fix(dart): parse Picoschema parenthetical types per spec

[pavelgj]

Fixes genkit-ai/genkit-dart#562

The Dart Picoschema converter treated parenthesized qualifiers as descriptions, so tags(array): string produced {type: string, description: "array"}. Some schemas also skipped conversion entirely (raw Picoschema was sent to the model) because isPicoschema only recognized bare primitive values. The spec runner never compared output, so spec/picoschema.yaml reported "all pass" while 9 of its 18 cases were actually wrong.

The converter now follows the JS reference implementation.

output:
  schema:
    tags(array, the tags): string, a tag   # {type: array, description: the tags, items: {type: string, description: a tag}}
    obj?(object, nested):                 # {type: [object, null], description: nested, properties: ..., additionalProperties: false}
      x: integer
    status(enum): [A, B]                  # {enum: [A, B]}
    (*): number                           # additionalProperties: {type: number}

Changes

• picoschema.dart: rewritten to mirror js/src/picoschema.ts:
  • (array|object|enum[, description]) qualifiers and (*) wildcards per spec; optional fields are nullable.
  • Top-level JSON Schema (type: ..., bare properties) is passed through.
  • New Picoschema.parse(schema, {schemas, schemaResolver}) resolves named schemas from schemas first, then the async resolver. toJsonSchema stays sync.
• Dotprompt._resolveMetadata always converts via Picoschema.parse, so DotpromptOptions.schemaResolver is now used for named schemas (it was ignored before).
• spec_test.dart checks output and passes the spec's schemas:.
• picoschema_test.dart: ports the JS cases and adds a regression group for genkit-dart#562.
• Version 1.1.0, CHANGELOG, PARITY, API docs. Also fixed the README (age: integer? -> age?: integer) and replaced gemini-pro in examples.

Behavior changes (strict, like other runtimes)

These now throw PicoschemaException:

Picoschema.toJsonSchema({'email(the email)': 'string'}); // use `email: string, the email`
Picoschema.toJsonSchema({'wild(*)': 'string'});          // use `(*): string`
Picoschema.toJsonSchema({'tags': 'string[]'});           // use `tags(array): string`
Picoschema.toJsonSchema({'s': 'a | b'});                 // use `s(enum): [a, b]`
Picoschema.toJsonSchema({'n': 'int'});                   // use `integer`
Picoschema.toJsonSchema({'a': 'Unknown'});               // was a silent {"$ref": "Unknown"}

Intentional deviations from JS (commented in code): a top-level any yields {} (JS yields the invalid {type: "any"}), and toJsonSchema(null) still returns {type: object}.

Testing

• dart analyze: clean; dart test: 237 passing.
• genkit-dart packages/genkit/test/ai/prompt_test.dart passes (141) against this branch via a local path override.

Follow-up in genkit-dart after publishing: bump to dotprompt: ^1.1.0, optionally use Picoschema.parse with the registry resolver in prompt_loader_io.dart, and add an (array) output to the testapps/prompts sample.

[huangjeff5]

This package is 1.1.0. The install snippet still says dotprompt: ^0.0.1.

^0.0.1 only allows 0.0.1. Someone following that page never gets this parser, so tags(array): string stays a string whose description is array.

I suggest dotprompt: ^1.1.0.

[pavelgj]

This package is 1.1.0. The install snippet still says dotprompt: ^0.0.1.

Good catch, updated both the API docs and the README to ^1.1.0. (a608d56)