Release 2.4.0

What's New

RIDDL has always required every value to be typed. 2.4.0 enforces that rule for the two values
whose spelling carries no type of their own: empty/none and prompt(…). A bare empty takes
the type of its position. Where nothing supplies a type, you write empty T, and T is a
type NAME, never a type expression. Declare type MaybeRecipe is Recipe? and write empty MaybeRecipe. See the Computational Model §0.3, ruling 6.

Features

• empty is checked at every position that types it. A bare empty is now held to the
  minimum-cardinality rule at every position that supplies a type:

  • set, append/remove and let
  • constructor, call, initiate and terminate arguments
  • put, return, require … with
  • store/upsert and update … set
  • comparison operands

  Fields typed inline (note: String?) are checked too. Before this release only four of those
  positions checked anything, and the rest accepted empty against types that cannot be empty.

• x == empty. empty can now be a comparison operand. The operand opposite it supplies its
  type.

• EmptyValue.typeRef and PromptValue.typeRef. A value's ascription is carried as the
  TypeRef it names. New non-deprecated readers: ascribedType and expressionAscription.
  A bare predefined type (Real, Boolean, TimeStamp) counts as a name; a parameterized one
  (String(1,30)) does not.

• Three new Errors:

  • value-ascription-not-a-name: empty String? or prompt(…) as T*. The 2.0 spelling was
    admitted by mistake; it still parses, so the message names the fix.
  • value-empty-untyped: a bare empty where nothing supplies a type, such as an untyped
    let, log empty or empty == empty.
  • value-empty-ascription-contradicts: an ascription that is not, syntactically, the
    position's declared type. A field typed inline takes a bare empty.

Bug Fixes

• update T set f = empty where … failed to parse. empty's optional ascription read the
  next word as a type name. put empty to …, append empty to … and store empty in … failed
  the same way. The guard after empty is now a word category rather than a list, and the EBNF
  carries the same rule.
• empty with an alias type was refused. empty MaybeNote with type MaybeNote is String?
  reported "requires at least one value", and so did a bare empty for any alias-typed
  constructor field. The cardinality check now reads through type names, including alias chains,
  and is safe on cyclic ones.
• riddlc dump --json misreported alias-typed fields. cardinality and acceptsEmpty came
  back as exactly-one and false for a field whose alias names an optional or a collection.
• Prettify dropped the keyword from empty record R, so the output re-parsed as a different
  AST.
• Messages.justWarnings returned advisories. It filtered by severity band, and Advisory
  sits inside that band while reporting isWarning = false. RiddlLib.ValidateResult (as a
  trailing defaulted field) and the JS facade both gain an advisories bucket. Before this, an
  advisory reached consumers only inside all. Blocking behaviour (--fail-on) was never
  affected.
• PromptValue had no JS export. Its @JSExportTopLevel annotation was attached to
  EmptyValue.

Deprecations

• EmptyValue.typeEx and PromptValue.typeEx are deprecated in favour of typeRef and
  ascribedType. Both still parse and are still populated, so existing readers see no change.

Internal

• BAST and JSON wire formats are unchanged; typeRef is derived on read, so files written by
  2.3.x load to the same AST the parser builds.
• EbnfReservedWordsTest keeps the grammar's hand-kept word lists in step with the parser.
• A corpus walk in the test suite prunes target/ instead of filtering it afterwards.
• Corpus check: all 191 riddl-models entry points validate with 0 errors, as they did on 2.3.1.