fix(#46): correct autotuning search space docs from 648 to 216

docs/notes/06-paper-outline.md

@@ -375,7 +375,7 @@ AutoTVM and Triton search over kernel implementation variants (tiling, unrolling
 
 ### 10.2 Quantum Adaptation
 
-Qiskit's transpiler offers routing methods (stochastic, VF2), layout methods (trivial, VF2Layout, dense), and optimization levels (0–3). But there is no systematic autotuning over these options for a specific circuit-device pair.
+Qiskit's transpiler offers routing methods (stochastic, sabre), layout methods (dense, VF2Layout), optimization levels (1–3), and scheduling methods (ASAP, ALAP, coherence-aware). But there is no systematic autotuning over these options for a specific circuit-device pair.
 
 We propose **quantum autotuning**: search over transpilation configurations, benchmark each on the target device, and select the optimal one.
 
@@ -525,7 +525,7 @@ Algorithm:
 **Objective:** Show that autotuning transpiler settings outperforms defaults.
 
 **Protocol:**
-1. Define search space: 3 routing × 3 layout × 4 opt_levels × 3 seeds × 2 fusion = 216 configurations
+1. Define search space: 2 routing × 2 layout × 3 opt_levels × 3 seeds × 2 fusion × 3 scheduling = 216 configurations
 2. For each benchmark circuit:
    a. Transpile with each configuration, estimate error
    b. Select top-5 configurations, run on hardware with 1024 shots (screening)

src/qc_compiler/autotuning.py

@@ -5,9 +5,9 @@
 device-optimal one, then caches results for similar circuit families.
 
 The search space includes:
-    - routing_method: stochastic, vf2, sabre
-    - layout_method: trivial, dense, vf2_layout
-    - optimization_level: 0, 1, 2, 3
+    - routing_method: stochastic, sabre
+    - layout_method: dense, vf2_layout
+    - optimization_level: 1, 2, 3
     - seed: 0, 1, 2 (randomness in routing/layout)
     - gate_fusion: on/off
     - scheduling_method: asap, alap, coherence_aware
@@ -260,9 +260,9 @@ def _generate_configurations(self) -> list[TranspileConfig]:
 
         Returns a reduced set of configurations that covers the most
         impactful parameter combinations. The full search space
-        (3 routing × 3 layout × 4 opt × 3 seed × 2 fusion × 3 scheduling
-        = 648 configs) is reduced by only varying the most impactful
-        parameters together.
+        (2 routing × 2 layout × 3 opt × 3 seed × 2 fusion × 3 scheduling
+        = 216 configs) varies the most impactful parameters
+        together.
 
         Returns:
             List of TranspileConfig objects to evaluate.

tests/test_autotuning.py

@@ -293,4 +293,31 @@ def test_search_with_backend_sets_best_circuit(self, tuner_with_backend):
             qc, circuit_family="test_best_circuit"
         )
         if result.best_circuit is not None:
-            assert result.best_circuit.num_qubits == qc.num_qubits
+            assert result.best_circuit.num_qubits == qc.num_qubits
+
+
+class TestSearchSpace:
+    """Regression tests for autotuning search space size (issue #46)."""
+
+    def test_search_space_size_is_216(self):
+        tuner = AutoTuner(cost_model=CostModel())
+        configs = tuner._generate_configurations()
+        assert len(configs) == 216
+
+    def test_search_space_covers_all_routing_methods(self):
+        tuner = AutoTuner(cost_model=CostModel())
+        configs = tuner._generate_configurations()
+        routing_methods = {c.routing_method for c in configs}
+        assert routing_methods == {"stochastic", "sabre"}
+
+    def test_search_space_covers_all_layout_methods(self):
+        tuner = AutoTuner(cost_model=CostModel())
+        configs = tuner._generate_configurations()
+        layout_methods = {c.layout_method for c in configs}
+        assert layout_methods == {"dense", "vf2_layout"}
+
+    def test_search_space_covers_all_optimization_levels(self):
+        tuner = AutoTuner(cost_model=CostModel())
+        configs = tuner._generate_configurations()
+        opt_levels = {c.optimization_level for c in configs}
+        assert opt_levels == {1, 2, 3}