Release 0.1.0a2

CHANGELOG.md

@@ -4,9 +4,12 @@ Notable changes per release. Versions follow the rules in
 [contributing](docs/contributing.md#plugin-api-and-versioning): the plugin API is stable,
 and a release that breaks it says so here.
 
-## 0.1.0a1 (unreleased)
+## 0.1.0a2 (2026-10-21)
 
-The first release, an alpha: the plugin API can still change before 0.1.0.
+The second alpha. The plugin API only grew (`overlap`, `computed_from`,
+`demand.cancelled()`, source matchers, parameter models), but an `ArrayInfo` given no
+translation now has voxel 0's corner at the origin, not its centre (under Sources): a
+plugin's source that meant its centre there passes `translation=(0.0,) * ndim`.
 
 ### Core and contrib
 
@@ -22,14 +25,8 @@ The first release, an alpha: the plugin API can still change before 0.1.0.
 
 ### Serving
 
-- Pipelines served as N5, Zarr v2 (OME-NGFF 0.4), Zarr v3 (OME-NGFF 0.5) and Neuroglancer
-  precomputed at once, and meshes made when fetched.
-- A per-stage chunk cache keyed by the pipeline's hash, and cache-busting digests in links.
-- Live edits through the REST API, Server-Sent Events, a control page and a
-  python-neuroglancer viewer that follows them.
-- Work bounded, shared between requests and ordered finest level first, then by when
-  each request arrived; work no client waits for any more is dropped, and a long op that
-  asks `demand.cancelled()` between its steps gives up once none does.
+- A long op that asks `demand.cancelled()` between its steps gives up once no client
+  waits for its work any more.
 - BLAS kept to one thread per call in the server (`threadpoolctl`), and affines applied
   without it: its threads fought the server's, and a three-field `warp://` read took ten
   times as long.
@@ -43,18 +40,6 @@ The first release, an alpha: the plugin API can still change before 0.1.0.
   A100, 13.5 s on 24 CPU threads); values within a grey level of the CPU's, cache keys the
   same. A sampled field is read in its own precision, not copied
   to float64.
-- `--token` for the control API, `--https` with a self-signed certificate.
-- zarr v2 and v3 chunks compressed with blosc (zstd, byte shuffle) by default, 10 to 30
-  times quicker to encode than gzip; `--compressor` chooses another. `--cache-gb` takes effect (an empty cache
-  given to the registry used to be replaced by the default 2 GiB one), and `0` turns the
-  cache off.
-- `serve --port 0` and `--ready-file` for launchers; the port is bound before it is
-  announced.
-- Datasets resolved by name on first request (`resolver`), routes from other packages
-  (`extra_routes`, the `chunkmirage.routes` entry point), and the app mounted under a prefix
-  in another app.
-- `chunkmirage.serve` and `Server`: an app served from Python on a port bound first, with a
-  callback once it accepts connections, in a background thread or this one.
 - `chunkmirage materialize` and `chunkmirage.materialize`: a dataset, or a box of it,
   computed and written into a store (a directory, `gs://`, `s3://`) as a frontend serves it,
   zarr v3, zarr v2 or N5, so it reads back as a stored dataset of its own; the metadata
@@ -78,17 +63,6 @@ The first release, an alpha: the plugin API can still change before 0.1.0.
   whether it is computed and whether the full-resolution result under it is (a preview
   told from a result while browsing), read afresh at every request;
   `GET /api/datasets/{name}/coverage` gives the counts per level.
-
-- zarr v2/v3, N5, precomputed and HDF5 on file, S3, GCS and HTTP; xarray arrays with their
-  coordinates and CF packing; cloud-optimized GeoTIFFs.
-- Computed sources: `synthetic://`, `scene://` (OME-Zarr 0.6 transformations), `warp://`,
-  `register://` (deformable registration on a GPU), `stitch://` (BigStitcher tiles),
-  `stack://` and `flip://`.
-- S3 read anonymously before credentials, GCS through its public URL when credentials
-  fail, metadata reads that give up after seconds, zarr compressors with members
-  tensorstore rejects.
-- Voxels placed as Neuroglancer places them: precomputed `voxel_offset` and funlib `offset`
-  are corners.
 - Stored arrays are opened and placed by worldarray (`worldarray>=0.1,<1`), the reader
   chunkmirage shares with cellmap-flow, cellmap-analyze and mesh-n-bone; chunkmirage keeps
   xarray coordinates, CF packing, the overrides and `kind`
@@ -109,8 +83,6 @@ The first release, an alpha: the plugin API can still change before 0.1.0.
   as precomputed their `voxel_offset` is 0 rather than a rounded half voxel. Explicit
   translations, and arrays that take their input's, are unchanged
   ([formats](docs/concepts/formats.md#stored-sources)).
-- Stored integers of 32 bits or more read as labels, booleans as masks, unless the spec's
-  `kind` says otherwise.
 
 ### Ops and pipelines
 
@@ -124,28 +96,12 @@ The first release, an alpha: the plugin API can still change before 0.1.0.
   cached as a stage of their own, and a served chunk is the weighted mean of every block
   reaching it, smooth where blocks computed on their own disagree and the same whatever
   order they are computed in; `register://`'s refined blocks blend with the same weights.
-
-- Pointwise ops, filters, morphology, connected components, spots, contacts, differences,
-  gradients, downsampling, normalized differences, slope and hillshade, with halos handled
-  by the pipeline.
-- Ops that return their block shaved by the halo or add leading axes; ops at one
-  resolution with the coarser levels downsampled from their output, or the nearest level
-  read as it is (`input_level`, `level_rtol`).
-- `cache_token()` for ops whose output depends on state outside their parameters, and
-  `DatasetRegistry.refresh` to read it again.
-- `slots` to bound how many calls of an op run at once.
 - `Pipeline.plan(level, box)`: what a read would compute, computing nothing: per stage, the
   chunks it is asked for, those held and those to compute, following `dataset://` into
   the datasets it reads and stopping at other sources.
-- Zero padding past the volume's edge (`padding`).
-- Channel axes read whole, so an op may change their length: select or combine channels.
-- A level shrunk by a whole factor for an op at one voxel size is the mean of each block,
-  not linear samples.
 
 ### Browser engine
 
-- Registration on WebGPU and pipelines of the package's own ops in Pyodide, in a gallery of
-  demos.
 - The engine as a base for a page of another project
   ([design](docs/design.md#another-page-on-the-engine)): Python loads only when a view
   needs it, a view with no ops served from the reader alone; `unserve()` forgets a view and
@@ -204,3 +160,65 @@ The first release, an alpha: the plugin API can still change before 0.1.0.
   behind every chunk a client waits for, in all the queue's slots but one, and never dropped
   as abandoned (`submit`, `claim` and `release` take `background`;
   [caching](docs/concepts/caching.md#order-of-work-and-requests-given-up-on)).
+
+## 0.1.0a1 (2026-10-05)
+
+The first release, an alpha: the plugin API can still change before 0.1.0.
+
+### Serving
+
+- Pipelines served as N5, Zarr v2 (OME-NGFF 0.4), Zarr v3 (OME-NGFF 0.5) and Neuroglancer
+  precomputed at once, and meshes made when fetched.
+- A per-stage chunk cache keyed by the pipeline's hash, and cache-busting digests in links.
+- Live edits through the REST API, Server-Sent Events, a control page and a
+  python-neuroglancer viewer that follows them.
+- Work bounded, shared between requests and ordered finest level first, then by when
+  each request arrived; work no client waits for any more is dropped.
+- `--token` for the control API, `--https` with a self-signed certificate.
+- zarr v2 and v3 chunks compressed with blosc (zstd, byte shuffle) by default, 10 to 30
+  times quicker to encode than gzip; `--compressor` chooses another. `--cache-gb` takes effect (an empty cache
+  given to the registry used to be replaced by the default 2 GiB one), and `0` turns the
+  cache off.
+- `serve --port 0` and `--ready-file` for launchers; the port is bound before it is
+  announced.
+- Datasets resolved by name on first request (`resolver`), routes from other packages
+  (`extra_routes`, the `chunkmirage.routes` entry point), and the app mounted under a prefix
+  in another app.
+- `chunkmirage.serve` and `Server`: an app served from Python on a port bound first, with a
+  callback once it accepts connections, in a background thread or this one.
+
+### Sources
+
+- zarr v2/v3, N5, precomputed and HDF5 on file, S3, GCS and HTTP; xarray arrays with their
+  coordinates and CF packing; cloud-optimized GeoTIFFs.
+- Computed sources: `synthetic://`, `scene://` (OME-Zarr 0.6 transformations), `warp://`,
+  `register://` (deformable registration on a GPU), `stitch://` (BigStitcher tiles),
+  `stack://` and `flip://`.
+- S3 read anonymously before credentials, GCS through its public URL when credentials
+  fail, metadata reads that give up after seconds, zarr compressors with members
+  tensorstore rejects.
+- Voxels placed as Neuroglancer places them: precomputed `voxel_offset` and funlib `offset`
+  are corners.
+- Stored integers of 32 bits or more read as labels, booleans as masks, unless the spec's
+  `kind` says otherwise.
+
+### Ops and pipelines
+
+- Pointwise ops, filters, morphology, connected components, spots, contacts, differences,
+  gradients, downsampling, normalized differences, slope and hillshade, with halos handled
+  by the pipeline.
+- Ops that return their block shaved by the halo or add leading axes; ops at one
+  resolution with the coarser levels downsampled from their output, or the nearest level
+  read as it is (`input_level`, `level_rtol`).
+- `cache_token()` for ops whose output depends on state outside their parameters, and
+  `DatasetRegistry.refresh` to read it again.
+- `slots` to bound how many calls of an op run at once.
+- Zero padding past the volume's edge (`padding`).
+- Channel axes read whole, so an op may change their length: select or combine channels.
+- A level shrunk by a whole factor for an op at one voxel size is the mean of each block,
+  not linear samples.
+
+### Browser engine
+
+- Registration on WebGPU and pipelines of the package's own ops in Pyodide, in a gallery of
+  demos.

docs/contributing.md

@@ -135,4 +135,6 @@ to the stable API waits for a major release.
    `pyproject.toml` and `chunkmirage/__init__.py` (`tests/test_docs.py` checks the three
    agree).
 2. Merge to `main` with CI green.
-3. Publish to PyPI only when the owners say so: `uv build && uv publish`.
+3. Publish to PyPI only when the owners say so: a GitHub release tagged `v<version>` runs
+   `.github/workflows/publish.yml`, which builds the wheel and uploads it (a manual run of
+   the workflow uploads to TestPyPI instead).

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "chunkmirage"
-version = "0.1.0a1"
+version = "0.1.0a2"
 description = "Spoof chunked array formats (zarr, n5, neuroglancer precomputed) over HTTP with on-the-fly processing, caching, and pluggable ops."
 readme = "README.md"
 license = { file = "LICENSE" }

src/chunkmirage/__init__.py

@@ -8,4 +8,4 @@
 from chunkmirage.sources import open_source
 
 __all__ = ["ArrayInfo", "Box", "Pipeline", "PipelineSpec", "Server", "create_app", "materialize", "open_source", "serve"]
-__version__ = "0.1.0a1"
+__version__ = "0.1.0a2"