[d-v-b]

🤖 AI text below 🤖

Added a context to zarr.ArrayIO and zarr.AsyncArrayIO: a value defining the data types, codecs and codec pipeline in scope for one array. Extension names in a stored document or in build's keyword arguments resolve through the context instead of the default registries; a kind the context leaves unset falls back to the defaults, while a kind it sets is a closed world, and a name outside it fails to resolve with an error naming the scope.

The type is the existing zarr.core.context.Context — the extensions-in-scope value the structured data type resolution already threads through nested reads — extended with codecs: Mapping[str, type[Codec]] and codec_pipeline: type[CodecPipeline], and exported from zarr.experimental (the ArrayIO classes themselves are public in the zarr namespace). An optional context= is threaded through the resolution functions: parse_dtype and the data-type getters, codec normalization (_resolve_codec, _parse_*_codec, parse_codecs), from_dict on both metadata classes, parse_array_metadata, create_codec_pipeline, and build_array_metadata. With no context, every path is byte-identical to before.

The layering and the document-model binding land in #419; this PR only adds the context. With the document-first binding the context is consumed only where the document resolves into classes, so opening an array whose codec names fall outside the context still works for document operations, and resolution fails when the data path is built.

Scope notes:

• Zarr format 2 filters and compressors are numcodecs objects, resolved by numcodecs; no context governs them (documented on from_dict and build_array_metadata).
• Codec instances pass through any context: it scopes name resolution, not values.
• Chunk key encodings and chunk grids are not yet context kinds; they can be added when their registries become extension points.

Stack

Stacked on #419 (claude/arrayio-lazy-metadata-48d8e5); merge that first. This PR is the claude/arrayio-context branch.

🤖 Generated with Claude Code