diff --git a/CHANGELOG.md b/CHANGELOG.md
index ddb511c..4205cb3 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,28 @@
 # Changelog
 
+## 0.4.0 (2026-10-22)
+
+cellmap-analyze's three asks, so it can call worldarray directly with no wrapper, and two
+of cellmap-flow's. No placement change and no unit table change.
+
+- `vol.read_onto(grid, begin=None, end=None, ...)` reads voxels `[begin, end)` of a
+  `Grid`'s spatial axes from the array, exactly that shape (Behaviour 20). It is the
+  `read` of the grid's index box at the grid's voxel size anchored at its corner, so at
+  the array's own voxel size nothing is resampled, whatever the sub-voxel offset, and
+  otherwise the array is resampled onto the grid. The grid's units are converted.
+- A pickled handle, and `reopen`, use the path resolved against the working directory at
+  open, so a worker that changes directory before unpickling opens the same store; it
+  used to raise `FileNotFoundError` for a relative path.
+- `write_index`, `write` and `write_chunk` take a scalar, which fills the box. It is
+  converted to the dtype only when its value survives (`0` into `uint8`; `300`, `-1`, `0.5`
+  and NaN raise), and the fill value over whole units leaves them absent.
+- `vol.source_translation`: the stored level's translation in `units`, beside
+  `source_voxel_size`, under resample and relabel too (cellmap-flow's `stored_placement`).
+- `create(exists="adopt")` also compares the keys of `attrs` with the existing array's
+  attributes and raises naming each that differs or is missing; keys not given are not
+  compared. An adopt given different `attrs` used to keep the old ones silently, as funlib's
+  `prepare_ds` kept an old offset (cellmap-flow's `_refuse_another_region`).
+
 ## 0.3.1 (2026-10-22)
 
 - `source_voxel_size` reports the stored level's voxel size under `on_mismatch="relabel"`,
diff --git a/README.md b/README.md
index 6b02152..556360c 100644
--- a/README.md
+++ b/README.md
@@ -50,6 +50,7 @@ six = wa.open(path, voxel_size=(6, 6, 6), on_mismatch="resample")
 
 # one volume read onto another volume's grid, for masks, contact sites, statistics
 onto = raw.read(box, voxel_size=labels.voxel_size, anchor=labels.corner)
+onto = raw.read_onto(grid, begin, end)   # voxels [begin, end) of a Grid, exactly that shape
 
 # index space, native units: what an engine or a viewer wants
 vol = wa.open(path, units=None)
@@ -100,8 +101,8 @@ s1 = wa.open_pyramid("out.zarr", mode="r+").add_level("s1", factors=2)   # place
 - **Plans outputs and blocks** (0.3): a store-less `Grid` with the same geometry as an
   array, so an output is planned, cut into blocks and only then created; `blocks` tiling an
   array or a box into write boxes with context-grown read boxes, in C order, as plain
-  picklable records a cluster job takes by id; the anisotropy-aware pyramid schedule.
-  Nothing is scheduled or run.
+  picklable records a cluster job takes by id; the anisotropy-aware pyramid schedule; every
+  input read onto that grid by its voxel indices (0.4). Nothing is scheduled or run.
 
 ## The placement rule
 
@@ -168,12 +169,12 @@ wa.units.abbreviation("nanometer") ; wa.store.join(path, "s0") ; wa.store.parent
 ```
 
 `WorldArray`: `ts`, `meta`, `grid`, `shape`, `dtype`, `axes`, `voxel_size`, `translation`,
-`corner`, `box`, `levels`, `level_name`, `level_path`, `source_voxel_size`, `resampled`,
+`corner`, `box`, `levels`, `level_name`, `level_path`, `source_voxel_size`, `source_translation`, `resampled`,
 `writable`, `write_shape`;
 `read(box, voxel_size=, padding=, method=, mask=, anchor=, transform=, post_transform=, select=, strict=)`,
-`read_index(begin, end, padding=, mask=, select=, transform=)`, `index_box(box)`,
+`read_index(begin, end, padding=, mask=, select=, transform=)`, `read_onto(grid, begin=, end=, ...)`, `index_box(box)`,
 `chunks_covering(begin, end, write=)`, `list_chunks()`, `chunk_key(index)`, `parse_chunk_key(key)`,
-`write_index(begin, end, data, partial=)`, `write(box, data, partial=)`, `reopen(context=)`.
+`write_index(begin, end, data, partial=)` and `write(box, data, partial=)` (`data` an array or a scalar), `reopen(context=)`.
 
 A box is a `(begin, end)` pair of floats in your units, half-open per axis, over the
 spatial axes or every axis; any object with `.begin`/`.end` or `.offset`/`.shape` also
diff --git a/docs/spec.md b/docs/spec.md
index 5744526..c656faa 100644
--- a/docs/spec.md
+++ b/docs/spec.md
@@ -149,8 +149,9 @@ vol.meta           # ArrayMeta in the store's own units
 vol.grid           # ArrayMeta in the requested units, every axis
 vol.shape, vol.dtype, vol.axes, vol.spatial, vol.level_name, vol.levels, vol.remote
                    # shape is the stored shape, or with a resample target the target voxels covering the data
-vol.source_voxel_size, vol.resampled   # the stored level's voxel size, also under resample and relabel; whether reads
-                                       #   resample to a target
+vol.source_voxel_size, vol.source_translation, vol.resampled
+                   # the stored level's voxel size and translation (0.4), also under resample and relabel; whether
+                   #   reads resample to a target
 vol.list_chunks()  # the chunk indices that exist in storage, by listing keys (zarr v2/v3, N5, unsharded precomputed)
 vol.path           # what the caller passed (a group, say); vol.level_path is the array actually opened
 vol.voxel_size     # tuple[float] in `units`, spatial axes, the array's order (zyx for a volume)
@@ -285,7 +286,8 @@ vol = wa.create(
     attrs=None,           # extra array attributes, merged into the array's own
     group_attrs=None,     # extra group attributes, merged; no one else's key is dropped or rewritten (Behaviour 12)
     exists="error",       # "error": anything already at `path` raises | "adopt": an identical array (shape, dtype, chunks,
-                          #   shards, placement within rtol 1e-6) is opened for writing; a different one raises naming each
+                          #   shards, placement within rtol 1e-6, and from 0.4 every key of `attrs` holding the value given)
+                          #   is opened for writing; a different one raises naming each
                                                     #   difference. There is no third value. An adopt writes no group document when the entry
                           #   already says the same (Behaviour 12)
     store_fill=False,     # True stores write units equal to fill_value instead of leaving them absent
@@ -337,6 +339,20 @@ wa.store.join(path, *parts) ; wa.store.parent(path) ; wa.store.basename(path)
 wa.units.abbreviation(unit) -> str   # the OME short form, "nm" for any spelling of nanometre the table knows
 ```
 
+### 0.4: reading onto a grid
+
+```python
+vol.read_onto(grid, begin=None, end=None, padding=None, method=None, mask=False, retries=None, transform=None,
+              post_transform=None, select=None)
+                          # voxels [begin, end) of a Grid's spatial axes (all of it when None), read from this array:
+                          #   exactly that shape, by the centre rule at the array's voxel size, else resampled onto the
+                          #   grid anchored at its corner (Behaviour 20)
+vol.write_index(begin, end, 0) ; vol.write(box, 0) ; vol.write_chunk(index, 0)
+                          # a scalar fills the box, converted to the dtype only when its value survives (Behaviour 13)
+vol.source_translation    # the stored level's translation in `units`, beside source_voxel_size, under resample and relabel
+wa.create(..., attrs={...}, exists="adopt")   # an adopt also compares the keys of `attrs` (Behaviour 11)
+```
+
 ## 3. Conventions
 
 These are the rules the package exists to hold in one place. They match chunkmirage's
@@ -682,9 +698,11 @@ Behaviours 10 to 16 are 0.2; 17 to 19 are 0.3.
     compression; cellmap-flow's blosc stays blosc by passing it. There is no overwrite: with
     `exists="error"` any document already at `path` raises `FileExistsError`; with
     `exists="adopt"` an array whose shape, dtype, chunks, shards and placement match
-    (placement within `rtol 1e-6`) is opened for writing and its dataset entry ensured, and
-    one that differs raises `ValueError` listing every differing field with the value found
-    and the value wanted (cellmap-flow's `_refuse_another_region`, registration-station's
+    (placement within `rtol 1e-6`), and whose attributes hold every key of `attrs` with the
+    value given as JSON stores it (keys not given are someone else's and not compared), is
+    opened for writing and its dataset entry ensured, and one that differs raises
+    `ValueError` listing every differing field and key with the value found and the value
+    wanted (cellmap-flow's `_refuse_another_region`, registration-station's
     missing check, which let a shorter last band's chunk grid win a creation race). A
     creation that loses a race, the store reporting the array present after the probe said
     absent, is re-read with backoff, 5 attempts at `0.5 s * attempt`, and then treated as an
@@ -697,7 +715,9 @@ Behaviours 10 to 16 are 0.2; 17 to 19 are 0.3.
     being the one it replaces. `create` returns a writable `WorldArray`;
     `wa.open(path, mode="r+")` opens an existing one with every rule of Conventions 3 to 5
     unchanged, and a handle pickles by its arguments and its context's spec, so each worker
-    process opens its own cheaply, on the shared context when the handle had one, else on
+    process opens its own cheaply (a local path goes resolved against the working directory
+    at open, so a worker that changes directory before unpickling, as cellmap-analyze's dask
+    workers do, opens the same store; `reopen` uses it too), on the shared context when the handle had one, else on
     a context rebuilt once per process from the same spec, so a capped context stays
     capped (mesh-n-bone's per-process cache and thread caps).
 
@@ -722,7 +742,12 @@ Behaviours 10 to 16 are 0.2; 17 to 19 are 0.3.
 
 13. **Aligned writes need no coordination.** `write_index(begin, end, data)` requires
     `data.shape == end - begin`, `data.dtype == vol.dtype` (no silent cast; `ValueError`),
-    and `[begin, end)` to be a union of whole write units: every bound a multiple of
+    or `data` a scalar (a number or a 0-d array) that fills the box: it is converted to the
+    dtype only when its value survives, so an integer or bool array takes whole numbers in
+    its range and a float array any number it does not overflow, rounded to its precision
+    (`0` into `uint8` is written; `300`, `-1`, `0.5` and NaN raise), and a scalar equal to
+    `fill_value` over whole units leaves them absent as an all-fill array does (cellmap-analyze
+    writes a constant for every empty or masked-out block). The box must be a union of whole write units: every bound a multiple of
     `write_shape` on its axis or the array's end on that axis, so the clipped last unit is
     allowed; otherwise `ValueError` names the axis, the bound and the unit size. A unit
     written whole is one put, so writers in different processes writing different units
@@ -809,6 +834,22 @@ Behaviours 10 to 16 are 0.2; 17 to 19 are 0.3.
     returns the OME short form for any spelling the table knows and raises for one it does
     not.
 
+20. **Reading onto a grid.** `vol.read_onto(grid, begin, end)` returns voxels `[begin, end)`
+    of a `Grid`'s spatial axes, all of it when `begin` and `end` are None, read from this
+    array, with exactly `end - begin` voxels on those axes; the array's other axes are read
+    as `read` reads them (whole, or pinned by `select`), the open-time `padding`, `method`,
+    `retries` and `transform` apply, and `mask` marks the grid voxels that held data. It is
+    `read(Box.from_index(begin, end, grid voxel size, grid translation), voxel_size=grid
+    voxel size, anchor=grid corner)`, so Behaviours 2 and 4 decide every voxel: at the
+    array's own voxel size (within level selection's rtol 1e-6) nothing is resampled,
+    whatever the offset between the grids, and each grid voxel takes the stored voxel whose
+    centre lies in it, never an interpolation, with `method="linear"` too; at any other
+    voxel size the array is resampled onto the grid, anchored at its corner. Adjoining
+    `[begin, end)` reads therefore agree with one read across both. The grid's spatial axes
+    must be named as the array's are (`ValueError` otherwise), and its units are converted
+    to the array's `units`. This is cellmap-analyze's per-operation grid: blocks come from
+    one `Grid`, every input is read onto it, and outputs are created on it.
+
 ## 5. Tests (behavioural; synthetic stores in `tmp_path`; one optional network test)
 
 Conformance, one tiny store each, checking `voxel_size`, `translation`, `axes`,
@@ -898,8 +939,10 @@ factor; missing ancestor documents down from the `.zarr` root and none above it,
 ancestor's format followed and a contradiction refused, an array in the way refused; an
 adopt writing no document and a changed placement writing one; a lost entry written again
 and the error after the attempts; `ensure_group_entry` on a foreign array, idempotent; a
-pickled handle keeping a capped context's spec and sharing the shared one; `codecs_of` by
-path; an all-fill write removing its own key.
+pickled handle keeping a capped context's spec and sharing the shared one, and a relative
+path reopening after a `chdir`; `codecs_of` by path; an all-fill write removing its own key;
+a scalar filling a box through `write_index`, `write` and `write_chunk`, refused when it
+does not fit the dtype.
 
 Grids (0.3): `blocks` partitioning with and without a box, `context` as an int, per axis
 and as before/after pairs, clipping on and off, `len`, `blocks[id]` and `index_of`;
@@ -908,6 +951,14 @@ and as before/after pairs, clipping on and off, `len`, `blocks[id]` and `index_o
 reproducing mesh-n-bone's anisotropic table (8, 8, 30 nm: z stays at 1 until y and x
 reach 32) and all 2s for isotropic data; `abbreviation` for every spelling in the table.
 
+Reading onto a grid (0.4): `read_onto` at the array's voxel size returning stored values
+under a sub-voxel offset with `method="linear"`, and the next voxel past a half-voxel one;
+onto a finer grid equal to the `read` it is defined as, in micrometres too, adjoining
+reads agreeing, the mask marking grid voxels before the data; its own grid at voxel sizes
+3.54, 4 and 5.3 returning the stored array exactly; a grid with other axis names refused.
+`source_translation` under relabel and resample and in micrometres; an adopt keeping keys
+it was not given and refusing a different or missing one, naming both values.
+
 Target: about 600 lines of test cases for 0.1 and about 300 more for 0.2 and 0.3, the
 store builders in `conftest.py` aside; no wording pins; no `/nrs`; no cluster.
 
diff --git a/pyproject.toml b/pyproject.toml
index 6b81ce4..dbf4ea0 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "worldarray"
-version = "0.3.1"
+version = "0.4.0"
 description = "Read zarr, N5 and Neuroglancer precomputed volumes in world coordinates through tensorstore: correct voxel placement, pyramid levels, padded region reads, resampling. numpy + tensorstore only."
 readme = "README.md"
 license = { file = "LICENSE" }
diff --git a/src/worldarray/__init__.py b/src/worldarray/__init__.py
index 350f76b..37e8d67 100644
--- a/src/worldarray/__init__.py
+++ b/src/worldarray/__init__.py
@@ -28,7 +28,7 @@
     parse_chunk_key,
 )
 
-__version__ = "0.3.1"
+__version__ = "0.4.0"
 
 __all__ = [
     "ArrayMeta",
diff --git a/src/worldarray/array.py b/src/worldarray/array.py
index 5988f54..cff63c7 100644
--- a/src/worldarray/array.py
+++ b/src/worldarray/array.py
@@ -36,7 +36,15 @@
     read_meta_light,
 )
 from .retry import Retry, read_with_retries
-from .store import Node, detect, is_remote, join, open_array, precomputed_scale_index
+from .store import (
+    Node,
+    detect,
+    is_remote,
+    join,
+    normalize_path,
+    open_array,
+    precomputed_scale_index,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -90,6 +98,9 @@ def __init__(
         if mode not in ("r", "r+"):
             raise ValueError("mode is 'r' (read) or 'r+' (read and write)")
         self.path = path
+        # where later opens find it (reopen, a pickle, the pyramid): a local path resolved
+        # against the working directory now, so a worker that changes directory opens the same store
+        self._where = normalize_path(path)
         self._mode = mode
         self._store_fill = bool(store_fill)
         self.padding = padding
@@ -139,8 +150,9 @@ def __init__(
             self._meta = self._rescaled_to_base(base_voxel_size, base_units)
         space, time = _unit_targets(units)
         self._grid = self._meta.in_units(space=space, time=time, strict=strict_units)
-        # the stored level's spatial voxel size, before any relabel renames the grid
+        # the stored level's spatial placement, before any relabel renames the grid
         self._stored_voxel_size = tuple(self._grid.voxel_size[i] for i in self._grid.spatial)
+        self._stored_translation = tuple(self._grid.translation[i] for i in self._grid.spatial)
 
         self._target: tuple[float, ...] | None = None
         if voxel_size is not None:
@@ -305,6 +317,13 @@ def source_voxel_size(self) -> tuple[float, ...]:
         is, whether reads resample it or ``on_mismatch="relabel"`` renames its grid."""
         return self._stored_voxel_size
 
+    @property
+    def source_translation(self) -> tuple[float, ...]:
+        """The stored level's translation (the centre of its voxel 0) on the spatial axes, in
+        ``units``: where the level is, whether reads resample it or ``on_mismatch="relabel"``
+        renames its grid."""
+        return self._stored_translation
+
     @property
     def resampled(self) -> bool:
         """Whether reads resample to a target voxel size (``on_mismatch="resample"``)."""
@@ -373,7 +392,7 @@ def levels(self) -> list[tuple[str, tuple[float, ...]]]:
     def levels_pyramid(self) -> Pyramid:
         if self._pyramid is None:
             self._pyramid = Pyramid(
-                self.path, context=self.context, units=self.units, select=self._select
+                self._where, context=self.context, units=self.units, select=self._select
             )
         return self._pyramid
 
@@ -422,14 +441,15 @@ def reopen(self, *, context: ts.Context | None = None, **overrides) -> WorldArra
             **overrides,
         }
         return WorldArray(
-            self.path, context=context if context is not None else self.context, **opts
+            self._where, context=context if context is not None else self.context, **opts
         )
 
     def __reduce__(self):
         """A handle pickles by its arguments and its context's recipe, and reopens in the
         receiving process (Behaviour 11): a worker opens its own store, cheaply, on a shared
         context when the handle had one, else on a context rebuilt from the same spec, so a
-        capped context stays capped."""
+        capped context stays capped. A local path goes as resolved at open, so a worker that
+        changed directory still finds the store."""
         opts = {
             **self._opts,
             "padding": self.padding,
@@ -439,7 +459,7 @@ def __reduce__(self):
             "retries": self.retries,
             "transform": self.transform,
         }
-        return (_reopen, (self.path, opts, recipe_of(self.context)))
+        return (_reopen, (self._where, opts, recipe_of(self.context)))
 
     # --- index space ------------------------------------------------------------
 
@@ -520,7 +540,9 @@ def write_index(
         or the array's end), so writers in different processes need no coordination; a
         unit written whole is one put. ``partial=True`` lifts the check and does tensorstore's
         read-modify-write of the touched units, unsafe while anything else writes the same
-        unit. No cast, no padding, no resampling; the write commits before returning."""
+        unit. ``data`` may be a scalar, which fills the box: it is converted to the array's
+        dtype only when the value survives (``0`` into ``uint8``, not ``300`` or ``0.5``).
+        No other cast, no padding, no resampling; the write commits before returning."""
         if not self.writable:
             raise ValueError(
                 f"{self.path} was opened read-only; open it with mode='r+' or create it"
@@ -535,9 +557,11 @@ def write_index(
         if len(begin) != len(shape) or len(end) != len(shape):
             raise ValueError(f"{self.path}: begin and end need one entry per axis of {shape}")
         wanted = tuple(e - b for b, e in zip(begin, end))
-        if data.shape != wanted:
+        if data.ndim == 0:
+            data = self._scalar(data)
+        elif data.shape != wanted:
             raise ValueError(f"{self.path}: data has shape {data.shape}; the box is {wanted}")
-        if data.dtype != self.dtype:
+        elif data.dtype != self.dtype:
             raise ValueError(
                 f"{self.path}: data is {data.dtype}, the array {self.dtype}; nothing is cast on write"
             )
@@ -555,6 +579,22 @@ def write_index(
                     )
         self._store[tuple(slice(b, e) for b, e in zip(begin, end))].write(data).result()
 
+    def _scalar(self, value: np.ndarray) -> np.ndarray:
+        """A 0-d ``value`` in the array's dtype, or ``ValueError`` when converting changes it:
+        an integer or bool array takes only whole numbers in its range, a float array any
+        number that does not overflow it (rounded to its precision)."""
+        dtype = self.dtype
+        with np.errstate(invalid="ignore", over="ignore"), warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            out = value.astype(dtype)
+            if dtype.kind in "biu":
+                kept = bool(out == value)
+            else:
+                kept = bool(np.isfinite(out) or not np.isfinite(value))
+        if not kept:
+            raise ValueError(f"{self.path}: {value.item()!r} does not fit {dtype}")
+        return out
+
     def write(self, box, data: np.ndarray, partial: bool = False) -> None:
         """Write ``data`` over a world box: the voxels Behaviour 2's centre rule selects, then
         ``write_index``. A box reaching outside the array raises; nothing is padded or
@@ -880,6 +920,57 @@ def read_index(b, e, pad):
                 data[~inside] = padding
         return _squeeze((data, inside) if mask else data, drop)
 
+    def read_onto(
+        self,
+        grid,
+        begin: Sequence[int] | None = None,
+        end: Sequence[int] | None = None,
+        padding: Any = None,
+        method: str | None = None,
+        mask: bool = False,
+        retries: Retry | None = None,
+        transform: Transform | None = None,
+        post_transform: Transform | None = None,
+        select: dict[str, int] | None = None,
+    ) -> np.ndarray | tuple[np.ndarray, np.ndarray]:
+        """Voxels ``[begin, end)`` of another grid (a ``Grid``), read from this array
+        (Behaviour 20): ``begin`` and ``end`` index the grid's spatial axes, all of it when
+        None, and the result has exactly that shape there; other axes are read as ``read``
+        reads them. At the array's own voxel size nothing is resampled, however the grids are
+        offset: each grid voxel takes the stored voxel whose centre lies in it. Otherwise the
+        array is resampled onto the grid by ``method``, anchored at the grid's corner. The
+        grid's units are converted to ``units``; its spatial axes must be named as this
+        array's are."""
+        sp, mine = grid.spatial, self._grid.spatial
+        names = tuple(grid.axes[i] for i in sp)
+        if names != tuple(self._grid.axes[i] for i in mine):
+            raise ValueError(
+                f"{self.path}: the grid's spatial axes are {names}; the array's "
+                f"{tuple(self._grid.axes[i] for i in mine)}"
+            )
+        vs, tr = [], []
+        for i, j in zip(sp, mine):
+            have, want = grid.units[i], self._grid.units[j]
+            f = 1.0 if have == want else U.factor(have, want)
+            vs.append(grid.voxel_size[i] * f)
+            tr.append(grid.translation[i] * f)
+        begin = (0,) * len(sp) if begin is None else tuple(int(b) for b in begin)
+        end = tuple(grid.shape[i] for i in sp) if end is None else tuple(int(e) for e in end)
+        if len(begin) != len(sp) or len(end) != len(sp):
+            raise ValueError(f"begin and end need one entry per spatial axis of the grid {names}")
+        return self.read(
+            Box.from_index(begin, end, vs, tr),
+            voxel_size=vs,
+            anchor=tuple(t - v / 2 for t, v in zip(tr, vs)),
+            padding=padding,
+            method=method,
+            mask=mask,
+            retries=retries,
+            transform=transform,
+            post_transform=post_transform,
+            select=select,
+        )
+
 
 def _squeeze(out, drop: tuple[int, ...]):
     if not drop:
@@ -1123,8 +1214,8 @@ def create(
     0.5) following the group's format (Behaviours 11 and 12). ``chunks`` is required.
     Nothing is ever deleted or overwritten: with ``exists="error"`` anything at ``path``
     raises ``FileExistsError``; with ``"adopt"`` an identical array (shape, dtype, chunks,
-    shards, placement) is opened for writing and a different one raises ``ValueError``
-    naming every difference. Returns a writable ``WorldArray``."""
+    shards, placement, and every key of ``attrs`` already holding the value given) is opened
+    for writing and a different one raises ``ValueError`` naming every difference. Returns a writable ``WorldArray``."""
     if grid is not None:
         # a Grid planned the output (Behaviour 18): its facts are the array's, chunks and
         # shards overridable; the placement arguments must not also be given
@@ -1165,6 +1256,7 @@ def adopt(normalized: str, fmt: str) -> WorldArray:
             shards=shards,
             voxel_size=vs,
             translation=tr,
+            attrs=attrs,
         )
         if diffs:
             raise ValueError(
diff --git a/src/worldarray/write.py b/src/worldarray/write.py
index d3e5884..f7fa190 100644
--- a/src/worldarray/write.py
+++ b/src/worldarray/write.py
@@ -803,9 +803,12 @@ def differences(
     shards: Sequence[int] | None,
     voxel_size: Sequence[float],
     translation: Sequence[float],
+    attrs: dict | None = None,
 ) -> list[str]:
     """How an existing array's ``ArrayMeta`` differs from what a caller wanted to create:
-    one line per field, empty when identical (placement within ``RTOL``)."""
+    one line per field, empty when identical (placement within ``RTOL``). Of the array's
+    attributes only the keys of ``attrs`` are compared, as JSON stores them; every other key
+    is someone else's."""
     out = []
     if tuple(meta.shape) != tuple(int(n) for n in shape):
         out.append(f"shape: found {tuple(meta.shape)}, wanted {tuple(shape)}")
@@ -826,4 +829,10 @@ def differences(
             out.append(
                 f"{what}: found {tuple(float(x) for x in f)}, wanted {tuple(float(x) for x in w)}"
             )
+    found_attrs = meta.attrs or {}
+    for key, wanted in json.loads(json.dumps(attrs or {})).items():
+        if key not in found_attrs:
+            out.append(f"attrs[{key!r}]: missing, wanted {wanted!r}")
+        elif found_attrs[key] != wanted:
+            out.append(f"attrs[{key!r}]: found {found_attrs[key]!r}, wanted {wanted!r}")
     return out
diff --git a/tests/test_read.py b/tests/test_read.py
index 91a6b2b..b27453f 100644
--- a/tests/test_read.py
+++ b/tests/test_read.py
@@ -554,6 +554,12 @@ def test_relabel_rescales_everything_by_the_ratio(tmp_path):
     resampled = wa.open(str(group / "s0"), voxel_size=(8, 8, 8), on_mismatch="resample")
     assert resampled.source_voxel_size == (4.0, 4.0, 4.0)
     assert resampled.voxel_size == (8.0, 8.0, 8.0)
+    for v in (
+        vol,
+        resampled,
+    ):  # where the level is, beside what it is (cellmap-flow's stored_placement)
+        assert v.source_translation == (10.0, 10.0, 10.0)
+    assert wa.open(str(group / "s0"), units="micrometer").source_translation == (0.01, 0.01, 0.01)
 
 
 def test_a_resampled_box_is_the_target_voxels_shape_counts(tmp_path):
diff --git a/tests/test_resample.py b/tests/test_resample.py
index 5e2c3b0..46e992e 100644
--- a/tests/test_resample.py
+++ b/tests/test_resample.py
@@ -139,3 +139,45 @@ def test_partial_block_mean_and_edge_fill(tmp_path):
         ((0, 0, 0), (48, 16, 16)), voxel_size=(8, 8, 8), method="nearest", padding="edge"
     )
     assert list(near[:, 0, 0]) == pytest.approx([1, 3, 5, 6, 6, 6])
+
+
+def test_read_onto_reads_another_grids_voxels(tmp_path):
+    """cellmap-analyze's per-operation grid (Behaviour 20): voxels [begin, end) of a grid,
+    exactly that many, from any array; at the array's own voxel size nothing is resampled,
+    however far under a voxel the grids are offset."""
+    path, data = _ramp(tmp_path, "float32")
+    vol = wa.open(path, method="linear")
+    near = wa.Grid((8, 4, 4), (4, 4, 4), (3, 3, 3))  # corner 1 nm past the array's
+    np.testing.assert_array_equal(vol.read_onto(near), data)  # stored values, not interpolated
+    far = wa.Grid(
+        (8, 4, 4), (4, 4, 4), (5, 2, 2)
+    )  # 3 nm past on z: each grid voxel holds the next centre
+    out = vol.read_onto(far)
+    assert out.shape == (8, 4, 4)
+    np.testing.assert_array_equal(out[:7], data[1:])
+    assert np.all(out[7] == 0)  # the last grid voxel holds no stored centre: the fill
+    finer = wa.Grid((16, 8, 8), (2, 2, 2), (-1, -1, -1), chunks=(4, 4, 4))  # corner -2
+    expect = vol.read(finer.box, voxel_size=(2, 2, 2), anchor=finer.corner)
+    np.testing.assert_array_equal(vol.read_onto(finer), expect)
+    halves = np.concatenate(
+        [vol.read_onto(finer, (0, 0, 0), (6, 8, 8)), vol.read_onto(finer, (6, 0, 0), (16, 8, 8))]
+    )
+    np.testing.assert_array_equal(halves, expect)  # adjoining reads agree with one read
+    in_um = wa.Grid((16, 8, 8), (0.002,) * 3, (-0.001,) * 3, units="micrometer")
+    np.testing.assert_array_equal(vol.read_onto(in_um), expect)
+    # the grid's voxels -1 and 0 are centred at -3 and -1 nm, before the data: fill, masked
+    out, inside = vol.read_onto(finer, (-1, 1, 1), (3, 3, 3), mask=True)
+    assert out.shape == (4, 2, 2) and not inside[:2].any() and inside[2:].all()
+    with pytest.raises(ValueError, match="spatial axes"):
+        vol.read_onto(wa.Grid((4, 4), (4, 4), (2, 2), axes=("y", "x")))
+
+
+def test_read_onto_its_own_grid_is_the_stored_array(tmp_path):
+    data = np.arange(7 * 6 * 5, dtype="uint16").reshape(7, 6, 5)
+    group = tmp_path / "odd.zarr"
+    write_zarr2_group(group, ome_multiscales("zyx", [("s0", (3.54, 4, 5.3), (1.0, -2.3, 7.7))]))
+    write_zarr2(group / "s0", data, chunks=(3, 3, 3))
+    vol = wa.open(str(group / "s0"))
+    np.testing.assert_array_equal(vol.read_onto(wa.Grid.from_meta(vol.grid)), data)
+    twice = wa.Grid.from_meta(vol.grid, voxel_size=(1.77, 2, 2.65))
+    assert vol.read_onto(twice).shape == twice.shape == (14, 12, 10)
diff --git a/tests/test_write.py b/tests/test_write.py
index ed70397..daff2cb 100644
--- a/tests/test_write.py
+++ b/tests/test_write.py
@@ -272,6 +272,19 @@ def test_create_never_overwrites_and_adopts_only_an_identical_array(tmp_path):
         wa.create(path, *args[:3], (3, 2, 2), chunks=(4, 4, 4), exists="adopt")
 
 
+def test_adopt_checks_the_attributes_it_was_given(tmp_path):
+    """cellmap-flow's _refuse_another_region: funlib's prepare_ds kept an old offset silently."""
+    path = str(tmp_path / "vol.zarr/s0")
+    args = ((8, 8, 8), "uint8", (4, 4, 4), (2, 2, 2))
+    wa.create(path, *args, chunks=(4, 4, 4), attrs={"offset": (0, 0, 0), "model": "a"})
+    again = wa.create(path, *args, chunks=(4, 4, 4), attrs={"offset": [0.0, 0, 0]}, exists="adopt")
+    assert again.meta.attrs["model"] == "a"  # keys not given are not compared
+    with pytest.raises(ValueError) as e:
+        wa.create(path, *args, chunks=(4, 4, 4), attrs={"model": "b", "run": 2}, exists="adopt")
+    assert "attrs['model']: found 'a', wanted 'b'" in str(e.value)
+    assert "attrs['run']: missing, wanted 2" in str(e.value)
+
+
 def test_create_keeps_everyone_elses_attributes(tmp_path):
     group = tmp_path / "shared.zarr"
     write_zarr3_group(
@@ -720,6 +733,44 @@ def test_ensure_group_entry_lists_an_array_the_package_did_not_create(tmp_path,
         )
 
 
+def test_a_pickled_relative_path_reopens_after_a_chdir(tmp_path, monkeypatch):
+    """cellmap-analyze's dask workers change directory before they unpickle."""
+    monkeypatch.chdir(tmp_path)
+    wa.create("b.zarr/seg/s0", (8, 8, 8), "uint8", (4, 4, 4), (2, 2, 2), chunks=(4, 4, 4))
+    vol = wa.open("b.zarr/seg/s0", mode="r+")
+    vol.write_index((0, 0, 0), (4, 4, 4), np.full((4, 4, 4), 9, np.uint8))
+    blob = pickle.dumps(vol)
+    (tmp_path / "elsewhere").mkdir()
+    monkeypatch.chdir(tmp_path / "elsewhere")
+    again = pickle.loads(blob)
+    assert again.path == str(tmp_path / "b.zarr/seg/s0")
+    assert int(again.read_index((0, 0, 0), (1, 1, 1))[0, 0, 0]) == 9
+    assert vol.reopen().level_path == again.level_path
+
+
+def test_a_scalar_fills_the_box(tmp_path):
+    vol = wa.create(
+        str(tmp_path / "vol.zarr/s0"), (8, 8, 8), "uint8", (1, 1, 1), (0.5,) * 3, chunks=(4, 4, 4)
+    )
+    vol.write_index((0, 0, 0), (4, 8, 8), 7)
+    vol.write(((4, 0, 0), (8, 4, 4)), np.uint8(3))
+    vol.write_chunk((1, 1, 1), 5)
+    got = vol.read_index((0, 0, 0), (8, 8, 8))
+    assert np.all(got[:4] == 7) and np.all(got[4:, :4, :4] == 3) and np.all(got[4:, 4:, 4:] == 5)
+    for bad in (300, -1, 0.5, float("nan"), True + 255):
+        with pytest.raises(ValueError, match="does not fit uint8"):
+            vol.write_index((0, 0, 0), (4, 4, 4), bad)
+    vol.write_index((0, 0, 0), (4, 4, 4), 0)  # the fill over a whole unit leaves it absent
+    assert (0, 0, 0) not in vol.list_chunks() and (0, 1, 0) in vol.list_chunks()
+    f = wa.create(
+        str(tmp_path / "f.zarr/s0"), (4,), "float32", (1,), (0.5,), axes=("x",), chunks=(4,)
+    )
+    f.write_index((0,), (4,), 0.1)  # rounded to the array's precision
+    assert np.all(f.read_index((0,), (4,)) == np.float32(0.1))
+    with pytest.raises(ValueError, match="does not fit float32"):
+        f.write_index((0,), (4,), 1e300)
+
+
 def test_a_pickled_handle_keeps_its_context(tmp_path):
     path = str(tmp_path / "vol.zarr/s0")
     wa.create(path, (8, 8, 8), "uint16", (4, 4, 4), (2, 2, 2), chunks=(4, 4, 4))
diff --git a/uv.lock b/uv.lock
index 9b6fbf0..1780bd0 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1247,7 +1247,7 @@ wheels = [
 
 [[package]]
 name = "worldarray"
-version = "0.3.0"
+version = "0.4.0"
 source = { editable = "." }
 dependencies = [
     { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },