Added this case to the skipped tests due to poorly-posed floating point errors causing larger divergence

Author: danieljvickers

toolchain/mfc/test/cases.py

@@ -1885,6 +1885,15 @@ def foreach_example():
                 # is platform-marginal (CPU goldens fail on most GPU lanes). The fast
                 # "Non-Newtonian -> IBM" suite case covers IBM+NN portably at 1e-12.
                 "2D_ibm_poiseuille_nn",
+                # Two immersed boundaries colliding across a periodic boundary via a stiff
+                # soft-sphere spring. The spring acts as a strong amplifier: it turns the
+                # ~1e-13 CPU/GPU floating-point difference in the hydrodynamic force (the
+                # order-dependent atomic surface-pressure integral) into an exponentially
+                # growing trajectory divergence, so the sharp-interface field fails the
+                # golden tolerance on GPU lanes even though both runs are individually
+                # reproducible. Not a correctness bug -- the case is genuinely chaotic at
+                # this stiffness, so it is not a portable regression target.
+                "3D_mibm_periodic_collision",
             ]
             if path in casesToSkip:
                 continue