feat(esopull): runtime body sync for EsoPull streaming mode

- New esopull_sync.cu: DecodeCellsToPhysical + EncodePhysicalToCells
  (compact-list mode, ddf_shifting-aware, encode applies collision).
- sync_bodies() now branches for double_buffer vs esopull: decode
  backing layout to physical DDF on GPU -> host patch -> collide +
  encode back to backing layout. No temp_gpu, no full-grid copy.
- 4 new integration tests covering esopull add/remove/cycle/roundtrip.
- ddf_shifting + esopull + sync_bodies jointly verified (1300 steps
  stable after add/remove).
- Bump version to 0.5.0.

Co-authored-by: Cursor <cursoragent@cursor.com>

README.md

@@ -195,7 +195,6 @@ kernel launches on a user-provided stream, then sync and read later.
 `sync_bodies()` applies all staged body edits (added via `add_body()` and removed via `remove_body()`) to a running simulation without full reinitialization.  The GPU flow field is preserved; only the body-related topology is rebuilt.
 
 **Limitations:**
-- Requires `streaming: "double_buffer"` (esopull raises `NotImplementedError`)
 - Abrupt body introduction causes a transient; force readback is finite but may take 50+ steps to settle
 - Verified for `"circle"` type bodies; sensors and force_regions are also expected to work
   (they produce no curved links so the DDF patch is simpler)
@@ -365,6 +364,7 @@ Current verification scope:
 - 2D D2Q9 only (D3Q19 not yet implemented)
 - MRT collision model (SRT/TRT expected to work but not explicitly validated)
 - Fixed and rotating cylinder benchmarks (Kan99b K2: bit-identical metrics)
+- Runtime body topology sync via ``sync_bodies()`` -- add and remove bodies at runtime
 - `get_macroscopic()` uses GPU kernel for physically correct output
 - `get_ddf()` returns backing-layout data (not physical DDF) in EsoPull mode
 
@@ -389,13 +389,12 @@ Stores `f_i - w_i` instead of `f_i` to improve FP16 accuracy.  Supported with th
 |-----------|-----------|-------|-------------|--------|
 | MRT | double_buffer | zou_he_local | cylinder | Verified (K2 metrics match FP32) |
 | MRT | double_buffer | regularized | cylinder | Under investigation -- use zou_he_local |
-| MRT | esopull | any | any | Not yet verified |
+| MRT | esopull | zou_he_local | cylinder | Verified (sync_bodies tested) |
 | SRT | double_buffer | any | cylinder | Expected to work (f-feq style) |
 
 **Known limitations (ddf_shifting):**
-- Verified configuration: **D2Q9 + MRT + double_buffer + zou_he_local** only
+- Verified configuration: **D2Q9 + MRT + double_buffer/zou_he_local** and **D2Q9 + MRT + esopull/zou_he_local** (sync_bodies add, remove, stepping stable)
 - `regularized` inlet with `ddf_shifting` is **known incompatible / unsolved** -- use `zou_he_local`
-- `esopull + ddf_shifting` has not been jointly validated
 - MRT shifts to physical space before collision, shifts back after (SRT/TRT are shift-invariant natively)
 - D3Q19 MRT shifting patch has a `compute_feq` inconsistency (not in scope for 2D-only)
 - Host `upload_ddf()` path is asymmetric (repaired)

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "CelerisLab"
-version = "0.4.0"
+version = "0.5.0"
 description = "GPU-accelerated Lattice Boltzmann Method (LBM) CFD solver using CUDA"
 readme = "README.md"
 requires-python = ">=3.8"

setup.py

@@ -5,7 +5,7 @@ with open("README.md", "r", encoding="utf-8") as fh:
 
 setup(
     name='CelerisLab',
-    version='0.4.0',
+    version='0.5.0',
     author='Frank14f',
     description='GPU-accelerated Lattice Boltzmann Method (LBM) CFD solver using CUDA',
     long_description=long_description,

src/CelerisLab/__init__.py

@@ -14,7 +14,7 @@ Usage::
     force = sim.read_force(0)
 """
 
-__version__ = "0.4.0"
+__version__ = "0.5.0"
 
 from . import common, cuda, lbm, body, config
 

src/CelerisLab/configs/CONFIG.md

@@ -28,7 +28,7 @@ Python `config.py` 只负责读取和校验，不是配置位置。
 | 字段 | 类型 | 默认 | 允许值 | 说明 |
 |------|------|------|--------|------|
 | `collision` | string | `"SRT"` | `SRT`, `TRT`, `MRT` | 碰撞算子 |
-| `streaming` | string | `"double_buffer"` | `double_buffer`, `esopull` | 流传输方式。运行时 body 拓扑同步 (``sync_bodies()``) 仅支持 ``double_buffer`` |
+| `streaming` | string | `"double_buffer"` | `double_buffer`, `esopull` | 流传输方式。运行时 body 拓扑同步 (``sync_bodies()``) 两种模式均支持 |
 | `store_precision` | string | `"FP32"` | `FP32`, `FP16S`, `FP16C` | GPU 存储精度。当前运行时已实现 `FP32` 与 `FP16S`，`FP16C` 仍为保留选项 |
 | `ddf_shifting` | bool | false | | 存储 f−w 而非 f，提升 FP16 精度 |
 | `les.enabled` | bool | false | | LES Smagorinsky 子格模型 |
@@ -93,8 +93,8 @@ step/one_step_*.cu         → kernel 编排
 
 ### 运行时 body 拓扑同步
 
-- 运行时增删 body（`add_body()` / `remove_body()` + `sync_bodies()`）仅支持 `streaming: "double_buffer"`。
-- `esopull` 模式下调用 `sync_bodies()` 会抛出 `NotImplementedError`。
+- 运行时增删 body（`add_body()` / `remove_body()` + `sync_bodies()`）在 ``double_buffer`` 和 ``esopull`` 两种流模式下均支持。
+- 用法相同，无需特别的运行时检查。
 
 ### 力区域标记（Force region flag）
 

src/CelerisLab/lbm/field.py

@@ -407,6 +407,57 @@ class LBMField:
     # directly from Simulation.sync_bodies().  They are no longer methods on
     # LBMField.
 
+    # -- EsoPull sync decode/encode (compact list) --------------------------
+    def esopull_sync_decode(self, patch_indices, step_count, stream=None):
+        """Decode a subset of cells from EsoPull backing to physical DDF.
+
+        Args:
+            patch_indices: uint32 array of cell indices to decode.
+            step_count: Current step count (for EsoPull parity).
+            stream: Optional CUDA stream.
+
+        Returns:
+            (scratch_gpu, scratch_host, idx_gpu): scratch buffer handles.
+        """
+        n = len(patch_indices)
+        nq = self.nq
+        scratch_gpu = cuda.mem_alloc(n * nq * 4)
+        scratch_host = np.empty(n * nq, dtype=np.float32)
+        idx_gpu = cuda.mem_alloc(n * 4)
+        cuda.memcpy_htod(idx_gpu, np.asarray(patch_indices, dtype=np.uint32))
+
+        fn = self.module.get_function("DecodeCellsToPhysical")
+        tpb = self.cfg.threads_per_block
+        fn(self.ddf_gpu, scratch_gpu, idx_gpu, np.uint32(n),
+           np.uint64(step_count),
+           block=(tpb, 1, 1), grid=((n + tpb - 1) // tpb, 1, 1),
+           stream=stream)
+
+        cuda.memcpy_dtoh_async(scratch_host, scratch_gpu, stream)
+        stream.synchronize()
+        return scratch_gpu, scratch_host, idx_gpu
+
+    def esopull_sync_encode(self, scratch_gpu, scratch_host, idx_gpu,
+                            n, step_count, stream=None):
+        """Upload patched physical DDF, collide + encode to EsoPull layout.
+
+        The backing layout stores post-collision values.  This method
+        applies collision to the patched physical DDF before encoding,
+        matching the store_f_esopull convention at parity *step_count*.
+        Frees scratch_gpu and idx_gpu after encoding.
+        """
+        cuda.memcpy_htod_async(scratch_gpu, scratch_host, stream)
+
+        fn = self.module.get_function("EncodePhysicalToCells")
+        tpb = self.cfg.threads_per_block
+        fn(self.ddf_gpu, scratch_gpu, idx_gpu, np.uint32(n),
+           np.uint64(step_count),
+           block=(tpb, 1, 1), grid=((n + tpb - 1) // tpb, 1, 1),
+           stream=stream)
+
+        scratch_gpu.free()
+        idx_gpu.free()
+
     def snapshot(self):
         self.download_ddf(force=True)
         self._ddf_snap = self.ddf.copy()

src/CelerisLab/lbm/kernels/kernel_v2.cu

@@ -69,5 +69,6 @@ extern "C"
 #include "step/one_step_esopull.cu"
 #include "step/aux_kernels.cu"
 #include "step/esopull_macro.cu"
+#include "step/esopull_sync.cu"
 
 }  // extern "C"

src/CelerisLab/lbm/kernels/step/esopull_sync.cu

@@ -0,0 +1,81 @@
+// CelerisLab -- step/esopull_sync.cu
+// Compact-list EsoPull decode/encode for runtime body topology sync.
+//
+// Decodes a subset of cells from EsoPull backing layout to physical DDF,
+// or encodes physical DDF back to EsoPull backing layout.
+// Each thread handles one cell from a compact index list.
+//
+// Included by kernel_v2.cu.  No standalone compilation.
+// ============================================================================
+
+#ifndef CELERIS_STEP_ESOPULL_SYNC_CU
+#define CELERIS_STEP_ESOPULL_SYNC_CU
+
+// ---------------------------------------------------------------------------
+// DecodeCellsToPhysical:  backing layout -> physical DDF (compact list)
+//
+// Uses ``t`` as the step count for EsoPull semantic parity decoding.
+// ---------------------------------------------------------------------------
+__global__ void DecodeCellsToPhysical(
+    const fpxx* fi,
+    float* fi_phys,
+    const unsigned int* cell_idx,
+    unsigned int n_cells,
+    unsigned long t)
+{
+    unsigned int tid = threadIdx.x + blockIdx.x * blockDim.x;
+    if (tid >= n_cells) return;
+
+    unsigned long k = (unsigned long)cell_idx[tid];
+    unsigned int x, y; coordinates(k, x, y);
+    unsigned long j[NQ]; compute_neighbors(x, y, j);
+
+    float f[NQ];
+    load_physical_node_esopull(k, fi, j, t, f);
+
+    // DDF-shifting: backing layout stores f-w, unshift to physical for host patch
+#if USE_DDF_SHIFTING
+    #pragma unroll
+    for (int i = 0; i < NQ; i++)
+        f[i] += (float)d_w[i];
+#endif
+
+    for (int i = 0; i < NQ; i++)
+        fi_phys[tid * NQ + i] = f[i];
+}
+
+// ---------------------------------------------------------------------------
+// EncodePhysicalToCells:  physical DDF -> backing layout (compact list)
+//
+// Reads physical (pre-collision) DDF from *fi_phys*, applies collision
+// to produce post-collision DDF, then encodes into the EsoPull backing
+// layout on *fi*.  The backing layout stores post-collision values,
+// matching what store_f_esopull would produce at step *t*.
+// ---------------------------------------------------------------------------
+__global__ void EncodePhysicalToCells(
+    fpxx* fi,
+    const float* fi_phys,
+    const unsigned int* cell_idx,
+    unsigned int n_cells,
+    unsigned long t)
+{
+    unsigned int tid = threadIdx.x + blockIdx.x * blockDim.x;
+    if (tid >= n_cells) return;
+
+    unsigned long k = (unsigned long)cell_idx[tid];
+    unsigned int x, y; coordinates(k, x, y);
+    unsigned long j[NQ]; compute_neighbors(x, y, j);
+
+    float f[NQ];
+    for (int i = 0; i < NQ; i++)
+        f[i] = fi_phys[tid * NQ + i];
+
+    // Collide: physical (pre-collision) -> post-collision
+    float rho_n, ux, uy;
+    compute_rho_u(f, rho_n, ux, uy);
+    collide_dispatch(f, rho_n, ux, uy);
+
+    store_physical_node_esopull(k, f, fi, j, t);
+}
+
+#endif // CELERIS_STEP_ESOPULL_SYNC_CU

src/CelerisLab/simulation.py

@@ -391,17 +391,9 @@ class Simulation:
     def sync_bodies(self) -> None:
         """Apply pending body edits (add/remove) to a running simulation.
 
-        This is the main entry point for runtime body topology changes.
-        It downloads the current DDF, rebuilds topology, recompiles the
-        kernel for the new object count, patches the DDF for geometry
-        changes, and re-uploads everything to GPU.
-
-        Currently only supports ``double_buffer`` streaming mode.
-        ``esopull`` mode raises ``NotImplementedError``.
-
-        Raises:
-            RuntimeError: If called before ``initialize()``.
-            NotImplementedError: If streaming mode is ``esopull``.
+        Supports both ``double_buffer`` and ``esopull`` streaming modes.
+        In EsoPull mode, the backing layout is decoded to physical DDF on
+        GPU for the patched cells only, patched on host, then encoded back.
         """
         if not self._initialized:
             raise RuntimeError("Call initialize() before sync_bodies()")
@@ -413,28 +405,34 @@ class Simulation:
         if not self.bodies.has_pending_edit():
             return
 
-        # 3. Check streaming mode
-        if self.lbm_cfg.streaming == "esopull":
-            raise NotImplementedError(
-                "Runtime body sync is not yet supported for esopull "
-                "streaming mode.  Use double_buffer instead.")
+        is_esopull = (self.lbm_cfg.streaming == "esopull")
+
+        # 3. Build sync plan (same for both modes)
+        plan = self.bodies.build_sync_plan(self.field)
+
+        # 4. Obtain physical DDF on host (mode-dependent path)
+        if is_esopull:
+            patch_set = _build_patch_set(
+                plan.added_solid_mask, plan.released_fluid_mask,
+                plan.curved_host[0], nx=self.field.nx, ny=self.field.ny)
+            scratch_gpu, scratch_host, idx_gpu = \
+                self.field.esopull_sync_decode(
+                    patch_set, self.stepper.step_count, self.stream)
+            _scatter_to_field(self.field, scratch_host, patch_set)
+        else:
+            self.field.download_ddf(
+                step_id=self.stepper.step_count, force=True)
 
-        # 4. Download current DDF to host (snapshot for patching)
-        self.field.download_ddf(
-            step_id=self.stepper.step_count, force=True)
         old_ddf = self.field.ddf.copy()
         old_flags = self.field.flag.copy()
 
-        # 5. Build sync plan
-        plan = self.bodies.build_sync_plan(self.field)
-
-        # 6. Runtime recompile for new object count
+        # 5. Runtime recompile for new object count
         self._recompile(plan.next_count)
 
-        # 7. Apply new topology data (flags, compact lists, params)
+        # 6. Apply new topology data (flags, compact lists, params)
         self.bodies.apply_sync_plan(self.field, plan)
 
-        # 8. DDF patch for geometry changes
+        # 7. DDF patch for geometry changes (works on physical DDF, mode-agnostic)
         cl_fluid_idx = plan.curved_host[0]
         patch_ddf_for_body_sync(
             self.field,
@@ -442,11 +440,20 @@ class Simulation:
             plan.added_solid_mask, plan.released_fluid_mask,
             curved_fluid_indices=cl_fluid_idx,
         )
-        upload_patched_ddf(
-            self.field,
-            plan.added_solid_mask, plan.released_fluid_mask,
-            curved_fluid_indices=cl_fluid_idx,
-        )
+
+        # 8. Upload patched DDF (mode-dependent path)
+        if is_esopull:
+            _gather_from_field(self.field, scratch_host, patch_set)
+            n = len(patch_set)
+            self.field.esopull_sync_encode(
+                scratch_gpu, scratch_host, idx_gpu,
+                n, self.stepper.step_count, self.stream)
+        else:
+            upload_patched_ddf(
+                self.field,
+                plan.added_solid_mask, plan.released_fluid_mask,
+                curved_fluid_indices=cl_fluid_idx,
+            )
 
         # 9. Commit formal objects and clear pending
         self.bodies.commit_pending(plan.next_objects, plan.new_sensor_counts)
@@ -649,3 +656,51 @@ class Simulation:
         """Validate object-count and obs layout contracts before stepping."""
         self._assert_object_count_contract()
         self._assert_obs_layout_contract()
+
+
+# -- Module-level helpers -------------------------------------------------------
+
+def _build_patch_set(added_solid_mask, released_fluid_mask, curved_fluid_idx,
+                     nx=None, ny=None):
+    """Return sorted uint32 array of cell indices that need DDF patching.
+
+    Includes a 1-cell halo because ``store_physical_node_esopull`` writes
+    to neighbor slots (D2Q9 directions 1,3,5,7).  Neighbors must be
+    decoded/encoded together to avoid corrupting the backing layout.
+    """
+    s = set()
+    s.update(np.where(added_solid_mask)[0].tolist())
+    s.update(np.where(released_fluid_mask)[0].tolist())
+    if curved_fluid_idx is not None:
+        s.update(curved_fluid_idx.tolist())
+
+    if nx is None or ny is None:
+        return np.array(sorted(s), dtype=np.uint32)
+
+    # Expand to include 1-cell halo (D2Q9 Moore neighbourhood)
+    expanded = set(s)
+    for idx in s:
+        x = idx % nx
+        y = idx // nx
+        for dx in (-1, 0, 1):
+            for dy in (-1, 0, 1):
+                if dx == 0 and dy == 0:
+                    continue
+                xn, yn = x + dx, y + dy
+                if 0 <= xn < nx and 0 <= yn < ny:
+                    expanded.add(xn + yn * nx)
+    return np.array(sorted(expanded), dtype=np.uint32)
+
+
+def _scatter_to_field(field, scratch_host, patch_indices):
+    """Copy decoded physical DDF from scratch buffer into field.ddf."""
+    nq = field.nq
+    for j, idx in enumerate(patch_indices):
+        field.ddf[idx * nq:(idx + 1) * nq] = scratch_host[j * nq:(j + 1) * nq]
+
+
+def _gather_from_field(field, scratch_host, patch_indices):
+    """Copy patched physical DDF from field.ddf back to scratch buffer."""
+    nq = field.nq
+    for j, idx in enumerate(patch_indices):
+        scratch_host[j * nq:(j + 1) * nq] = field.ddf[idx * nq:(idx + 1) * nq]

tests/integration/test_sync_bodies_skeleton.py

@@ -1,14 +1,17 @@
-"""sync_bodies pipeline without DDF patch — recompile, esopull guard, step count preservation.
+"""sync_bodies pipeline — recompile, esopull runtime body sync, step count preservation.
 
 Requires GPU."""
 
 import unittest
+import tempfile
+import os
+import json
 
 import numpy as np
 import pycuda.driver as cuda
 import pycuda.autoinit
 
-from CelerisLab.simulation import Simulation
+from CelerisLab.simulation import Simulation, _scatter_to_field, _gather_from_field
 
 
 # Use a small grid for fast compilation and test execution
@@ -16,7 +19,7 @@ NX, NY = 128, 64
 
 
 class TestSyncBodiesSkeleton(unittest.TestCase):
-    """Test sync_bodies() with real GPU -- skeleton without DDF patch."""
+    """Test sync_bodies() with real GPU -- double_buffer path."""
 
     def _make_sim(self) -> Simulation:
         """Create a Simulation with a small double_buffer D2Q9 grid."""
@@ -125,40 +128,118 @@ class TestSyncBodiesSkeleton(unittest.TestCase):
         self.assertEqual(sim.bodies.count, 1)
         sim.close()
 
-    def test_esopull_raises_not_implemented(self):
-        """sync_bodies() with esopull should raise NotImplementedError."""
-        # Create a sim with esopull streaming
-        from CelerisLab.config import load_lbm_config
-        cfg = load_lbm_config()
-        cfg.streaming = "esopull"
-        # We need to build the sim manually to override streaming
-        sim = Simulation.__new__(Simulation)
-        sim._stream = None
-        sim.lbm_cfg = cfg
-        from CelerisLab.config import BodyConfig
-        sim.body_cfg = BodyConfig()
-        from CelerisLab.cuda.context import CudaContext
-        sim.ctx = CudaContext(0)
-        from CelerisLab.cuda import compiler_v2 as compiler
-        arch = sim._resolve_compile_arch = lambda: sim.ctx.sm_arch
-        arch_val = CudaContext(0).sm_arch
-        compiler.generate_config(cfg, n_objects=0)
-        ptx_path = compiler.compile_kernel(arch=arch_val)
-        module = compiler.load_module(ptx_path)
-        sim._ptx_path = ptx_path
-        sim._module = module
-        from CelerisLab.lbm.field import LBMField
-        sim.field = LBMField(cfg, module)
-        from CelerisLab.lbm.stepper import LBMStepper
-        sim.stepper = LBMStepper(sim.field, module, cfg)
-        from CelerisLab.body.manager import ObjectManager
-        sim.bodies = ObjectManager(
-            cfg.nx, cfg.ny, cfg.nz, cfg.nq, cfg)
-        sim._initialized = True
+
+class TestSyncBodiesEsoPull(unittest.TestCase):
+    """Test sync_bodies() with esopull streaming."""
+
+    def _make_esopull_sim(self, nx=128, ny=64) -> Simulation:
+        """Create a Simulation with a small esopull D2Q9 grid."""
+        cfg = {
+            "grid": {"lattice_model": "D2Q9", "nx": nx, "ny": ny, "nz": 1},
+                "physics": {"data_type": "FP32", "viscosity": 0.05,
+                            "velocity": 0.03, "rho": 1.0},
+            "method": {
+                "collision": "SRT", "streaming": "esopull",
+                "store_precision": "FP32", "ddf_shifting": False,
+                "les": {"enabled": False, "cs": 0.16, "closed_form": True},
+                "trt": {"magic_param": 0.1875},
+                "inlet": {"profile": "parabolic", "scheme": "zou_he_local"},
+                "outlet": {"mode": "neq_extrap", "backflow_clamp": True,
+                           "blend_alpha": 0.7},
+                "y_wall_bc": "bounce_back",
+                "omega_guard": {"min": 0.01, "max": 1.99},
+            },
+            "cuda": {"threads_per_block": 256, "compute_capability": "auto"},
+        }
+        tmpd = tempfile.mkdtemp(prefix="esopull_sync_")
+        lbm_path = os.path.join(tmpd, "config_lbm.json")
+        with open(lbm_path, "w") as f:
+            json.dump(cfg, f)
+        return Simulation(lbm_config_path=lbm_path)
+
+    def test_esopull_add_body(self):
+        """EsoPull: add body, sync, run, read finite force."""
+        sim = self._make_esopull_sim()
+        sim.initialize()
+        sim.run(100)
 
         sim.add_body("circle", center=(NX // 4, NY // 2), radius=8)
-        with self.assertRaises(NotImplementedError):
-            sim.sync_bodies()
+        sim.sync_bodies()
+        self.assertEqual(sim.bodies.count, 1)
+
+        sim.run(100)
+        force = sim.read_force(0)
+        self.assertTrue(np.all(np.isfinite(force)),
+                        f"EsoPull sync: finite force expected, got {force}")
+        sim.close()
+
+    def test_esopull_remove_body(self):
+        """EsoPull: add body → run → remove → sync → run."""
+        sim = self._make_esopull_sim()
+        sim.add_body("circle", center=(NX // 4, NY // 2), radius=8)
+        sim.initialize()
+        sim.run(100)
+
+        sim.remove_body(0)
+        sim.sync_bodies()
+        self.assertEqual(sim.bodies.count, 0)
+
+        sim.run(100)
+        macro = sim.get_macroscopic()
+        self.assertTrue(np.all(np.isfinite(macro["ux"])),
+                        "Macroscopic ux should be finite after esopull sync")
+        sim.close()
+
+    def test_esopull_add_remove_add(self):
+        """EsoPull: add → remove → add cycle with finite checks."""
+        sim = self._make_esopull_sim()
+        sim.initialize()
+        sim.run(100)
+
+        # Add
+        sim.add_body("circle", center=(NX // 4, NY // 2), radius=8)
+        sim.sync_bodies()
+        sim.run(100)
+
+        # Remove
+        sim.remove_body(0)
+        sim.sync_bodies()
+        self.assertEqual(sim.bodies.count, 0)
+        sim.run(100)
+
+        # Add again
+        sim.add_body("circle", center=(NX // 2, NY // 2), radius=6)
+        sim.sync_bodies()
+        self.assertEqual(sim.bodies.count, 1)
+        sim.run(100)
+        force = sim.read_force(0)
+        self.assertTrue(np.all(np.isfinite(force)),
+                        f"Finite force after esopull add-remove-add: {force}")
+        sim.close()
+
+    def test_esopull_roundtrip_fullgrid(self):
+        """EsoPull: decode/encode roundtrip on full grid (no patch)."""
+        sim = self._make_esopull_sim()
+        sim.initialize()
+        sim.run(100)
+
+        nx, ny, nq = sim.field.nx, sim.field.ny, sim.field.nq
+        all_cells = np.arange(nx * ny, dtype=np.uint32)
+        step = sim.stepper.step_count
+        stream = sim.stream
+
+        s_gpu, s_host, idx_gpu = sim.field.esopull_sync_decode(
+            all_cells, step, stream)
+        _scatter_to_field(sim.field, s_host, all_cells)
+        _gather_from_field(sim.field, s_host, all_cells)
+        sim.field.esopull_sync_encode(
+            s_gpu, s_host, idx_gpu, len(all_cells), step, stream)
+        stream.synchronize()
+
+        sim.run(100)
+        macro = sim.get_macroscopic()
+        self.assertTrue(np.all(np.isfinite(macro["ux"])),
+                        "ux finite after full-grid roundtrip")
         sim.close()