diff --git a/examples/3D_ibm_neighborhood_radius/README.md b/examples/3D_ibm_neighborhood_radius/README.md
new file mode 100644
index 000000000..f635700ef
--- /dev/null
+++ b/examples/3D_ibm_neighborhood_radius/README.md
@@ -0,0 +1,77 @@
+# Automatic `ib_neighborhood_radius` when ranks are not cubes
+
+`ib_neighborhood_radius` is a count of **rank hops**. A rank keeps an immersed-boundary patch only while the
+patch's centroid lies inside its own subdomain grown outward by that many hops
+(`s_get_neighbor_bounds`, `f_neighborhood_ranks_own_location`), and drops it otherwise. So the radius has to be
+large enough that the body is reachable within that many hops **in every direction** — and the hop that needs
+the most is the one stepping across the *thinnest* rank.
+
+When the radius is not set in the case file, MFC chooses it from the body's half-extent divided by a rank
+width. The width it used was assembled the wrong way round:
+
+```fortran
+local_rank_width = -1._wp
+do each direction:
+    local_rank_width = max(local_rank_width, <this rank's extent in that direction>)
+call s_mpi_allreduce_min(local_rank_width, min_rank_width)
+```
+
+Each rank reports its **widest** extent, and the minimum is taken over ranks. On a decomposition where every
+rank is long in one direction and thin in another, the reported width is the long one, the radius comes out too
+small, and ranks that should have kept the patch drop it.
+
+## The case
+
+A thin plate in a long, narrow channel: 20 chords by 8 by 6, with 400 x 50 x 50 cells chosen so MFC's topology
+search settles on **16 x 2 x 2** at 64 ranks. That is an ordinary shape for a wake, a jet or a channel — the
+flow direction resolved far more finely than the cross-stream ones — and it makes the ranks strongly
+anisotropic:
+
+| direction | ranks | extent per rank |
+| --- | --- | --- |
+| x | 16 | **1.250** |
+| y | 2 | 4.000 |
+| z | 2 | 3.000 |
+
+The plate is 1.0 x 2.4 x 0.1, so `s_get_ib_bound` (geometry 9, the cuboid's half-diagonal) returns **1.3010**.
+Crossing that at 1.250 per hop needs `ceil(1.1 * 1.3010 / 1.250) = 2` hops. The old width of 4.000 gives
+`ceil(1.1 * 1.3010 / 4.000) = 1`.
+
+## Running it
+
+```
+./mfc.sh run examples/3D_ibm_neighborhood_radius/case.py -n 64
+```
+
+and read the line MFC prints at start-up:
+
+```
+Automatic choice of ib_neighborhood_radius selected:  N
+```
+
+| | printed radius |
+| --- | --- |
+| before | **1** |
+| after | **2** |
+
+Both runs complete; the case is 1 M cells and takes a few minutes on two CPU nodes. `SUMMARY=1 python3
+case.py` prints the half-extent, the rank extents and the arithmetic above without running anything.
+
+## It is not only this case
+
+The same two production grids that motivated the fix, measured from their own `lustre_*_cb.dat`:
+
+| case | topology | old width | old radius | new width | new radius |
+| --- | --- | --- | --- | --- | --- |
+| gust encounter, 128 ranks | 16 x 2 x 4 | 2.051 | 1 | 1.052 | **2** |
+| flapping wing, 128 ranks | 8 x 4 x 4 | 1.745 | 1 | 1.027 | **2** |
+
+Both pick 1 where 2 is required.
+
+## Scope
+
+The new width is never larger than the old one, so the chosen radius never decreases: the change can only make
+the neighbourhood more conservative, at the cost of more hops in the force reduction. Cases that set
+`ib_neighborhood_radius` explicitly are untouched, and so is any decomposition whose ranks are close to cubic,
+where the widest and narrowest extents coincide — which is why a uniform grid with a balanced topology shows no
+difference.
diff --git a/examples/3D_ibm_neighborhood_radius/case.py b/examples/3D_ibm_neighborhood_radius/case.py
new file mode 100644
index 000000000..1669458c6
--- /dev/null
+++ b/examples/3D_ibm_neighborhood_radius/case.py
@@ -0,0 +1,132 @@
+#!/usr/bin/env python3
+"""Automatic ib_neighborhood_radius on a decomposition whose ranks are not cubes.
+
+A thin plate held in a long, narrow channel. The domain is 20 chords long and 8 by 6 across, and the cell
+counts (400 x 50 x 50) are chosen so MFC's topology search settles on 16 x 2 x 2 at 64 ranks. That makes the
+x ranks 1.25 chords wide while the y and z ranks are 4.0 and 3.0 -- an ordinary situation for a wake, a jet or
+a channel, where the flow direction is resolved far more finely than the cross-stream ones.
+
+`ib_neighborhood_radius` is deliberately left unset, so MFC chooses it at start-up and prints
+
+    Automatic choice of ib_neighborhood_radius selected:  N
+
+The radius counts rank hops, and the body must be reachable within that many hops in *every* direction, so
+the hop that matters is the one crossing the thinnest rank. The plate's half-extent is 1.301 chords and the
+thinnest rank is 1.250 wide, so it needs 2 hops, not 1.
+
+See README.md for the measured before/after and the argument.
+"""
+
+import json
+import math
+import os
+
+Re, Ma, gamma, U, rho = 1000.0, 0.1, 1.4, 1.0, 1.0
+cs = U / Ma
+P = rho * cs**2 / gamma
+c = 1.0  # chord
+SPAN = float(os.environ.get("SPAN", 2.4))  # spanwise length of the plate
+THICK = 0.1 * c
+
+M = int(os.environ.get("M", 400))  # cells in x; 400/16 ranks = 25, the stencil minimum
+N = int(os.environ.get("N", 50))  # cells in y; 50/2 ranks = 25
+PZ = int(os.environ.get("PZ", 50))  # cells in z; 50/2 ranks = 25
+TSTOP = float(os.environ.get("TSTOP", 0.5))
+
+x0, x1 = -10.0 * c, 10.0 * c
+y0, y1 = -4.0 * c, 4.0 * c
+z0, z1 = -3.0 * c, 3.0 * c
+
+case = {
+    "run_time_info": "T",
+    "parallel_io": "T",
+    "prim_vars_wrt": "T",
+    "ib_state_wrt": "T",
+    "format": "silo",
+    "precision": "double",
+    "x_domain%beg": x0,
+    "x_domain%end": x1,
+    "y_domain%beg": y0,
+    "y_domain%end": y1,
+    "z_domain%beg": z0,
+    "z_domain%end": z1,
+    "m": M - 1,
+    "n": N - 1,
+    "p": PZ - 1,
+    "cyl_coord": "F",
+    "cfl_adap_dt": "T",
+    "cfl_target": 0.4,
+    "n_start": 0,
+    "t_save": TSTOP,
+    "t_stop": TSTOP,
+    "num_patches": 1,
+    "num_fluids": 1,
+    "model_eqns": "5eq",
+    "alt_soundspeed": "F",
+    "mpp_lim": "F",
+    "mixture_err": "T",
+    "time_stepper": "rk3",
+    "weno_order": 5,
+    "weno_eps": 1.0e-10,
+    "weno_Re_flux": "T",
+    "weno_avg": "T",
+    "avg_state": "arithmetic",
+    "mapped_weno": "T",
+    "null_weights": "F",
+    "mp_weno": "F",
+    "riemann_solver": "hllc",
+    "low_Mach": 2,
+    "wave_speeds": "direct",
+    "viscous": "T",
+    "fd_order": 4,
+    "patch_icpp(1)%geometry": 9,
+    "patch_icpp(1)%x_centroid": 0.0,
+    "patch_icpp(1)%y_centroid": 0.0,
+    "patch_icpp(1)%z_centroid": 0.0,
+    "patch_icpp(1)%length_x": x1 - x0,
+    "patch_icpp(1)%length_y": y1 - y0,
+    "patch_icpp(1)%length_z": z1 - z0,
+    "patch_icpp(1)%vel(1)": U,
+    "patch_icpp(1)%vel(2)": 0.0,
+    "patch_icpp(1)%vel(3)": 0.0,
+    "patch_icpp(1)%pres": P,
+    "patch_icpp(1)%alpha_rho(1)": rho,
+    "patch_icpp(1)%alpha(1)": 1.0,
+    "fluid_pp(1)%gamma": 1.0 / (gamma - 1.0),
+    "fluid_pp(1)%eos": "ideal_gas",
+    "fluid_pp(1)%Re(1)": Re / (c * U),
+    "bc_x%beg": -7,
+    "bc_x%grcbc_in": "T",
+    "bc_x%vel_in(1)": U,
+    "bc_x%vel_in(2)": 0.0,
+    "bc_x%vel_in(3)": 0.0,
+    "bc_x%pres_in": P,
+    "bc_x%alpha_rho_in(1)": rho,
+    "bc_x%alpha_in(1)": 1.0,
+    "bc_x%end": -8,
+    "bc_y%beg": -8,
+    "bc_y%end": -8,
+    "bc_z%beg": -8,
+    "bc_z%end": -8,
+    # ib_neighborhood_radius deliberately NOT set: this case exists to exercise the automatic choice
+    "ib": "T",
+    "num_ibs": 1,
+    "patch_ib(1)%geometry": 9,  # cuboid
+    "patch_ib(1)%x_centroid": 0.0,
+    "patch_ib(1)%y_centroid": 0.0,
+    "patch_ib(1)%z_centroid": 0.0,
+    "patch_ib(1)%length_x": c,
+    "patch_ib(1)%length_y": SPAN,
+    "patch_ib(1)%length_z": THICK,
+    "patch_ib(1)%slip": "F",
+    "patch_ib(1)%moving_ibm": 0,
+}
+
+if __name__ == "__main__":
+    if os.environ.get("SUMMARY"):
+        bound = 0.5 * math.sqrt(c**2 + SPAN**2 + THICK**2)
+        print(f"plate half-extent (s_get_ib_bound, geometry 9): {bound:.4f}")
+        print(f"rank extents at 64 ranks (16 x 2 x 2): x {(x1 - x0) / 16:.3f}, " f"y {(y1 - y0) / 2:.3f}, z {(z1 - z0) / 2:.3f}")
+        print(f"hops needed across the thinnest rank: ceil(1.1 * {bound:.4f} / {(x1 - x0) / 16:.3f}) = " f"{max(1, math.ceil(1.1 * bound / ((x1 - x0) / 16)))}")
+    else:
+        print(json.dumps(case, indent=4))
diff --git a/src/simulation/m_start_up.fpp b/src/simulation/m_start_up.fpp
index 6a5434f5b..f1faba9af 100644
--- a/src/simulation/m_start_up.fpp
+++ b/src/simulation/m_start_up.fpp
@@ -1542,10 +1542,14 @@ contains
                 max_ib_bound = max(max_ib_bound, particle_cloud(k)%radius)
             end do
 
-            ! determine the upper bound on the size
-            local_rank_width = -1._wp
+            ! Narrowest rank extent, over every direction as well as every rank. The radius is a count of rank
+            ! hops, so the distance one hop covers is the extent of the rank it steps over, and the direction
+            ! needing the most hops to span the body is the one whose ranks are thinnest. Reducing over each
+            ! rank's widest extent first reports the wrong number whenever ranks are anisotropic, which is the
+            ! norm on a stretched grid or an elongated domain.
+            local_rank_width = huge(0._wp)
             #:for X, ID, DIM in [('x', 1, 'm'), ('y', 2, 'n'), ('z', 3, 'p')]
-                if (num_dims >= ${ID}$) local_rank_width = max(local_rank_width, abs(${X}$_cb(${DIM}$) - ${X}$_cb(-1)))
+                if (num_dims >= ${ID}$) local_rank_width = min(local_rank_width, abs(${X}$_cb(${DIM}$) - ${X}$_cb(-1)))
             #:endfor
             call s_mpi_allreduce_min(local_rank_width, min_rank_width)
 
diff --git a/toolchain/mfc/test/cases.py b/toolchain/mfc/test/cases.py
index fff3b4552..1639a8d84 100644
--- a/toolchain/mfc/test/cases.py
+++ b/toolchain/mfc/test/cases.py
@@ -3167,6 +3167,8 @@ def foreach_example():
                 # the transverse momentum drifts past the 1e-3 Example tolerance across compilers
                 # (nvhpc passes; Intel and CCE disagree by ~2e-3 absolute). No single golden is portable.
                 "2D_hybrid_slab",
+                # Needs its 16 x 2 x 2 rank topology; the Example suite runs it on one rank and a shrunken grid.
+                "3D_ibm_neighborhood_radius",
             ]
             if path in casesToSkip:
                 continue