Time-to-first-token latency (ns)
+- **BenchmarkConcurrent_{Model}** — N goroutines generating simultaneously, report aggregate tok/s
+
+Models: Gemma3-4B-Q4_K_M, Llama-3.1-8B-Q4_K_M, Qwen-2.5-7B-Q4_K_M. Each loads once per benchmark function (load time excluded via `b.StopTimer`/`b.StartTimer`). VRAM logged after load via `GetVRAMInfo()`.
+
+Run: `go test -tags rocm -bench . -benchtime 3x -timeout 600s`
+
+## 2. Flash Attention (Manual)
+
+No code changes. Build a second llama-server without `-DGGML_HIP_ROCWMMA_FATTN=ON`, run the benchmark suite twice with `ROCM_LLAMA_SERVER_PATH` pointing at each binary, record results in FINDINGS.md.
+
+## 3. Parallel Slots
+
+Add `ParallelSlots int` to go-inference's `LoadConfig` and `WithParallelSlots(n int) LoadOption`. Pass `--parallel N` to llama-server in `startServer` when > 0. Benchmark concurrent requests with 1 slot vs 4 slots.
+
+## 4. Testing
+
+Benchmarks are the tests — they produce the numbers. Existing integration tests remain as correctness checks. All results documented in FINDINGS.md.