bench: faithful Claude Code transcript replay + repro & build-flag fix for empty-response bug

.gitmodules

@@ -1,4 +1,7 @@
 [submodule "dflash/deps/llama.cpp"]
 	path = dflash/deps/llama.cpp
-	url = https://github.com/Luce-Org/llama.cpp.git
-	branch = luce-dflash
+	# Temporarily pointing at easel's fork (branch fix/cuda-vmm-pool-extension-race)
+	# while https://github.com/Luce-Org/llama.cpp-dflash-ggml/pull/4 is in review.
+	# Revert to Luce-Org/llama.cpp.git once that PR is merged.
+	url = https://github.com/easel/llama.cpp-dflash-ggml.git
+	branch = fix/cuda-vmm-pool-extension-race

dflash/CMakeLists.txt

@@ -119,6 +119,11 @@ if(DFLASH27B_TESTS)
         target_include_directories(smoke_load_draft PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src)
         target_link_libraries(smoke_load_draft PRIVATE dflash27b ggml ggml-cuda)
     endif()
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/spike_thin_copy.cpp")
+        add_executable(spike_thin_copy test/spike_thin_copy.cpp)
+        target_include_directories(spike_thin_copy PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src)
+        target_link_libraries(spike_thin_copy PRIVATE ggml ggml-cuda)
+    endif()
     if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/smoke_draft_graph.cpp")
         add_executable(smoke_draft_graph test/smoke_draft_graph.cpp)
         target_include_directories(smoke_draft_graph PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/src)