Manifest parsing

metewand-core parses a version-1 metewand.yaml document into domain types for datasets, problems, implementations, environments, experiments, execution policies, and observation policies. The parser is pure: the caller supplies the source text and its logical path, and parsing performs no filesystem or network access.

use std::path::Path;

use metewand_core::manifest::{DatasetDefinition, parse_manifest};

let source = r#"
version: 1
name: "example"
datasets:
  fixed:
    output-schema: "schemas/dataset.json"
    sources: ["datasets/fixed/**"]
"#;

let manifest = parse_manifest(Path::new("metewand.yaml"), source)?;
assert!(matches!(
    manifest.datasets["fixed"],
    DatasetDefinition::Fixed(_)
));

# Ok::<(), metewand_core::manifest::ManifestParseError>(())

Metewand-owned YAML option keys use kebab-case, including output-schema, execution-policies, and max-concurrency. Underscore spellings are rejected. Enum values such as one_shot and prepare_and_execute, benchmark-defined names, parameter keys, and family-owned semantics keep their declared spelling. This unreleased version-1 spelling change preserves canonical internal records and content identities.

The typed boundary enforces the version-1 manifest shape: required fields, runner and environment variants, version constants, name and path syntax, nonempty and unique lists required by the public schema, safe integer bounds, and the one-shot capability and observation-policy restrictions. Every Metewand-owned object rejects unknown fields, including nested workers, parameter axes, and experiment cases.

Execution policies accept max-concurrency as a positive safe integer. Its default is 1, which keeps performance measurements serial. Zero, negative, fractional, string, and out-of-range values fail parsing before execution.

Literal parameter defaults and axis values become CanonicalValues as they are parsed. Values must fit Metewand's language-neutral JSON domain: strings, booleans, nulls, safe integers, finite numbers, arrays, and mappings with string keys. Non-finite numbers and unsupported YAML tags are rejected. An untagged date or time remains a string. Nulls are preserved in parameter values, while nulls in typed fields such as datasets fail the required mapping shape. Optional typed fields may be omitted, but an explicit null, including an empty YAML value, is rejected. For example, omitting enforcement uses its default; enforcement: null and enforcement: are invalid.

YAML syntax

Manifests and problem contracts use a single YAML mapping document. Explicit mapping keys must be unique strings. Anchors, aliases, and merge keys are supported, with bounded alias expansion. After alias and merge expansion, mappings and sequences may nest at most 24 levels. Explicit fields override merged fields, and earlier mappings in a merge sequence take precedence. Tags that introduce non-JSON types and external includes are unsupported.

version: 1
name: "shared-settings"
environments:
  local: { kind: "local" }
execution-policies:
  serial: &serial
    version: 1
    max-concurrency: 1
  parallel:
    <<: *serial
    max-concurrency: 2

The CLI defaults to metewand.yaml. An explicit --manifest accepts any filename, including .yml, and always reads YAML. TOML configuration is no longer accepted. Native ecosystem files such as Cargo.toml, pyproject.toml, and Julia's Manifest.toml retain their own formats.

Diagnostics

ManifestParseError retains the logical source path and the decoder's byte range as a one-indexed SourceSpan. Columns count Unicode scalar values, and the end position is exclusive. This representation can feed the shared typed diagnostic envelope without reparsing human-readable error text.

Repository paths cross the typed boundary only in normalized lexical form. They are relative UTF-8 paths with / separators and no empty, ., or .. components; absolute paths, backslashes, and trailing separators are rejected. Parsing does not resolve those paths against a repository root, inspect referenced schemas or problem contracts, expand source bundles, or follow symlinks. The pure parameter-namespace validator performs the first structural ownership checks. After the caller loads problem contracts, the pure manifest compatibility validator checks their relationships with datasets, implementations, and policies. metewand-runtime performs the filesystem-sensitive source-bundle expansion.

The complete version-1 YAML fixture is fixtures/manifest/v1/complete.yaml.