add compute profiler: CPU bottleneck detection per SemiWiki article

src/upgrades/rsi-guard.ts

@@ -5,6 +5,13 @@
  *   Sakana AI — RSI Lab (rsi-lab.sakana.ai)
  *   Anthropic — "When AI builds itself" (anthropic.com/institute/recursive-self-improvement)
  *   Stephen Wolfram — "Games between Programs" (writings.stephenwolfram.com)
+ *   SemiWiki — "Agentic AI Demands More Than GPUs" (semiwiki.com)
+ *
+ * Key insight from SemiWiki: CPU orchestration, not GPU inference, is the
+ * bottleneck for agentic AI. Safety checks, tool orchestration, verification,
+ * and sub-agent coordination all run on CPU. Our system's safety stack
+ * (content gate, decomp guard, prompt adapter) runs before any model call —
+ * this is efficient: we block unsafe tasks without wasting GPU cycles.
  *
  * Three RSI bottlenecks this module addresses:
  *
@@ -285,3 +292,78 @@ export class GoalPersistenceGuard {
     return lines.join("\n");
   }
 }
+
+// ── Compute Profiler (SemiWiki insight) ──────────────
+
+export interface ComputeSample {
+  timestamp: string;
+  phase: "safety_check" | "model_inference" | "tool_execution" | "verification" | "orchestration";
+  durationMs: number;
+  isCpu: boolean;
+}
+
+export interface ComputeProfile {
+  totalCpuMs: number;
+  totalGpuMs: number;
+  cpuRatio: number;
+  bottleneckPhase: string;
+  recommendation: string;
+}
+
+export class ComputeProfiler {
+  private store: StateStore;
+
+  constructor(store?: StateStore) {
+    this.store = store ?? new StateStore();
+    this.store.ensureSubDir("compute");
+  }
+
+  /** Record a compute sample */
+  record(sample: ComputeSample): void {
+    this.store.append("compute", "samples.jsonl", sample);
+  }
+
+  /** Profile recent compute usage */
+  profile(samples: number = 50): ComputeProfile {
+    const recent = this.store.readLines<ComputeSample>("compute", "samples.jsonl")
+      .slice(-samples);
+
+    if (recent.length === 0) {
+      return { totalCpuMs: 0, totalGpuMs: 0, cpuRatio: 0, bottleneckPhase: "unknown", recommendation: "No data yet" };
+    }
+
+    const cpuSamples = recent.filter((s) => s.isCpu);
+    const gpuSamples = recent.filter((s) => !s.isCpu);
+    const totalCpuMs = cpuSamples.reduce((s, c) => s + c.durationMs, 0);
+    const totalGpuMs = gpuSamples.reduce((s, c) => s + c.durationMs, 0);
+    const totalMs = totalCpuMs + totalGpuMs;
+
+    // Find the phase that consumes the most time
+    const phaseTimes = new Map<string, number>();
+    for (const s of recent) {
+      phaseTimes.set(s.phase, (phaseTimes.get(s.phase) ?? 0) + s.durationMs);
+    }
+    const bottleneckPhase = [...phaseTimes.entries()].sort((a, b) => b[1] - a[1])[0]?.[0] ?? "unknown";
+
+    const cpuRatio = totalMs > 0 ? totalCpuMs / totalMs : 0;
+    const recommendation = cpuRatio > 0.6
+      ? `CPU bottleneck (${(cpuRatio * 100).toFixed(0)}% of time). Consider: parallel safety checks, vectorized tool execution, async verification.`
+      : `GPU dominates (${((1 - cpuRatio) * 100).toFixed(0)}% of time). Consider: larger batch sizes, model quantization.`;
+
+    return { totalCpuMs, totalGpuMs, cpuRatio, bottleneckPhase, recommendation };
+  }
+
+  /** Get compute report */
+  getReport(): string {
+    const p = this.profile();
+    return [
+      "Compute Profile (SemiWiki bottleneck analysis)",
+      "",
+      `  CPU time:   ${(p.totalCpuMs / 1000).toFixed(1)}s (${(p.cpuRatio * 100).toFixed(0)}%)`,
+      `  GPU time:   ${(p.totalGpuMs / 1000).toFixed(1)}s (${((1 - p.cpuRatio) * 100).toFixed(0)}%)`,
+      `  Bottleneck: ${p.bottleneckPhase}`,
+      "",
+      `  ${p.recommendation}`,
+    ].join("\n");
+  }
+}