hecks 0.2.0 → 0.3.0

data/lib/hecks/bluebook/dsl/attribute_collector.rb

@@ -0,0 +1,327 @@
+module Hecks
+  module Bluebook
+    module DSL
+      module AttributeCollector
+        ListOf = Struct.new(:type)
+        OneOf  = Struct.new(:values)
+
+        UNSET = Object.new.freeze
+        private_constant :UNSET
+
+        def attributes = @attributes ||= []
+
+        # Value objects synthesised from inline closed sets, collected here and
+        # installed by whoever owns value objects (the aggregate).
+        def closed_sets = @closed_sets ||= []
+
+        # `admits:` names a closed set that is ALREADY DECLARED elsewhere —
+        #
+        #   attribute :op, String, admits: "Vocabulary::QueryComparator"
+        #
+        # which is the difference between it and `one_of`: `one_of` SYNTHESISES
+        # a fresh value object named for the attribute, so it can only ever name
+        # something new. `admits` points at a set the language already holds, so
+        # the same set can be named from many places without being written twice.
+        #
+        # QUALIFIED, because a closed set is a value object INSIDE an aggregate
+        # and `reference_to` reaches heads only — so this is text, checked where
+        # it is read rather than by reference resolution.
+        #
+        # AND WRITTEN AS TEXT, not as the constant path `Vocabulary::QueryComparator`
+        # it reads like. The constant spelling was tried — `ConstShim` returning
+        # a Module so Ruby's `::` reaches a second `const_missing` — and it
+        # cannot hold: `Facade::Surface` installs every aggregate name as a
+        # TOP-LEVEL constant, so the moment any facade exists, `Vocabulary`
+        # resolves to that module and the shim is never asked. A spelling that
+        # works only until a facade is built is worse than a quoted one.
+        #
+        # THE TYPE POSITION TAKES A BARE CONSTANT, ALWAYS REQUIRED (ADR 0025,
+        # "Attributes") — omitting it (a mint default of String) and quoting
+        # it as text both refuse now. Neither had a real reason left: no
+        # corpus attribute ever omitted the type, and `ConstShim#const_missing`
+        # (S0b) already resolves a bare, not-yet-declared constant to the
+        # SAME forward reference the quoted form existed for — a bareword
+        # `Name` reaches a value object named "Name" declared later in the
+        # same block exactly as `"Name"` used to, `spell`'s own `to_s`
+        # renders either one identically. Neither form appears in any frozen
+        # era text (checked directly), so both are refused unconditionally —
+        # nothing for `MetaValidator.shadow_parsing?` to answer for.
+        # RENAMED FROM `attribute` — item #13's full metaprogrammed
+        # dispatch (slice 3, whole-project table-unification survey).
+        # The word `attribute` itself is no longer a real method any
+        # builder answers directly: every (context, word) Keyword row
+        # for it carries `calls: "attribute_impl"`, and `GenericDispatch`
+        # forwards the whole call here untouched — this method's own
+        # body is exactly what `attribute` always was, unchanged, just
+        # reached generically now rather than by Ruby's own direct
+        # method lookup. `attribute_collector_spec.rb` (`AttributeCollector
+        # has no method without a test` — dsl_coverage_spec.rb) and the
+        # bootstrap fallback (`GenericDispatch::BOOTSTRAP_CALLS_FALLBACK`) both
+        # name this same string; they must never drift apart.
+        def attribute_impl(name, type = UNSET, default: nil, optional: false, pattern: nil,
+                           admits: nil, one_of: nil)
+          # moved to the language: FieldName invariant, on Root.Attribute
+
+          if type.equal?(UNSET)
+            raise Malformed, "#{name} declares no type — attribute :#{name}, SomeType is required, " \
+                             "there is no default"
+          end
+
+          # `list_of("X")` carries the same quoted text one level down — a
+          # bare constant is required there too, checked before it unwraps
+          # below, not after (once unwrapped, a plain String in the `type:`
+          # slot looks identical to one that legitimately belongs there).
+          quoted = type.is_a?(ListOf) ? type.type : type
+          if quoted.is_a?(::String)
+            raise Malformed, "#{name}'s type #{quoted.inspect} is quoted text — give the bare constant " \
+                             "(#{quoted}) instead; a forward reference to a value object declared later " \
+                             "in the same block already resolves without quoting"
+          end
+
+          refuse_unshared_pattern(name, pattern) if pattern
+
+          type = synthesise_closed_set(name, type) if type.is_a?(OneOf)
+          list = type.is_a?(ListOf)
+          attributes << Attribute.new(
+            name:     name,
+            type:     list ? type.type : type,
+            list:     list,
+            default:  default,
+            optional: optional,
+            pattern:  pattern,
+            admits:   admits
+          )
+
+          install_inline_closed_set(name, one_of) if one_of
+        end
+
+        # RENAMED FROM `list_of` — item #13's full metaprogrammed
+        # dispatch (slice 5). Called in an attribute's own TYPE
+        # position (`attribute :x, list_of(Y)`), never through a `def
+        # list_of` any ONE builder answers as its own word — reached
+        # via `WordGate#word_gate_dispatch`'s new "Type"-context
+        # fallback, the same one `one_of_impl` below uses. Bootstrap-
+        # reachable (every core chapter's own list-typed attributes use
+        # it), so `GenericDispatch::BOOTSTRAP_CALLS_FALLBACK` carries a
+        # SINGLE `["Type", "list_of"]` entry rather than one per calling
+        # context — the bootstrap branch checks that key too now, same
+        # reasoning as the ordinary fallback.
+        def list_of_impl(type) = ListOf.new(type)
+
+        # `reference_to Account` MINTS `:account` — no `_id` — the default
+        # every `reference_to`-shaped word (`AggregateBuilder`,
+        # `EntityBuilder`, `CommandBuilder#cross_reference`,
+        # `QueryBuilder`, `PortOperationBuilder`, all `include
+        # AttributeCollector`) shares, since dropping the suffix is what
+        # makes `/` a real traversal operator possible at all (ADR 0025,
+        # "References" — a hop crosses at `/`, a field walk crosses at
+        # `.`; deriving the hop name from `account_id` is what made an
+        # explicit operator impossible before).
+        #
+        # LEGACY UNDER SHADOW-PARSING (S0a's own bridge): frozen era text
+        # was minted under the OLD default, and `EraGuard.shadow_parse`
+        # reconstructs a held aggregate's shape to DIFF against the
+        # current one — reading that text through the NEW default would
+        # silently reconstruct the WRONG historical name, not merely
+        # refuse a spelling the way S1's `identified_by` legacy forms do.
+        # This is a mint DEFAULT changing, not a syntax being removed, so
+        # there is nothing to refuse here — only a fork in what gets
+        # minted when `as:` is omitted.
+        private def default_reference_name(target)
+          suffix = MetaValidator.shadow_parsing? ? "_id" : ""
+          :"#{Naming.snake(target)}#{suffix}"
+        end
+
+        # A closed set declared INLINE on the attribute:
+        #
+        #   attribute :status, one_of("open", "shut")
+        #
+        # Desugars to a value object named for the attribute, so it goes through
+        # exactly the machinery a hand-written one_of does — and so the
+        # attribute's type is still a DECLARED value object, which is now a
+        # structural rule rather than a predicate.
+        #
+        # An earlier reading of this spelling parsed it and threw the values
+        # away: the attribute became a plain String and the closed set meant
+        # nothing, in a construct that looked supported. The desugaring is
+        # pinned now — the same bluebook must always yield the same IR.
+        # RENAMED FROM `one_of` — item #13's full metaprogrammed dispatch
+        # (slice 5), same reasoning as list_of_impl above. SAME NAME as
+        # `ValueObjectBuilder#one_of_impl`'s own override on purpose —
+        # that method's own `super(*values)` call (the no-block, bare
+        # type-position case) resolves by METHOD NAME up the ancestor
+        # chain, and renaming only one side would silently break it.
+        def one_of_impl(*values) = OneOf.new(values)
+
+        private
+
+        def relationship_attribute(target, kind, name, optional: false, list: false)
+          attributes << Attribute.new(
+            name:         name,
+            type:         Reference.new(target),
+            list:         list,
+            optional:     optional,
+            relationship: kind
+          )
+        end
+
+        # `one_of:` NAMES A CLOSED SET ON THE FIELD ITSELF (ADR 0025,
+        # "Attributes") — joining `pattern:`/`admits:` where value
+        # constraints already live, for the one context where it means
+        # anything: a `value_object` block, where it replaces the old
+        # `one_of do member ... end` wrapper for a SINGLE-FIELD set (a
+        # multi-field set still writes bare `member` lines, unwrapped).
+        # `ValueObjectBuilder` overrides this; every other includer
+        # (Aggregate/Entity/Command/Query/PortOperation) inherits this
+        # refusal — those contexts already have the unchanged type-position
+        # form (`attribute :status, one_of("open", "shut")`) for an
+        # anonymous inline set, so `one_of:` naming a *field* there would be
+        # a second spelling of the same idea, not a new one.
+        def install_inline_closed_set(name, _values)
+          raise Malformed,
+                "#{name}'s one_of: only means something inside a value_object — name a closed set " \
+                "with the type-position one_of(...) instead"
+        end
+
+        # A pattern is refused AT DECLARATION, not when a value first meets it :
+        # a regex whose meaning depends on which engine reads it is a defect in
+        # the bluebook, and a bluebook that loads is one whose patterns carry
+        # one meaning. PatternSubset says which those are, and why each is refused.
+        def refuse_unshared_pattern(name, pattern)
+          rejection = PatternSubset.validate(pattern)
+          return unless rejection
+
+          raise Malformed,
+                "#{name}'s pattern #{pattern.inspect} uses a #{rejection.construct} — " \
+                "#{rejection.reason}"
+        end
+
+        def synthesise_closed_set(name, one_of)
+          type = Naming.pascal(name)
+          closed_sets << ValueObject.declare(
+            name:       type,
+            attributes: [Attribute.new(name: :value, type: "String")],
+            members:    one_of.values.map { |value| { value: value.to_s } },
+            closed_set: true
+          )
+          type
+        end
+
+        # Every selected identity head contributes all of its recursively
+        # scalar leaves in declaration order. Named, inline and compound-key
+        # declarations therefore share one flattening rule.
+        def resolve_identity_field!(field, value_objects, context_name)
+          attr = attributes.find { |a| a.name == field }
+
+          # `:id` OR AN `_id`-SUFFIXED NAME WITH NO MATCHING ATTRIBUTE is
+          # not a typo to refuse — it is the language's own WALK-PARENT/
+          # FALLBACK-IDENTITY convention. The `_id` suffix is the meta-
+          # domain's own (the `Command`/`Entity`/`Aggregate` etc. records
+          # identify by `owner_id`/`bluebook_id`/`aggregate_id`, supplied
+          # by the judge's replay rather than declared locally — see
+          # `behavior.bluebook`'s own "owner_id is not a declared
+          # attribute" comment); bare `:id` is `Instance#materialize_
+          # identity!`'s own fallback name (`@aggregate.identified_by ||
+          # :id`) made explicit rather than left implicit — declaring
+          # `identified_by :id` says out loud what omitting `identified_by`
+          # entirely already meant. `reference_to` mints the `_id` suffix
+          # for the same reason `attr.reference?` below resolves bare: an
+          # `_id`(-shaped) name is already a scalar by the language's own
+          # spelling convention, walk-supplied or locally declared alike.
+          return [field.to_s] if attr.nil? && (field.to_s == "id" || field.to_s.end_with?("_id"))
+
+          raise Malformed, "#{context_name}.identified_by :#{field} names no attribute #{context_name} declares" unless attr
+
+          identity_paths_for_attribute(attr, value_objects, context_name, field.to_s, [])
+        end
+
+        # A named or inline identity mints one structured field, then expands
+        # each scalar leaf beneath it into the existing path-shaped IR.
+        def resolve_identity_type!(type, as, insert_at, value_objects, context_name)
+          target = Naming.demodulise(type.respond_to?(:hecks_name) ? type.hecks_name : type)
+          matches = value_objects.select { |value_object| value_object.hecks_name.to_s == target }
+          if matches.size > 1
+            raise Malformed, "#{context_name}.identified_by names duplicate value object #{target}"
+          end
+
+          vo = type.respond_to?(:attributes) ? type : matches.first
+          raise Malformed, "#{context_name}.identified_by names #{target}, which is not a declared value object" unless vo
+          if vo.attributes.empty?
+            raise Malformed, "#{context_name}.identified_by names #{target}, which declares no attributes"
+          end
+
+          field = (as || Naming.snake(target)).to_sym
+          if attributes.any? { |attribute| attribute.name == field }
+            raise Malformed,
+                  "#{context_name}.identified_by #{target} mints :#{field}, but that attribute is already declared"
+          end
+
+          # `Attribute.new` DIRECTLY, not the public `attribute(...)` DSL
+          # entry — `target` is `Naming.demodulise`'d TEXT, not a bareword
+          # the bluebook author typed (the type position's own quoted-text
+          # refusal is about DSL source, not internal minting), and `vo`
+          # ITSELF can't be passed either: it is the real, already-built
+          # `ValueObject` subclass, permanently anonymous from Ruby's own
+          # `to_s` (only `hecks_name` carries its name) — `Attribute#spell`
+          # would demodulise it to "#<Class:0x...>", not "PizzaName".
+          attributes << Attribute.new(name: field, type: target)
+          # MOVED to `insert_at` — the attribute count AT THE MOMENT
+          # `identified_by` was actually called, captured by the caller —
+          # not left where `attribute` just appended it. Resolution happens
+          # at BUILD time, after every other attribute in the block has
+          # already run, so appending would put the identity field LAST
+          # regardless of where `identified_by` was actually written.
+          # Most real bluebooks write it first (insert_at 0); one (a
+          # ScheduledPayment corpus member) writes it after a reference_to
+          # and an attribute — this matches either, and whatever a person
+          # hand-writing `attribute field, Type` at that exact point,
+          # the way this used to be required, would have produced.
+          attributes.insert(insert_at, attributes.pop)
+          vo.attributes.flat_map do |attribute|
+            identity_paths_for_attribute(attribute, value_objects, context_name,
+                                         "#{field}.#{attribute.name}", [target])
+          end
+        end
+
+        def identity_paths_for_attribute(attribute, value_objects, context_name, path, visited)
+          if attribute.list?
+            raise Malformed,
+                  "#{context_name}'s identity member #{path} is a list — an identity member must be scalar"
+          end
+          if attribute.optional?
+            raise Malformed,
+                  "#{context_name}'s identity member #{path} is optional — an identity must be wholly known"
+          end
+
+          return [path] if attribute.reference?
+
+          nested = value_objects.find { |value_object| value_object.hecks_name.to_s == attribute.type.to_s }
+          return [path] unless nested
+
+          if visited.include?(nested.hecks_name.to_s)
+            cycle = [*visited, nested.hecks_name.to_s].join(" -> ")
+            raise Malformed, "#{context_name}'s identity value objects form a cycle: #{cycle}"
+          end
+
+          # A BARE FIELD DERIVES ONE SCALAR. `identified_by :ref` (or one leg
+          # of a compound `identified_by :a, :b`) names a single field, and
+          # deriving ITS path only makes sense while every value object along
+          # the way wraps exactly one field itself — the same "single-field
+          # value object" ADR 0025 names this shape after. A multi-field
+          # value object here used to expand silently into every one of its
+          # own fields, minting an unannounced compound key nothing declared.
+          if nested.attributes.size != 1
+            candidates = nested.attributes.map(&:name).join(", ")
+            raise Malformed,
+                  "#{context_name}.identified_by :#{path} names #{nested.hecks_name}, which has " \
+                  "#{nested.attributes.size} field#{'s' unless nested.attributes.size == 1} (#{candidates})"
+          end
+
+          member = nested.attributes.first
+          identity_paths_for_attribute(member, value_objects, context_name, "#{path}.#{member.name}",
+                                       [*visited, nested.hecks_name.to_s])
+        end
+      end
+    end
+  end
+end

data/lib/hecks/bluebook/dsl/binding_proxy.rb

@@ -0,0 +1,71 @@
+module Hecks
+  module Bluebook
+    module DSL
+      class BindingProxy
+        def self.namespace(domain, collector)
+          Module.new do
+            define_singleton_method(:const_missing) do |aggregate|
+              BindingProxy.new("#{domain}::#{aggregate}", collector)
+            end
+          end
+        end
+
+        def initialize(fqn, collector)
+          @fqn       = fqn
+          @collector = collector
+        end
+
+        # THE AGGREGATE-SCOPED PORT — `Payments::Payment.port("Gateway") do
+        # ... end`, the same receiver a plain bind like `.persisted_by(...)`
+        # already reaches, because a port belongs to exactly one aggregate
+        # the same way a bind does. A REAL method, not method_missing : its
+        # shape (a name and a block building operations) has nothing to do
+        # with `Bind`, so it does not belong in that generic verb path.
+        def port(name, &block)
+          domain, aggregate_name = @fqn.split("::")
+          aggregate_ir = Hecks.current_registry.bluebook(domain)&.aggregate(aggregate_name) or
+            raise Malformed, "#{@fqn} declares no such aggregate — a port needs one to belong to"
+
+          # See HecksagonBuilder#port's own comment on why this resolver
+          # swap is needed : ConstShim's active resolver is one global for
+          # the whole dynamic extent, and it is currently this file's own
+          # BindingProxy-minting one, which would turn a bare `Pizza` inside
+          # `reference_to Pizza` into another BindingProxy instead of a name.
+          built = ConstShim.with(->(const) { const }) { DomainPortBuilder.build(name, owner: aggregate_name, &block) }
+
+          # A `verb`-shaped port is a plain `Port` — the exact struct
+          # `Hecks.port`'s own top-level method registers, so it goes
+          # through the same `add_port` the registry already answers for
+          # that call. It belongs to no aggregate IR the way an
+          # operations-shaped `DomainPort` does; the aggregate above
+          # was only needed to resolve `owner` for the operations branch.
+          if built.is_a?(Port)
+            Hecks.current_registry.add_port(built)
+            return self
+          end
+
+          built.operations.each do |operation|
+            operation.attributes.select(&:reference?).each { |attribute| attribute.type.declared_in = aggregate_ir }
+          end
+          aggregate_ir.add_port(built)
+          self
+        end
+
+        def method_missing(verb, *args, **kwargs, &block)
+          @collector << Bind.new(
+            aggregate: @fqn,
+            verb:      verb.to_s,
+            adapter:   args.first.to_s,
+            role:      kwargs[:role]&.to_s
+          )
+          block&.call
+          self
+        end
+
+        def respond_to_missing?(_name, _include_private = false) = true
+
+        def to_s = @fqn
+      end
+    end
+  end
+end