Reset new topology groups to full binary search

src/aituner/llm.py

@@ -301,8 +301,9 @@ def build_prompt(
         "",
         "The primary cross-topology comparison metric is request_rate_per_gpu, not raw request_rate.",
         "The proposal should beat the incumbent on request_rate_per_gpu under the 95%+ SLO target.",
-        "The evaluator uses the best feasible sampling_u from the same tp_dp_product group when it exists. For a different tp_dp_product group, it uses a scaled lower floor instead of reusing the global incumbent directly.",
-        "Do not assume a configuration with fewer GPUs must start from the global incumbent sampling_u.",
+        "The evaluator uses the best feasible sampling_u from the same tp_dp_product group when it exists.",
+        "If a tp_dp_product group has no history yet, the evaluator starts from the study's original search.low and runs a full binary search for that group.",
+        "Do not assume a configuration with fewer GPUs should inherit the global incumbent sampling_u.",
     ]
     return "\n".join(sections)
 

src/aituner/store.py

@@ -253,12 +253,6 @@ def _derive_search_floor(*, study: StudySpec, state: StudyState, parallel_size:
         group_incumbent.get("best_sampling_u"), (int, float)
     ):
         candidate = float(group_incumbent["best_sampling_u"])
-    elif (
-        isinstance(state.best_sampling_u, (int, float))
-        and isinstance(state.best_parallel_size, int)
-        and state.best_parallel_size > 0
-    ):
-        candidate = float(state.best_sampling_u) * float(parallel_size) / float(state.best_parallel_size)
     else:
         candidate = low
     return min(high, max(low, candidate))