fix(rccl): drop the GDR flush case the NCCL sync never wired up

projects/rccl/test/transport/NetIbMPI/GdrFlushTests.cpp

@@ -217,34 +217,12 @@ TEST_F(GdrFlushTest, RepeatedFlush_NoFaultBurst) {
         EXPECT_EQ(flush, ncclSuccess) << "no flush in the burst may raise a QP async-fatal";
 }
 
-// Regression guard - force the removed scratchpad RDMA_WRITE back. On a dma-buf
-// scratchpad this must reproduce the fault (flush no longer succeeds), proving
-// the WRITE is the culprit and that removing it is the fix.
-//
-// Disabled: the NCCL v2.31.2-1 sync (#10785) added this case but not the two
-// things it calls. RunRecvFlushBurst has no forceWrite parameter and there is no
-// gdrSupported(), so a --debug --tests_build --enable-mpi-tests build of develop
-// does not compile once ENABLE_FAULT_INJECTION is on. Restoring the forceWrite
-// hook is a decision for whoever owns the GDR flush fence, not something to
-// guess at here, so the case is guarded off rather than half-fixed.
-#if 0
-TEST_F(GdrFlushTest, ForcedScratchpadWrite_ReproducesFault) {
-    SKIP_UNLESS_MPI_PREREQS(kExactTwoProcesses, kExactTwoProcesses,
-                                          false, kMinGpusPerNode, kNoNodeLimit);
-    if (!cuMemEnabledEnv()) GTEST_SKIP() << "Requires NCCL_CUMEM_ENABLE=1 (dma-buf scratchpad target)";
-    if (!scratchpadFlushEnabled())
-        GTEST_SKIP() << "Requires the scratchpad flush enabled (RCCL_GDR_FLUSH_GPU_MEM_NO_RELAXED_ORDERING=1)";
-    AssertInitAndGetDevices(nullptr);
-    if (!(gdrPtrSupport() & NCCL_PTR_CUDA)) GTEST_SKIP() << "GDR (NCCL_PTR_CUDA) not supported on this device";
-
-    ncclResult_t forced = ncclSuccess;
-    RunRecvFlushBurst(/*iterations=*/1, /*verifyData=*/false,
-                      &forced);
-    if (MPIEnvironment::world_rank == 0)
-        EXPECT_NE(forced, ncclSuccess)
-            << "forced scratchpad RDMA_WRITE on a dma-buf buffer should fault the flush QP";
-}
-#endif  // disabled pending the forceWrite hook
+// The NCCL v2.31.2-1 sync added ForcedScratchpadWrite_ReproducesFault: force the
+// removed scratchpad RDMA_WRITE on a dma-buf flush QP and expect the flush to
+// fault. That case never compiled: gdrSupported() does not exist, and
+// RunRecvFlushBurst has no forceWrite parameter. If a production forceWrite
+// hook is added, restore the case (skip with gdrPtrSupport(), three-arg burst
+// plus the hook) rather than leaving the regression untested.
 
 }  // namespace