Supported Surface

This page is the 0.2.0 support contract for dag-ml. It separates production-facing surfaces from conformance fixtures and backlog work. It does not change any public ABI, JSON schema, Rust, Python or WASM signature.

Support Levels

Level

Meaning

Supported

Included in the release scope and covered by CI gates.

Conformance

Stable enough for binding authors and integration tests, but not a complete production adapter.

Experimental

Public shape may exist, but release notes must call out limitations.

Backlog

Not part of the release promise.

dag-ml Surface

Area

Level

Notes

Graph, campaign and execution-plan contracts

Supported

Rust validation, JSON Schemas, C ABI discovery and CLI validation are gated.

Fold identity, OOF joins and leakage refusal

Supported

Sample-id joins, validation-only OOF, group/origin/repetition guards and D9 multisource negative cases are tested.

Deterministic selection and replay bundle validation

Supported

Plan/controller/data/artifact fingerprints and selection metric levels are validated.

Research provenance export

Supported

RO-Crate, PROV and OpenLineage exports are generated from validated internal contracts.

C ABI JSON contract helpers

Supported

Header snapshot, C conformance and non-mock replay paths are gated.

Runtime process adapter protocol

Supported

JSONL frames, describe handshake, timeouts, retries and worker pools are covered by CLI tests.

Python and WASM JSON-contract bindings

Supported

Wheel/package metadata and smoke tests are CI-gated; object-native Python DSL frontend is not included.

Pipeline DSL JSON compiler

Supported

Canonical JSON plus nirs4all-compatible serialized JSON descriptors are covered.

Direct Python/YAML object DSL frontend

Backlog

Host object resolution remains binding-owned.

sklearn production process adapter

Conformance

The reference adapter is tested and useful for integration, but release notes must list supported estimator families and persistence limits.

prospectr and mdatools process adapters

Conformance

Tested reference adapters for selected R operators; stateful msc, simca and mcrals remain backlog.

SpectroChemPy and Orange-Spectroscopy adapters

Backlog

Tracked in docs/HOST_ADAPTER_BACKLOG.md.

EXPLAIN phase execution through host adapters

Experimental

Contracts and mock replay exist; production adapter dispatch is not a final-release promise.

Controller-side task batching

Backlog

Parallel scalar scheduling is supported; batch requests/static subgraphs remain future hardening.

dag-ml-data Dependency

dag-ml 0.2.0 consumes the sibling dag-ml-data contracts through JSON-identical schemas and fixtures. The supported cross-repo contract for this release is:

  • CoordinatorDataPlanEnvelope v1;

  • FeatureFusionSelector v1;

  • CoordinatorBranchView v1;

  • FittedAdapterRef v1 as a data-side replay/persistence contract;

  • shared FoldSet, conformance pack and parity-oracle manifests.

The following dag-ml-data capabilities are required for release validation but remain provider-specific at runtime:

  • host-side execution of branch_view modes by_metadata, by_tag and by_filter;

  • materialize/predict signal-type enforcement once dag-ml carries expected signal type through replay;

  • production provider arenas beyond the in-memory conformance provider.

Public-Signature Policy

For the 0.2.0 release window:

  • no C ABI symbol, struct layout, JSON schema id/version, Rust public function, Python facade function or WASM export changes without an explicit contract entry and ABI/schema snapshot update;

  • if such a change is accepted, downstream chains such as nirs4all-core, nirs4all-web and browser/Python smoke packages must be rebuilt before tag;

  • documentation, CI jobs, tests and private benchmark helpers are allowed when they do not alter exported signatures.

Post-0.2.0 Backlog

  1. Keep the dag-ml-capi AddressSanitizer lane green and extend it beyond library unit tests when C ABI lifecycle coverage expands.

  2. Extend the initial performance probes to replay cache export and process worker pools.

  3. Raise public Rust documentation coverage toward the ADR target in follow-up hardening.