diff --git a/docs/plans/2026-02-19-phase4-implementation.md b/docs/plans/2026-02-19-phase4-implementation.md
new file mode 100644
index 0000000..f86bea9
--- /dev/null
+++ b/docs/plans/2026-02-19-phase4-implementation.md
@@ -0,0 +1,422 @@
+# Phase 4: Performance Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Add parallel inference slots, a Go benchmark suite measuring decode speed / TTFT / concurrent throughput, and document flash attention results.
+
+**Architecture:** Add `ParallelSlots` to go-inference's `LoadConfig`, wire `--parallel N` through go-rocm's `startServer`, then benchmark three models with `testing.B` and `b.ReportMetric`.
+
+**Tech Stack:** Go testing.B, go-inference options, llama-server `--parallel` flag
+
+---
+
+### Task 1: Add ParallelSlots to go-inference
+
+Add `ParallelSlots int` field to `LoadConfig` and a `WithParallelSlots(n int) LoadOption` constructor in the go-inference package.
+
+**Files:**
+- Modify: `/home/claude/Code/core/go-inference/options.go`
+
+**Step 1: Add field and option to go-inference**
+
+In `/home/claude/Code/core/go-inference/options.go`, add to `LoadConfig` struct (after `GPULayers`):
+
+```go
+ParallelSlots int // Number of concurrent inference slots (0 = server default)
+```
+
+Add option function after `WithGPULayers`:
+
+```go
+// WithParallelSlots sets the number of concurrent inference slots.
+// Higher values allow parallel Generate/Chat calls but increase VRAM usage.
+// 0 or unset uses the server default (typically 1).
+func WithParallelSlots(n int) LoadOption {
+	return func(c *LoadConfig) { c.ParallelSlots = n }
+}
+```
+
+**Step 2: Verify go-inference builds**
+
+Run: `cd /home/claude/Code/core/go-inference && go build ./...`
+Expected: Clean
+
+**Step 3: Commit go-inference**
+
+```bash
+cd /home/claude/Code/core/go-inference
+git add options.go
+git commit -m "feat: add ParallelSlots to LoadConfig for concurrent inference
+
+Co-Authored-By: Virgil <virgil@lethean.io>"
+```
+
+---
+
+### Task 2: Wire parallel slots through go-rocm
+
+Pass `--parallel N` to llama-server when `ParallelSlots > 0`. Update `startServer` and `LoadModel`.
+
+**Files:**
+- Modify: `server.go` (add parallelSlots param to startServer)
+- Modify: `backend.go` (pass cfg.ParallelSlots)
+
+**Step 1: Add parallelSlots to startServer**
+
+In `server.go`, change the `startServer` signature from:
+
+```go
+func startServer(binary, modelPath string, gpuLayers, ctxSize int) (*server, error) {
+```
+
+to:
+
+```go
+func startServer(binary, modelPath string, gpuLayers, ctxSize, parallelSlots int) (*server, error) {
+```
+
+In the args-building section (after the `--ctx-size` block at line ~106), add:
+
+```go
+if parallelSlots > 0 {
+	args = append(args, "--parallel", strconv.Itoa(parallelSlots))
+}
+```
+
+**Step 2: Update LoadModel in backend.go**
+
+Change the `startServer` call from:
+
+```go
+srv, err := startServer(binary, path, cfg.GPULayers, ctxLen)
+```
+
+to:
+
+```go
+srv, err := startServer(binary, path, cfg.GPULayers, ctxLen, cfg.ParallelSlots)
+```
+
+**Step 3: Update TestStartServer_RetriesOnProcessExit in server_test.go**
+
+The test calls `startServer` directly. Update from:
+
+```go
+_, err := startServer("/bin/false", "/nonexistent/model.gguf", 999, 0)
+```
+
+to:
+
+```go
+_, err := startServer("/bin/false", "/nonexistent/model.gguf", 999, 0, 0)
+```
+
+**Step 4: Run tests**
+
+Run: `go test ./...`
+Expected: All tests PASS
+
+Run: `go vet ./...`
+Expected: Clean
+
+**Step 5: Commit**
+
+```bash
+git add server.go backend.go server_test.go
+git commit -m "feat: pass --parallel N to llama-server for concurrent inference slots
+
+Co-Authored-By: Virgil <virgil@lethean.io>"
+```
+
+---
+
+### Task 3: Benchmark Suite
+
+Go benchmark tests measuring decode speed, time-to-first-token, and concurrent throughput across three models.
+
+**Files:**
+- Create: `rocm_benchmark_test.go`
+
+**Step 1: Write the benchmark file**
+
+Create `rocm_benchmark_test.go`:
+
+```go
+//go:build rocm
+
+package rocm
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"forge.lthn.ai/core/go-inference"
+)
+
+// benchModels lists the models to benchmark.
+// Each loads ~3-9 GB of VRAM, so they run sequentially (one at a time).
+var benchModels = []struct {
+	name string
+	path string
+}{
+	{"Gemma3-4B-Q4_K_M", "/data/lem/gguf/LEK-Gemma3-4B-Q4_K_M.gguf"},
+	{"Llama3.1-8B-Q4_K_M", "/data/lem/gguf/LEK-Llama-3.1-8B-Q4_K_M.gguf"},
+	{"Qwen2.5-7B-Q4_K_M", "/data/lem/gguf/LEK-Qwen-2.5-7B-Q4_K_M.gguf"},
+}
+
+func skipBenchIfUnavailable(b *testing.B) {
+	b.Helper()
+	if _, err := os.Stat("/dev/kfd"); err != nil {
+		b.Skip("no ROCm hardware")
+	}
+	if _, err := findLlamaServer(); err != nil {
+		b.Skip("llama-server not found")
+	}
+}
+
+// loadBenchModel loads a model for benchmarking. Caller must defer m.Close().
+// Stops the benchmark timer during loading so load time isn't measured.
+func loadBenchModel(b *testing.B, path string, opts ...inference.LoadOption) inference.TextModel {
+	b.Helper()
+	if _, err := os.Stat(path); err != nil {
+		b.Skipf("model not available: %s", path)
+	}
+
+	b.StopTimer()
+	backend := &rocmBackend{}
+	defaults := []inference.LoadOption{inference.WithContextLen(2048)}
+	m, err := backend.LoadModel(path, append(defaults, opts...)...)
+	if err != nil {
+		b.Fatalf("load model: %v", err)
+	}
+
+	if vram, err := GetVRAMInfo(); err == nil {
+		b.Logf("VRAM after load: %d MiB used / %d MiB total",
+			vram.Used/(1024*1024), vram.Total/(1024*1024))
+	}
+
+	b.StartTimer()
+	return m
+}
+
+// BenchmarkDecode measures token generation speed (tok/s).
+// Generates 128 tokens per iteration, reports tokens/second.
+func BenchmarkDecode(b *testing.B) {
+	skipBenchIfUnavailable(b)
+
+	for _, model := range benchModels {
+		b.Run(model.name, func(b *testing.B) {
+			m := loadBenchModel(b, model.path)
+			defer m.Close()
+
+			const maxTok = 128
+
+			b.ResetTimer()
+			for range b.N {
+				ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
+				var count int
+				for range m.Generate(ctx, "Explain the theory of relativity in detail.", inference.WithMaxTokens(maxTok)) {
+					count++
+				}
+				cancel()
+
+				if count > 0 {
+					tokPerSec := float64(count) / (b.Elapsed().Seconds() / float64(b.N))
+					b.ReportMetric(tokPerSec, "tok/s")
+				}
+			}
+		})
+	}
+}
+
+// BenchmarkTTFT measures time-to-first-token latency.
+// Reports the time from request start to first token received.
+func BenchmarkTTFT(b *testing.B) {
+	skipBenchIfUnavailable(b)
+
+	for _, model := range benchModels {
+		b.Run(model.name, func(b *testing.B) {
+			m := loadBenchModel(b, model.path)
+			defer m.Close()
+
+			b.ResetTimer()
+			for range b.N {
+				ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+				start := time.Now()
+				var ttft time.Duration
+				for range m.Generate(ctx, "Hello", inference.WithMaxTokens(1)) {
+					if ttft == 0 {
+						ttft = time.Since(start)
+					}
+				}
+				cancel()
+
+				if ttft > 0 {
+					b.ReportMetric(float64(ttft.Microseconds()), "µs/first-tok")
+				}
+			}
+		})
+	}
+}
+
+// BenchmarkConcurrent measures throughput with multiple goroutines.
+// Uses 4 parallel slots and 4 goroutines generating simultaneously.
+func BenchmarkConcurrent(b *testing.B) {
+	skipBenchIfUnavailable(b)
+
+	for _, model := range benchModels {
+		b.Run(model.name, func(b *testing.B) {
+			m := loadBenchModel(b, model.path, inference.WithParallelSlots(4))
+			defer m.Close()
+
+			const numWorkers = 4
+			const maxTok = 32
+
+			prompts := []string{
+				"The capital of France is",
+				"The capital of Germany is",
+				"The capital of Italy is",
+				"The capital of Spain is",
+			}
+
+			b.ResetTimer()
+			for range b.N {
+				var wg sync.WaitGroup
+				var mu sync.Mutex
+				totalTokens := 0
+
+				wg.Add(numWorkers)
+				start := time.Now()
+				for i := range numWorkers {
+					go func(idx int) {
+						defer wg.Done()
+						ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+						defer cancel()
+
+						count := 0
+						for range m.Generate(ctx, prompts[idx], inference.WithMaxTokens(maxTok)) {
+							count++
+						}
+
+						mu.Lock()
+						totalTokens += count
+						mu.Unlock()
+					}(i)
+				}
+				wg.Wait()
+				elapsed := time.Since(start)
+
+				if totalTokens > 0 {
+					b.ReportMetric(float64(totalTokens)/elapsed.Seconds(), "tok/s-aggregate")
+					b.ReportMetric(float64(totalTokens), "total-tok")
+				}
+			}
+		})
+	}
+}
+```
+
+**Step 2: Run benchmarks**
+
+Run: `go test -tags rocm -bench BenchmarkDecode/Gemma3 -benchtime 1x -timeout 120s -v`
+Expected: PASS with tok/s metric reported
+
+Run full suite (takes several minutes — each model loads separately):
+`go test -tags rocm -bench . -benchtime 3x -timeout 600s -v`
+Expected: All benchmarks PASS
+
+**Step 3: Verify unit tests still pass**
+
+Run: `go test ./...`
+Expected: All PASS (benchmark file is `//go:build rocm`, won't run without tag)
+
+**Step 4: Commit**
+
+```bash
+git add rocm_benchmark_test.go
+git commit -m "feat: benchmark suite for decode speed, TTFT, and concurrent throughput
+
+Co-Authored-By: Virgil <virgil@lethean.io>"
+```
+
+---
+
+### Task 4: Flash Attention Comparison (Manual)
+
+Build a second llama-server without flash attention and compare performance. No code changes — just benchmarks and documentation.
+
+**Step 1: Build llama-server without flash attention**
+
+```bash
+cd /home/claude/llama.cpp
+cmake -B build-nofa \
+    -DGGML_HIP=ON \
+    -DAMDGPU_TARGETS=gfx1100 \
+    -DCMAKE_BUILD_TYPE=Release
+cmake --build build-nofa --parallel $(nproc) -t llama-server
+```
+
+Note: This is the same as the original build but WITHOUT `-DGGML_HIP_ROCWMMA_FATTN=ON`.
+
+**Step 2: Run benchmarks with flash attention (current binary)**
+
+```bash
+go test -tags rocm -bench BenchmarkDecode -benchtime 3x -timeout 600s -v 2>&1 | tee /tmp/bench-fa.txt
+```
+
+**Step 3: Run benchmarks without flash attention**
+
+```bash
+ROCM_LLAMA_SERVER_PATH=/home/claude/llama.cpp/build-nofa/bin/llama-server \
+go test -tags rocm -bench BenchmarkDecode -benchtime 3x -timeout 600s -v 2>&1 | tee /tmp/bench-nofa.txt
+```
+
+**Step 4: Record results in FINDINGS.md**
+
+Add a section to FINDINGS.md documenting the comparison.
+
+---
+
+### Task 5: Update TODO.md and FINDINGS.md
+
+**Files:**
+- Modify: `TODO.md` (mark Phase 4 items done)
+- Modify: `FINDINGS.md` (add benchmark results, flash attention comparison)
+
+**Step 1: Update TODO.md**
+
+Mark all Phase 4 items `[x]` with commit references and dates.
+
+**Step 2: Update FINDINGS.md**
+
+Add a `## 2026-02-19: Phase 4 Performance (Charon)` section with:
+- Benchmark results table (tok/s, TTFT, VRAM for each model)
+- Flash attention comparison
+- Concurrent throughput numbers
+- Notes on parallel slots and VRAM impact
+
+**Step 3: Commit**
+
+```bash
+git add TODO.md FINDINGS.md
+git commit -m "docs: Phase 4 complete — benchmarks, flash attention, parallel slots
+
+Co-Authored-By: Virgil <virgil@lethean.io>"
+```
+
+---
+
+## Summary
+
+| Task | What | Files | Test Type |
+|------|------|-------|-----------|
+| 1 | ParallelSlots in go-inference | go-inference/options.go | Build check |
+| 2 | Wire --parallel to llama-server | server.go, backend.go, server_test.go | Unit |
+| 3 | Benchmark suite | rocm_benchmark_test.go | Benchmark (GPU) |
+| 4 | Flash attention comparison | (manual) | Benchmark (GPU) |
+| 5 | Documentation | TODO.md, FINDINGS.md | N/A |