Add study tune loop and smoke configs

configs/examples/dash0_smoke_proposals/aggressive.json

@@ -0,0 +1,19 @@
+{
+  "observation": "Push batching further after validating the balanced layout.",
+  "diagnosis": "If TTFT remains under control, a larger admission window should maximize achieved request rate.",
+  "config_patch": {
+    "env_patch": {},
+    "flag_patch": {
+      "tensor-parallel-size": 4,
+      "max-num-seqs": 24,
+      "max-num-batched-tokens": 98304,
+      "gpu-memory-utilization": 0.92,
+      "block-size": 64
+    }
+  },
+  "expected_effects": [
+    "Highest throughput among the smoke candidates if memory is sufficient",
+    "More pressure on TTFT, so the binary search should locate the safe threshold"
+  ],
+  "why_not_previous_failures": "Keeps tp=4 and block-size stable while only expanding batching and memory utilization."
+}

configs/examples/dash0_smoke_proposals/balanced.json

@@ -0,0 +1,19 @@
+{
+  "observation": "Increase batching once tp=4 is fixed.",
+  "diagnosis": "Throughput should improve if the engine can admit more concurrent prefills without violating TTFT.",
+  "config_patch": {
+    "env_patch": {},
+    "flag_patch": {
+      "tensor-parallel-size": 4,
+      "max-num-seqs": 16,
+      "max-num-batched-tokens": 65536,
+      "gpu-memory-utilization": 0.9,
+      "block-size": 64
+    }
+  },
+  "expected_effects": [
+    "Higher feasible sampling_u than the conservative baseline",
+    "Better token throughput if memory headroom is sufficient"
+  ],
+  "why_not_previous_failures": "Raises batching in a controlled step instead of jumping directly to the most aggressive setting."
+}

configs/examples/dash0_smoke_proposals/conservative.json

@@ -0,0 +1,19 @@
+{
+  "observation": "Start from a safe tp=4 layout and conservative batching.",
+  "diagnosis": "The first pass should verify multi-GPU launch and avoid queueing collapse from over-batching.",
+  "config_patch": {
+    "env_patch": {},
+    "flag_patch": {
+      "tensor-parallel-size": 4,
+      "max-num-seqs": 8,
+      "max-num-batched-tokens": 32768,
+      "gpu-memory-utilization": 0.85,
+      "block-size": 64
+    }
+  },
+  "expected_effects": [
+    "Stable startup on 4x H20",
+    "Low risk of OOM during the first binary-search probes"
+  ],
+  "why_not_previous_failures": "This is the initial baseline proposal."
+}