hecks 0.2.0 → 1.0.0

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

@@ -0,0 +1,1087 @@
+require_relative "word_gate"
+module Hecks
+  module Bluebook
+    module DSL
+      class BluebookBuilder
+        GRAMMAR_CONTEXT = "Bluebook"
+
+        include WordGate
+
+        attr_reader :classification
+
+        def initialize(name, version: nil)
+          @name       = name
+          @version    = version
+          @aggregates       = []
+          @read_models      = []
+          @policies         = []
+          @process_managers = []
+          # THE ROOT of the CHAPTER-WIDE given pool — one level wider
+          # than `AggregateBuilder`'s own `@entity_named_givens` (S10
+          # extended across an aggregate's whole entity tree, earlier
+          # this arc). See `#given`'s own comment for what this closes.
+          @chapter_named_givens = {}
+          # EVERY BARE CHAPTER-GIVEN REFERENCE THIS CHAPTER'S OWN FILES
+          # LEFT UNRESOLVED SO FAR — threaded into every aggregate the
+          # same way `@chapter_named_givens` is. See
+          # `AggregateBuilder#pending_chapter_given`'s own comment for
+          # what queues here and `#resolve_pending_chapter_givens!`,
+          # below, for where it drains.
+          @chapter_pending_givens = []
+          # ONE LEVEL WIDER STILL — the CHAPTER-WIDE, ENTITY-SCOPED pool
+          # (the piece analogue of `@chapter_named_givens`, above). See
+          # `EntityBuilder#given_impl`'s own comment for what this
+          # closes; `docs/implemented/resolution-rules/
+          # chapter-entity-given.md` for the full algorithm.
+          @chapter_entity_named_givens   = {}
+          @chapter_entity_pending_givens = []
+        end
+
+        # Chapter metadata belongs to the composed folder, not whichever file
+        # happened to sort first. A later concept file may be the one carrying
+        # the version header; adopt it into the already-open chapter builder.
+        # Two different versions are a real contradiction, not load order.
+        def adopt_version(version)
+          return if version.nil?
+          if @version && @version.to_s != version.to_s
+            raise Malformed, "#{@name} declares both version #{@version.inspect} and #{version.inspect}"
+          end
+
+          @version = version
+        end
+        private :adopt_version
+
+        def vision(value)
+          # moved to the language: Vision invariant, on Chapter.Declare
+
+          @vision = value
+        end
+
+        # A domain's own identity can change — this names what it used to
+        # be, so the storage layer can recognize its own history under the
+        # old name instead of minting a brand-new lineage from nothing.
+        def formerly_known_as(value) = @formerly_known_as = value.to_s
+
+        # A SUB-LANGUAGE NAMES WHERE IT LANDS. ADR 0026's own seam: the core
+        # grammar does not name its extension points, so this chapter names
+        # ITSELF onto them instead — the core contexts (e.g. "Query",
+        # "ReadModel") whose own admitted words this chapter's `Syntax`
+        # aggregate contributes rows for. Variadic, and accumulating across
+        # calls the same reason `identified_by`/`group_by` are: nothing here
+        # requires one call to name every context at once.
+        # RENAMED FROM `attaches_to` — item #13's full metaprogrammed
+        # dispatch (slice 4c). Not bootstrap-reachable (only sub-language
+        # chapters like Paging use it; the CORE chapters never describe
+        # themselves with it).
+        def attaches_to_impl(*contexts) = (@attaches_to ||= []).concat(contexts.map(&:to_s))
+
+        def core       = @classification = :core
+        def supporting = @classification = :supporting
+        def generic    = @classification = :generic
+
+        # `@chapter_named_givens` is threaded into every aggregate this
+        # chapter builds — see `AggregateBuilder#given`'s own comment
+        # for the sharing this enables; NOT a new top-level DSL word
+        # itself (an aggregate's own EXISTING `given` already both
+        # declares locally and write-throughs here as a side effect,
+        # the identical shape `EntityBuilder#given`'s own write-through
+        # to its owner aggregate's pool already takes — no new spelling
+        # for "declare a precondition," one level wider, same word).
+        # RENAMED FROM `aggregate` — item #13's full metaprogrammed
+        # dispatch (slice 4c). Bootstrap-reachable (every core/attached
+        # chapter's own top-level shape is written with it), so also
+        # named in GenericDispatch::BOOTSTRAP_CALLS_FALLBACK.
+        def aggregate_impl(name, &block)
+          @aggregates << AggregateBuilder.build(name, chapter_named_givens:          @chapter_named_givens,
+                                                      chapter_pending_givens:        @chapter_pending_givens,
+                                                      chapter_entity_named_givens:   @chapter_entity_named_givens,
+                                                      chapter_entity_pending_givens: @chapter_entity_pending_givens, &block)
+        end
+
+        # `read_model` is the word (ADR 0025 reverts `report` — the IR
+        # construct, the registry API, and the docs filename all said
+        # `read_model` the whole time; no era was ever minted under
+        # `report`, so this is history and source agreeing again). `report`
+        # stays answered under `MetaValidator.shadow_parsing?` (S0a's own
+        # bridge) for the same reason `has_many` does — frozen era text
+        # that used it must keep booting; live source refuses it, naming
+        # the replacement.
+        def read_model(name, &block)
+          # A read model gathers heads from SEVERAL aggregates, so no single head
+          # declares it — the chapter does. Its owner is stamped in `build`, where
+          # the chapter namespace exists.
+          @read_models << ReadModelBuilder.build(name, &block)
+        end
+
+        def report(name, &block)
+          return read_model(name, &block) if MetaValidator.shadow_parsing?
+
+          raise Malformed, "report is gone — read_model is the word now"
+        end
+
+        def policy(name, &block)
+          @policies << PolicyBuilder.build(name, &block)
+        end
+
+        def process_manager(name, &block)
+          @process_managers << ProcessManagerBuilder.build(name, &block)
+        end
+
+        def build
+          # The chapter is the top of the construct chain — `Bluebook` is a
+          # ROOT, and its constructor stamps every aggregate and read model with
+          # itself as owner, so every `hecks_fqn` below resolves by walking up
+          # to it. No constants are installed at load time : the public door is
+          # a per-boot projection, installed by `Loader.bind_runtime` once a
+          # dispatcher exists to close over (facade/surface.rb).
+          bluebook = Bluebook::Chapter.new(name: @name, version: @version, vision: @vision,
+                                           aggregates: @aggregates,
+                                           read_models: @read_models,
+                                           policies: @aggregates.flat_map(&:policies) + @policies,
+                                           process_managers: @process_managers,
+                                           classification: @classification,
+                                           formerly_known_as: @formerly_known_as,
+                                           attaches_to: @attaches_to || [])
+
+          # SAME REASON, SAME GATE — a bare chapter-given may still be
+          # pending (see `AggregateBuilder#pending_chapter_given`) if a
+          # file that would resolve it hasn't loaded yet; resolving now
+          # would see the same incomplete `@chapter_named_givens`
+          # `validate_assembled!` below would. Deferred to
+          # `MetaValidator.judge_deferred!` the same way, and BEFORE
+          # `validate_assembled!` there — nothing downstream should ever
+          # read an unresolved placeholder's fields.
+          resolve_pending_chapter_givens! unless MetaValidator.deferring?
+          resolve_pending_chapter_entity_givens! unless MetaValidator.deferring?
+
+          # A CHAPTER MAY BE SPLIT ACROSS FILES (see `self.build`'s own
+          # comment). Every check below needs the WHOLE chapter present —
+          # a hop, a projection, a correlation key or an event shape can
+          # equally name a construct declared in a file that has not
+          # loaded yet, and `@aggregates`/`@process_managers` here are
+          # only ever as complete as whatever has loaded SO FAR. So,
+          # exactly like `MetaValidator.call` below, this is skipped
+          # while `MetaValidator.defer` is loading the chapter's files
+          # and run once instead — by `MetaValidator.judge_deferred!`,
+          # against the fully assembled chapter — after the last one
+          # loads. A single-file chapter (still the common case) never
+          # sees `deferring?` true here at all, so its own checks still
+          # run inline, exactly as before.
+          self.class.validate_assembled!(bluebook) unless MetaValidator.deferring?
+
+          # The language judges the bluebook, in the language. Last, so the
+          # meta-domain sees a fully built IR — the whole-document rules need
+          # every declaration present, which is why they cannot be givens fired
+          # at declaration time.
+          MetaValidator.call(bluebook)
+        end
+
+        # THE OTHER HALF OF A CHAPTER-WIDE `given` REFERENCE —
+        # `AggregateBuilder#pending_chapter_given` recognised an
+        # unresolved bare reference and deferred it here, unable to
+        # check further: a later file in this SAME chapter might still
+        # declare the real thing. Runs once every file has loaded,
+        # against the now-complete `@chapter_named_givens` pool — the
+        # IDENTICAL lookup `reference_named_chapter_given` already does,
+        # just late enough to see every aggregate's own declarations,
+        # not only the ones loaded before the referencing one.
+        #
+        # MUTATES each placeholder `Given` IN PLACE rather than
+        # replacing it — it is already embedded, by Ruby object
+        # reference, in the referencing aggregate's own `preconditions`
+        # and in any command (same aggregate) that separately
+        # bare-referenced the same description, so there is nothing
+        # downstream holding a second, now-stale copy to update.
+        # Instance-level (not `self.`, unlike `validate_assembled!`) —
+        # unlike that battery, this needs `@chapter_named_givens` itself,
+        # which only exists on the builder instance still open for this
+        # chapter (`MetaValidator.judge_deferred!` reaches it via
+        # `registry.bluebook_builder(name)`, guaranteed already present).
+        def resolve_pending_chapter_givens!
+          @chapter_pending_givens.each do |entry|
+            resolved = resolve_pending_chapter_given(entry)
+            entry[:placeholder].description = resolved.description
+            entry[:placeholder].canonical   = resolved.canonical
+            entry[:placeholder].predicate   = resolved.predicate
+          end
+          @chapter_pending_givens.clear
+        end
+
+        def resolve_pending_chapter_given(entry)
+          description = entry[:description]
+          candidates  = RuleReference.resolve_owner_keyed(@chapter_named_givens, description)
+
+          if entry[:declared_by]
+            candidates[entry[:declared_by]] ||
+              raise(Malformed,
+                    "#{entry[:aggregate]}'s given #{description.inspect} names no precondition " \
+                    "#{entry[:declared_by]} declares in this chapter — #{entry[:declared_by]} " \
+                    "either hasn't declared #{description.inspect}, or declared_by: named the " \
+                    "wrong aggregate")
+          elsif candidates.size == 1
+            candidates.values.first
+          elsif candidates.empty?
+            raise(Malformed,
+                  "#{entry[:aggregate]}'s given #{description.inspect} names no precondition " \
+                  "any aggregate in this chapter ever declares — declare it once with a block " \
+                  "(some aggregate's own given(#{description.inspect}) { ... })")
+          else
+            raise(Malformed,
+                  "#{entry[:aggregate]}'s given #{description.inspect} is ambiguous in this " \
+                  "chapter — #{candidates.keys.join(', ')} each declare a DIFFERENT predicate " \
+                  "under this same description; name which one with declared_by: (e.g. " \
+                  "given(#{description.inspect}, declared_by: #{candidates.keys.first}))")
+          end
+        end
+        private :resolve_pending_chapter_given
+
+        # THE ENTITY-SCOPED ANALOGUE, one level down — see
+        # `#resolve_pending_chapter_givens!`'s own comment; identical
+        # shape, resolved against `@chapter_entity_named_givens` instead.
+        def resolve_pending_chapter_entity_givens!
+          @chapter_entity_pending_givens.each do |entry|
+            resolved = resolve_pending_chapter_entity_given(entry)
+            entry[:placeholder].description = resolved.description
+            entry[:placeholder].canonical   = resolved.canonical
+            entry[:placeholder].predicate   = resolved.predicate
+          end
+          @chapter_entity_pending_givens.clear
+        end
+
+        def resolve_pending_chapter_entity_given(entry)
+          description = entry[:description]
+          candidates  = RuleReference.resolve_owner_keyed(@chapter_entity_named_givens, description)
+
+          if entry[:declared_by]
+            candidates[entry[:declared_by]] ||
+              raise(Malformed,
+                    "#{entry[:entity]}'s given #{description.inspect} names no precondition " \
+                    "#{entry[:declared_by]} declares in this chapter — #{entry[:declared_by]} " \
+                    "either hasn't declared #{description.inspect}, or declared_by: named the " \
+                    "wrong piece")
+          elsif candidates.size == 1
+            candidates.values.first
+          elsif candidates.empty?
+            raise(Malformed,
+                  "#{entry[:entity]}'s given #{description.inspect} names no precondition " \
+                  "any piece in this chapter ever declares — declare it once with a block " \
+                  "(some piece's own given(#{description.inspect}) { ... })")
+          else
+            raise(Malformed,
+                  "#{entry[:entity]}'s given #{description.inspect} is ambiguous across the " \
+                  "chapter's own pieces — #{candidates.keys.join(', ')} each declare a DIFFERENT " \
+                  "predicate under this same description; name which one with declared_by: (e.g. " \
+                  "given(#{description.inspect}, declared_by: #{candidates.keys.first.inspect}))")
+          end
+        end
+        private :resolve_pending_chapter_entity_given
+
+        # EVERY WHOLE-CHAPTER CHECK, IN ONE PLACE — the battery `#build`
+        # used to run inline, now a pure function of an assembled
+        # `Bluebook::Chapter` so `MetaValidator.judge_deferred!` can run
+        # it too, once, on a chapter whose files have ALL loaded (see
+        # `#build`'s own comment for why that split exists at all).
+        # Public, not `private_class_method`'d, for exactly that second
+        # caller — `MetaValidator` needs to reach this with no builder
+        # instance in hand, only the chapter `judge_deferred!` already
+        # read back out of the registry.
+        def self.validate_assembled!(bluebook)
+          # moved to the language: an attribute type is a reference to its Shape,
+          # so an undeclared value object fails reference resolution
+          validate_reference_value_objects!(bluebook.aggregates)
+          validate_correlation_keys!(bluebook.process_managers, bluebook.aggregates)
+          validate_no_bidirectional_references!(bluebook.aggregates)
+          unless MetaValidator.shadow_parsing?
+            validate_event_shapes!(bluebook.aggregates)
+            validate_with_projections!(bluebook.policies, bluebook.process_managers, bluebook.aggregates)
+          end
+
+          # Every hop AggregateBuilder#seal_query_field recognised and
+          # deferred gets checked for real here — the earliest point a
+          # hop CAN be checked, for exactly the reason
+          # validate_no_bidirectional_references! above already gives:
+          # `Bluebook.new` just stamped `hecks_owner` on every
+          # aggregate, so `Reference#resolve` finally has a chapter to
+          # walk. Before this line every target in the file (including
+          # ones declared ABOVE the aggregate doing the asking) would
+          # have resolved to nil.
+          infer_hop_query_arguments!(bluebook)
+          validate_query_hops!(bluebook)
+
+          # Same precondition, same reason: a `projects` declaration's
+          # own reference cannot resolve until every aggregate in the
+          # chapter is real and owner-stamped (S12, ADR 0025).
+          validate_projected_fields!(bluebook)
+        end
+
+        # AN ENTITY COMMAND MAY NOT NAME ITSELF AS ITS ROOT.
+        #
+        # That is the whole of what is left here, and it needs saying plainly
+        # because the sentence this used to raise — "references must target
+        # aggregate heads" — was never what it checked.
+        #
+        # `CommandBuilder#reference_to` sets `references` ONLY when the target's
+        # bare name equals the owner's ; anything else becomes a reference
+        # ATTRIBUTE. So on an aggregate command `references` is always a copy of
+        # that aggregate's own name, and looking it up in an index of aggregates
+        # is a TAUTOLOGY — that branch never refused anything and structurally
+        # could not. Verified across all eight golden chapters before deleting it.
+        #
+        # On a PIECE's command the owner is the entity, and an entity is not a
+        # head, so what this actually refuses is `reference_to <its own name>`
+        # written inside `entity do … end`. A piece is reached THROUGH its
+        # aggregate ; a command on one addresses the aggregate, never the piece.
+        #
+        # Reference ATTRIBUTES are the language's business now — offered as the
+        # head's own id and resolved as references, so `Aggregate.Reference` and
+        # `Command.Reference` refuse an undeclared head with no predicate at all.
+        def self.validate_reference_value_objects!(aggregates)
+          heads = aggregates.map(&:hecks_name)
+
+          violations = aggregates.flat_map do |aggregate|
+            aggregate.entities.flat_map do |entity|
+              entity.commands.filter_map do |command|
+                next unless command.references
+                next if heads.include?(command.references.to_s)
+
+                "#{aggregate.hecks_name}.#{entity.hecks_name}.#{command.hecks_name} names itself as its root"
+              end
+            end
+          end
+
+          return if violations.empty?
+
+          raise Malformed,
+                "an entity command is addressed through its aggregate; #{violations.uniq.join('; ')}"
+        end
+
+        # EVENTS ARE FIRST-CLASS BY CONVENTION, NOT BY DECLARATION (ADR
+        # 0025, "events and reactions" — "a domain event is a value
+        # object with its own attributes, not a label"). No new `event
+        # do ... end` construct exists to hand-author and keep in step
+        # with every emitting command by hand — an event's own known
+        # shape IS whichever command(s) declare `emits` for its name, and
+        # this is the ONE thing that has to hold for that convention to
+        # mean anything: every command that emits a given name has to
+        # agree on what it carries. An event is one fact; a fact does not
+        # carry two different truths depending on who is telling it.
+        #
+        # STRUCTURAL fields only (name/type/list/optional) — `pattern:`/
+        # `admits:`/`default:` are refinements ON a field, not a second
+        # claim about what the payload holds, so two emitting commands
+        # are free to differ there without actually disagreeing about
+        # the event's own shape.
+        def self.validate_event_shapes!(aggregates)
+          event_emitters(aggregates).each do |event_name, pairs|
+            next if pairs.size == 1
+
+            shapes = pairs.map { |(owner, command)| event_shape(command, owner_aggregate(owner, aggregates)) }.uniq
+            next if shapes.size == 1
+
+            named = pairs.map { |(owner, command)| "#{owner}.#{command.hecks_name}" }.sort
+            raise Malformed,
+                  "#{event_name.inspect} is emitted with different shapes by #{named.join(' and ')} — " \
+                  "an event is one fact, and every command that emits it must declare the same fields"
+          end
+        end
+
+        # THE "EXPENSIVE HALF" the ADR names: "with: { account: :account }
+        # projecting into a reaction that has no declared contract ...
+        # breaks at dispatch rather than at load." Checked here, now that
+        # `validate_event_shapes!` (above) guarantees at most one real
+        # shape per event name, and command references being first-class
+        # (`Naming.command_ref`) means the TARGET side is a real
+        # resolvable command, not a string that might be a typo.
+        #
+        # SAME-CHAPTER ONLY, ON PURPOSE — a `with:` whose source event or
+        # target command lives outside this chapter (an `across` policy
+        # reacting to another domain's event entirely) is silently left
+        # unchecked rather than refused: there is nothing here yet to
+        # check it against, and "unresolvable" is not the same claim as
+        # "wrong."
+        #
+        # A FOR_EACH POLICY'S SOURCE ISN'T THE EVENT AT ALL — a fan-out
+        # `with:`'s symbols read the QUERY ROW `for_each` answers
+        # (FreezeAccountsOnSuspension's own comment: "`account` is the
+        # key the fan-out merges for each row"), which this has no shape
+        # for; the SOURCE half is skipped for those, the TARGET half
+        # (does the dispatched command actually declare the field) still
+        # runs, since that half is true regardless of where the value
+        # came from.
+        def self.validate_with_projections!(policies, process_managers, aggregates)
+          lookup = command_lookup(aggregates)
+          heads  = correlation_heads(process_managers)
+
+          policies.each do |policy|
+            next if policy.with_spec.to_a.empty?
+
+            source_event = policy.for_each.to_s.empty? ? policy.on_event : nil
+            check_with_spec!(policy.trigger_command, source_event, policy.with_spec, lookup,
+                             "#{policy.name}'s trigger", aggregates, heads)
+          end
+
+          process_managers.each do |pm|
+            pm.handlers.each do |handler|
+              handler.dispatches.each do |dispatch|
+                next if dispatch.with_spec.to_a.empty?
+
+                check_with_spec!(dispatch.command_name, handler.event_type, dispatch.with_spec, lookup,
+                                 "#{pm.name}'s dispatch #{dispatch.command_name}", aggregates, heads, pm: pm)
+              end
+            end
+          end
+        end
+
+        # `pm:` is present only for a process manager's own dispatch — a
+        # saga leg's source symbol resolves against the CURRENT triggering
+        # event first, same as a policy, but falls all the way back to the
+        # saga's own MEMORY when the current event does not carry it
+        # (`SagaInterpreter#dispatch_args`, its own last `else`) — and
+        # memory starts as the OPENING event's payload
+        # (`SagaInterpreter#instance = { ..., memory: event.payload }`,
+        # never updated after), never the leg's own. Settlement's own
+        # comment names exactly this: "the credit leg reads a destination
+        # no event carried" — `AccountDebited` never declares `:reference`,
+        # only `TransferRequested` (`pm.starts_on`) does, and that is
+        # where the value is genuinely still coming from.
+        def self.check_with_spec!(command_ref, event_name, with_spec, lookup, label, aggregates, correlation_heads, pm: nil)
+          target        = lookup[command_ref]
+          source_shape  = event_name && event_shape_for(event_name, aggregates)
+          memory_shape  = pm && event_shape_for(pm.starts_on, aggregates)
+          correlation   = pm && pm.correlates_by && pm.correlation_head
+          # A POLICY'S SOURCE ALSO CARRIES THE EMITTER'S OWN IDENTITY —
+          # `PolicyInterpreter#emitter_identity`, the runtime half of this.
+          # An entity command's event never declares its aggregate's
+          # identity (it arrives through `reference_to`, not an
+          # `attribute`), so before this a policy on `KnightCaptured` could
+          # not spell `with: { label: :label }` at all — "reads :label off
+          # KnightCaptured, which does not declare it" — and chess's
+          # AdvancePly grew optional, unread attributes just to survive a
+          # wholesale forward. Policies only: a saga leg's own source is
+          # `SagaInterpreter#dispatch_args`, which merges no such thing.
+          identity_sources = pm.nil? && event_name ? event_identity_heads_for(event_name, aggregates) : []
+
+          with_spec.each do |field, source|
+            raise Malformed, "#{label}'s with: names #{field.inspect}, which #{command_ref} does not declare" if target && !command_declares?(target, field, aggregates, correlation_heads)
+
+            next unless source.is_a?(::Symbol)
+            next if source == correlation
+            next if identity_sources.include?(source)
+            next unless source_shape || memory_shape
+
+            found = [source_shape, memory_shape].compact.any? { |shape| shape.any? { |name, *| name == source } }
+            next if found
+
+            raise Malformed, "#{label}'s with: reads :#{source} off #{event_name.inspect}, which does not declare it"
+          end
+        end
+
+        # A command's OWN `reference_to` (bare, no `as:`) never lands in
+        # `attributes` — `CommandBuilder#reference_to`'s self-reference
+        # branch sets `command.references` instead (S2), and mints no new
+        # field at all. What addresses it is not one name but the SAME
+        # SET `CommandInterpreter::ArgumentGate#refuse_unknown_arguments`
+        # already accepts at dispatch time — `:id`, the owning aggregate's
+        # own `identity_heads` (real corpus proof — `Account.Debit`
+        # dispatched everywhere as `number: ...`, `Account`'s own
+        # `identified_by`), AND `Naming.reference_key(command.references)`
+        # (real corpus proof — `FreezeAccountsOnSuspension`'s `for_each`
+        # fan-out, whose own comment reads "`account` is the key the
+        # fan-out merges for each row it answers"). Both are simultaneously
+        # legal there, not context-dependent alternatives, so both are
+        # legal here : this mirrors that gate rather than re-deriving a
+        # narrower rule that would refuse one of two real, already-shipped
+        # dispatch conventions.
+        def self.command_declares?(command, field, aggregates, correlation_heads)
+          return true if command.attributes.any? { |a| a.name == field }
+          return true if field == :id
+          return true if correlation_heads.include?(field)
+          return false unless command.references
+
+          referenced = aggregates.find { |a| a.hecks_name == command.references }
+          return false unless referenced
+
+          referenced.identity_heads.include?(field) || Naming.reference_key(command.references) == field
+        end
+
+        # THE FOURTH addressing key `ArgumentGate#refuse_unknown_arguments`
+        # accepts, alongside `:id`/`identity_heads`/`reference_key` — every
+        # saga in THIS domain's own `correlates_by` head, carried through
+        # every dispatch as pure passthrough (Settlement's own comment:
+        # "`reference:` carries the correlation forward... this is pure
+        # passthrough, not an addressing key"). A command declaring none of
+        # its attributes named this is not a gap; the correlation key rides
+        # through commands that never read it, same as it does at runtime.
+        def self.correlation_heads(process_managers)
+          process_managers.filter_map { |pm| pm.correlates_by && pm.correlation_head }
+        end
+
+        # Every command this chapter declares, an aggregate's own AND
+        # every entity nested inside one, paired with a name for what
+        # declares it — shared by `validate_event_shapes!` and
+        # `validate_with_projections!`'s own command lookup, the same
+        # reach `HecksagonBuilder#commands_in` needs one level up (S8).
+        def self.each_command(aggregates)
+          return enum_for(:each_command, aggregates) unless block_given?
+
+          aggregates.each do |aggregate|
+            aggregate.commands.each { |command| yield aggregate.hecks_name, command }
+            aggregate.entities.each do |entity|
+              entity.commands.each { |command| yield "#{aggregate.hecks_name}.#{entity.hecks_name}", command }
+            end
+          end
+        end
+
+        # NOT MEMOISED — this used to be `@event_emitters ||=` on the
+        # builder instance, which is safe for a one-file chapter but
+        # wrong for one split across several: the FIRST file's build()
+        # call would compute and cache it from whatever `@aggregates`
+        # held at that moment, and every later file's own validation
+        # would keep reading that same stale snapshot, silently missing
+        # any command a later file adds. Recomputed fresh every call
+        # instead — this walks the whole chapter once per `#build`, not
+        # a hot path worth memoising at that cost.
+        def self.event_emitters(aggregates)
+          each_command(aggregates).each_with_object(Hash.new { |h, k| h[k] = [] }) do |(owner, command), index|
+            command.emits.each { |event_name| index[event_name] << [owner, command] }
+          end
+        end
+
+        # STRUCTURAL, NOT NOMINAL. Two commands on two different
+        # aggregates that both `emits "SameEvent"` are free to type a
+        # field through two DIFFERENT, locally-scoped wrapper value
+        # objects (e.g. one aggregate's own `value: SomeText` vs
+        # another's `value: OtherText`, exactly the per-aggregate "own
+        # text VO" convention every aggregate in this grammar already
+        # follows for everything from `RuleText` to `FieldRef`) without
+        # actually disagreeing about the event's shape — comparing
+        # `a.type` by NAME would flag that as a violation for no real
+        # reason: an event is one fact, and two isomorphic wrapper types
+        # tell an identical one. So a value-object type is unwrapped to
+        # its OWN attribute shape (recursively — a wrapper could itself
+        # wrap another) before comparing, and only a primitive type
+        # (nothing left to unwrap) or two VOs that truly differ once
+        # unwrapped still counts as a real mismatch. `owner` carries the
+        # type's `value_object` lookup — a command's own attributes only
+        # know their type's NAME, never the aggregate that declared it,
+        # and two sibling aggregates in one chapter each keep a
+        # same-named VO private to themselves, so the unwrap has to ask
+        # the SAME aggregate the field's own command belongs to, never a
+        # neighbor's.
+        def self.event_shape(command, owner)
+          command.attributes.map { |a| [a.name, unwrap_shape(owner, a.type.to_s), a.list?, a.optional?] }.sort
+        end
+
+        def self.unwrap_shape(owner, type_name, seen = [])
+          return type_name if owner.nil? # owner couldn't be resolved -- compare by name, same as before this unwrap existed
+          return type_name if Attribute::PRIMITIVES.include?(type_name)
+          return type_name if seen.include?(type_name) # a self-referential VO bottoms out on its own name, not an infinite unwrap
+
+          shape = owner.value_object(type_name)
+          return type_name unless shape # not this owner's own VO (a reference type, say) -- nothing further to unwrap
+
+          shape.attributes.map { |a| [a.name, unwrap_shape(owner, a.type.to_s, seen + [type_name]), a.list?, a.optional?] }.sort
+        end
+
+        # `owner` (from `each_command`) is a plain STRING — the aggregate's
+        # `hecks_name` alone, or `"Aggregate.Entity"` for an entity's own
+        # command. Either way the VALUE OBJECTS a command's fields can be
+        # typed with are the AGGREGATE's own (`Entity` carries no
+        # `value_object` lookup of its own — the whole rest of this file
+        # already resolves hop/type lookups only at the aggregate level,
+        # e.g. `validate_hop_tail!`'s `target.value_object(type)`), so only
+        # the first segment ever matters here.
+        def self.owner_aggregate(owner, aggregates)
+          aggregates.find { |a| a.hecks_name == owner.to_s.split(".").first }
+        end
+
+        def self.event_shape_for(event_name, aggregates)
+          pairs = event_emitters(aggregates).fetch(event_name.to_s, [])
+          return nil if pairs.empty?
+
+          owner_name, command = pairs.first
+          event_shape(command, owner_aggregate(owner_name, aggregates))
+        end
+
+        # The identity heads of the aggregate that emits `event_name` — an
+        # entity's event is stamped with its OWNING aggregate's identity
+        # (`Event#id` is the parent's), so an owner spelled "Game.Knight"
+        # answers Game's heads.
+        def self.event_identity_heads_for(event_name, aggregates)
+          pairs = event_emitters(aggregates).fetch(event_name.to_s, [])
+          return [] if pairs.empty?
+
+          owner_name, = pairs.first
+          aggregate = owner_aggregate(owner_name, aggregates)
+          return [] unless aggregate
+
+          heads = aggregate.identity_heads.map(&:to_sym)
+          # AN ENTITY'S EVENT ALSO CARRIES THE PIECE'S OWN IDENTITY — the
+          # args a piece was addressed by are the args its event announces
+          # (`Emission#emit`: `payload: args`), so `id`-shaped heads are
+          # genuinely there at runtime even though no `attribute` line on
+          # the entity command declares them.
+          entity_names = owner_name.to_s.split(".").drop(1)
+          entity = entity_names.reduce(aggregate) { |owner, name| owner&.entities&.find { |e| e.hecks_name == name } }
+          heads + (entity ? entity.identity_heads.map(&:to_sym) : [])
+        end
+
+        def self.command_lookup(aggregates)
+          each_command(aggregates).each_with_object({}) do |(owner, command), index|
+            index["#{owner}.#{command.hecks_name}"] = command
+          end
+        end
+
+        # A REFERENCE RING IS NOT A MODELLING CHOICE, IT IS A MISSING ONE
+        # — a DDD aggregate is a consistency boundary precisely because
+        # something outside it can only ever point IN, by id, never the
+        # other way. A caller must be able to reason about one aggregate
+        # alone ; a ring back to where it started means no aggregate in
+        # it is a boundary anyone can reason about without the rest of
+        # the ring, and the whole ring is really one aggregate wearing
+        # several names.
+        #
+        # Checked at the bluebook level, not inside `AggregateBuilder`
+        # itself, because seeing a cycle needs every end declared — an
+        # aggregate finishes building long before it can know whether
+        # some later aggregate in the same file points back at it.
+        #
+        # ACYCLIC WITHIN A CHAPTER (ADR 0025, "References") — widened
+        # from the direct pair (A -> B -> A) this used to catch alone to
+        # any ring, however long (A -> B -> C -> A), the same DFS
+        # coloring a reference graph needs for any cycle. A cross-chapter
+        # reference is UNREACHABLE here rather than unchecked:
+        # `Reference#resolve` is scoped to its own chapter by
+        # construction, so a target this chapter never declares is a
+        # dangling name, not an edge — `edges.key?` below is what keeps
+        # the walk from ever leaving this chapter's own aggregates.
+        # Self-reference stays legal (`parent.parent.name` for a
+        # hierarchy is real and safe) — excluded the same way the
+        # direct-pair check already excluded it.
+        def self.validate_no_bidirectional_references!(aggregates)
+          edges = aggregates.each_with_object({}) do |aggregate, index|
+            index[aggregate.hecks_name] = aggregate.reference_targets.uniq.reject { |target| target == aggregate.hecks_name }
+          end
+
+          cycle = find_reference_cycle(edges)
+          return unless cycle
+
+          ring = "#{cycle.join(' -> ')} -> #{cycle.first}"
+          raise Malformed,
+                "reference cycle: #{ring} — an aggregate points at another by id, and a " \
+                "ring back to where it started means no aggregate in it is a boundary " \
+                "anyone can reason about alone ; break the ring, or let one side be found " \
+                "through a query instead of a reference pointing back"
+        end
+
+        # Plain DFS with a visiting/done coloring, over the reference
+        # graph THIS chapter's own aggregates declare. Returns the ring
+        # itself (in the order it closes), or nil.
+        def self.find_reference_cycle(edges)
+          state = {}
+
+          edges.each_key do |start|
+            cycle = reference_cycle_from(start, edges, state, [])
+            return cycle if cycle
+          end
+
+          nil
+        end
+
+        def self.reference_cycle_from(node, edges, state, path)
+          return nil if state[node] == :done
+          return path[path.index(node)..] if state[node] == :visiting
+
+          state[node] = :visiting
+          path.push(node)
+
+          edges[node].each do |target|
+            next unless edges.key?(target) # a name this chapter never declares is dangling, not an edge
+
+            found = reference_cycle_from(target, edges, state, path)
+            return found if found
+          end
+
+          path.pop
+          state[node] = :done
+          nil
+        end
+
+        # THE OTHER HALF OF A HOP — AggregateBuilder#seal_query_field
+        # recognised the HEAD of a dotted where-field that names one of
+        # its own references and deferred it here, unable to check
+        # further: it cannot yet resolve what the reference points AT.
+        # This runs once every aggregate exists in one chapter, so it
+        # can.
+        #
+        # Only WHERE clauses ever reach here — a hop on ORDER BY is
+        # refused outright, immediately, back in seal_query_field
+        # itself (that answer never needed the target's shape).
+        #
+        # AN ENTITY'S OWN QUERIES DID reach `EntityBuilder#reference_to`
+        # (added after this comment first claimed otherwise — S9, ADR
+        # 0025) without ever reaching HERE: tier-1 sealing
+        # (`AggregateBuilder#query_surfaces`) already recognises a hop
+        # on an entity's own field and DEFERS it exactly like an
+        # aggregate's, but nothing ever walked entity queries at tier 2
+        # to check the deferral — a bad hop, or even a well-formed one,
+        # built silently and then matched nothing at runtime
+        # (`QueryInterpreter#entity_rows` reads an element's fields by
+        # literal hash key, never follows a reference). Refused outright
+        # here instead of taught to follow the hop for real: no corpus
+        # member needs an entity query to cross a reference yet, and a
+        # named refusal beats a runtime that resolves nothing while
+        # looking like it might.
+        def self.validate_query_hops!(bluebook)
+          bluebook.aggregates.each do |aggregate|
+            aggregate.queries.each do |query|
+              query.wheres.each do |clause|
+                next unless QuerySpecification::HopPath.hop_head?(clause.field, aggregate.attributes)
+
+                validate_hop_clause!(aggregate, query, clause)
+              end
+            end
+
+            aggregate.entities.each { |entity| refuse_entity_query_hops!(aggregate, entity) }
+          end
+        end
+
+        # The chapter-wide half of AggregateBuilder's local query-argument
+        # inference. A hop cannot resolve while its aggregate is still being
+        # built; here every Reference has an owner and target, so a symbolic
+        # comparison can inherit the type of the scalar it compares without a
+        # duplicate query-local declaration.
+        def self.infer_hop_query_arguments!(bluebook)
+          bluebook.aggregates.each do |aggregate|
+            aggregate.queries.each do |query|
+              query.wheres.each do |clause|
+                name = clause.value
+                next unless name.is_a?(Symbol)
+                next if query.attribute(name)
+                next unless QuerySpecification::HopPath.hop_head?(clause.field, aggregate.attributes)
+
+                plan = QuerySpecification::HopPath.plan(clause.field, aggregate.attributes)
+                next if plan.refusal || plan.hops.empty?
+
+                target = plan.hops.last.target
+                head, *nested = plan.tail.to_s.split(".")
+                leaf = if nested.empty? && target.lifecycle&.field.to_s == head
+                         Attribute.new(name: name, type: String)
+                       else
+                         root = target.attributes.find { |candidate| candidate.name.to_s == head }
+                         found = root && QuerySpecification::FieldPath.leaf_attribute(root, nested) do |type|
+                           target.value_object(type)
+                         end
+                         found && Attribute.new(name: name, type: found.type, list: found.list?)
+                       end
+                next unless leaf
+
+                query.attributes << leaf
+              end
+            end
+          end
+        end
+
+        def self.refuse_entity_query_hops!(aggregate, entity)
+          entity.queries.each do |query|
+            query.wheres.each do |clause|
+              next unless QuerySpecification::HopPath.hop_head?(clause.field, entity.attributes)
+
+              raise Malformed,
+                    "#{aggregate.hecks_name}::#{entity.hecks_name}.#{query.hecks_name} asks about " \
+                    "#{clause.field}, which hops through #{entity.hecks_name}'s own reference — " \
+                    "an entity query does not follow a hop the way an aggregate's own does; ask " \
+                    "through the aggregate's own query instead, or open the target directly"
+            end
+          end
+        end
+
+        def self.validate_hop_clause!(aggregate, query, clause)
+          plan = QuerySpecification::HopPath.plan(clause.field, aggregate.attributes)
+
+          case plan.refusal
+          when :unresolvable
+            # HopPath.plan pushes even an unresolved hop onto `hops`
+            # before reporting this, specifically so `target_name` —
+            # real, known at declaration, independent of whether
+            # `resolve` succeeded — is always here to name.
+            raise Malformed,
+                  "#{aggregate.hecks_name}.#{query.hecks_name} asks about #{clause.field}, " \
+                  "which hops to #{plan.hops.last.target_name}, which this chapter never " \
+                  "declares — a hop into an aggregate this chapter cannot see resolves to " \
+                  "nothing, and a where that resolves to nothing matches nothing and " \
+                  "refuses nothing"
+          when :too_deep
+            raise Malformed,
+                  "#{aggregate.hecks_name}.#{query.hecks_name} asks about #{clause.field}, " \
+                  "whose hop chain reaches #{QuerySpecification::HopPath::MAX_HOPS} " \
+                  "references deep without landing — a chain this long is refused as a " \
+                  "likely mistake, not a structural limit"
+          end
+
+          target = plan.hops.last.target
+          validate_hop_tail!(aggregate, query, clause, target, plan.tail)
+        end
+
+        # The same three-way answer seal_query_field gives for its OWN
+        # aggregate's fields — landing on a real scalar (fine), landing
+        # on a value object (refused by name), or naming nothing at all
+        # (refused by name) — asked instead of the hop's TARGET aggregate,
+        # since that is whose shape the tail actually has to answer for.
+        def self.validate_hop_tail!(aggregate, query, clause, target, tail)
+          name, *nested = tail.to_s.split(".")
+          attribute = target.attributes.find { |candidate| candidate.name.to_s == name }
+          return validate_hop_comparator!(aggregate, query, clause, target, attribute, nested) if
+            nested.empty? && (attribute || target.lifecycle&.field.to_s == name)
+          return validate_hop_comparator!(aggregate, query, clause, target, attribute, nested) if
+            nested.any? && attribute &&
+            QuerySpecification::FieldPath.scalar_leaf?(attribute, nested) { |type| target.value_object(type) }
+
+          if nested.any? && attribute &&
+             !QuerySpecification::FieldPath.leaf_attribute(attribute, nested) { |type| target.value_object(type) }.nil?
+            raise Malformed,
+                  "#{aggregate.hecks_name}.#{query.hecks_name} asks about #{clause.field}, " \
+                  "which hops to #{target.hecks_name} and then asks about #{tail}, which " \
+                  "lands on a value object, not a scalar — a dotted query path ends on a " \
+                  "scalar member, or the engines answer it differently"
+          end
+
+          raise Malformed,
+                "#{aggregate.hecks_name}.#{query.hecks_name} asks about #{clause.field}, " \
+                "which hops to #{target.hecks_name} and then asks about #{tail}, which " \
+                "#{target.hecks_name} never declares — a query over a field that does " \
+                "not exist matches nothing and refuses nothing"
+        end
+
+        # A WHERE hop with an ordered comparator is legitimate ("client
+        # whose balance > 500") — AggregateBuilder#seal_ordered_comparator
+        # already deferred this exact check for the same reason every
+        # other hop check is deferred, and this is where it gets asked,
+        # against the hop's TARGET instead of the querying aggregate.
+        def self.validate_hop_comparator!(aggregate, query, clause, target, attribute, nested)
+          return unless AggregateBuilder::ORDERED_COMPARATORS.include?(clause.op.to_s.to_sym)
+          return if attribute &&
+                    QuerySpecification::FieldPath.numeric?(attribute, nested) { |type| target.value_object(type) }
+
+          held = attribute ? "holds no number" : "is the lifecycle field, which holds text"
+          raise Malformed,
+                "#{aggregate.hecks_name}.#{query.hecks_name} compares #{clause.field} with " \
+                "#{clause.op} after hopping to #{target.hecks_name}, but the field it lands " \
+                "on #{held} — an ordered comparison needs a numeric field, and over " \
+                "anything else the adapters answer differently or not at all"
+        end
+
+        # THE TARGET HALF of `projects` validation (S12, ADR 0025) —
+        # `AggregateBuilder#seal_projected_fields` already checked the
+        # LOCAL half at declare time (the reference names a real
+        # `reference_to` on THIS aggregate); this checks the reference
+        # actually resolves to a real aggregate in this chapter, and
+        # that aggregate really declares `remote_field` as a scalar.
+        #
+        # Reuses `QuerySpecification::HopPath` rather than re-deriving
+        # hop resolution a second way — `"reference/remote_field"` is
+        # the same single-hop shape a query's own `/`-spelled hop
+        # resolves, even though `projects`'s own DSL spelling is dotted
+        # (`from: :"customer.status"`): two constructs, two spellings,
+        # one resolution primitive. A single hop can never reach
+        # HopPath::MAX_HOPS, so :too_deep is structurally unreachable
+        # here and is not special-cased.
+        def self.validate_projected_fields!(bluebook)
+          bluebook.aggregates.each do |aggregate|
+            aggregate.projected_fields.each { |field| validate_projected_field!(aggregate, field) }
+          end
+        end
+
+        def self.validate_projected_field!(aggregate, field)
+          plan = QuerySpecification::HopPath.plan("#{field.reference}/#{field.remote_field}", aggregate.attributes)
+
+          if plan.refusal == :unresolvable
+            raise Malformed,
+                  "#{aggregate.hecks_name}.projects :#{field.name} reads through :#{field.reference}, " \
+                  "which hops to #{plan.hops.last.target_name}, which this chapter never declares — " \
+                  "a projection through an aggregate this chapter cannot see resolves to nothing"
+          end
+
+          target = plan.hops.last.target
+          remote_attribute = target.attributes.find { |candidate| candidate.name.to_s == plan.tail }
+
+          # THE WORKED EXAMPLE ITSELF (ADR 0025) reads through a
+          # LIFECYCLE field — banking's Customer.status is `lifecycle
+          # :status`, never a plain `attribute` — the same fallback
+          # validate_hop_tail! already gives a query's own hop tail. A
+          # lifecycle field is always a plain string by construction ;
+          # nothing further to check once it matches by name.
+          return if remote_attribute.nil? && target.lifecycle&.field.to_s == plan.tail
+
+          # A PROJECTION MAY CHAIN THROUGH ANOTHER PROJECTION (S12, ADR
+          # 0025's own boundary rule, followed through) — `target`'s
+          # OWN projected fields live in `projected_fields`, a
+          # separate list from `attributes`, so a match there is
+          # invisible to the check above even though it names a real,
+          # always-current, stored field. `Transfer.projects
+          # :source_customer_status, from: :"source.customer_status"`
+          # reads Account's own already-projected `customer_status`
+          # this way — Account is one hop from Customer, Transfer is
+          # one hop from Account, and neither aggregate needs to know
+          # about the other's target two hops away. A projected
+          # field's remote value is always a scalar by construction
+          # (`RebuildSweep.remote_value` never copies a reference, a
+          # value object, or a list), so nothing further to check once
+          # it matches by name — same reasoning the lifecycle
+          # fallback just above already applies.
+          return if remote_attribute.nil? && target.projected_fields.any? { |f| f.name.to_s == plan.tail }
+
+          unless remote_attribute
+            raise Malformed,
+                  "#{aggregate.hecks_name}.projects :#{field.name} reads #{target.hecks_name}'s own " \
+                  "#{plan.tail.inspect}, which #{target.hecks_name} never declares"
+          end
+
+          return if projectable_scalar?(target, remote_attribute)
+
+          raise Malformed,
+                "#{aggregate.hecks_name}.projects :#{field.name} reads #{target.hecks_name}'s own " \
+                "#{plan.tail.inspect}, which is not a scalar — a projected field copies a single " \
+                "value, never a reference, a value object, or a list"
+        end
+
+        def self.projectable_scalar?(target, attribute)
+          !attribute.list? && !attribute.reference? && target.value_object(attribute.type).nil?
+        end
+
+        # `correlates_by` NAMES A SCALAR, NOW CHECKED RATHER THAN TRUSTED.
+        #
+        # ProcessManagerBuilder#validate! already refuses a bare, undotted
+        # spelling — a SYNTACTIC guarantee that the declaration cannot leave
+        # the question open. It cannot go further: a process manager is built
+        # in isolation, before this chapter's aggregates exist to check
+        # against. Here, with the whole document assembled, the dotted path
+        # is walked for real — against whichever command actually emits an
+        # event this process manager reacts to — so a path that still lands
+        # on a value object is refused before the runtime ever has to decide
+        # what a non-scalar correlation key even means: a saga keys off this
+        # value directly, and a value object carries no guaranteed-stable
+        # identity to key on the way a scalar does.
+        #
+        # A command that does not declare the path's first segment at all is
+        # silently skipped, not refused — correlation has two other fallback
+        # tiers below the payload dig (a correlation stamp, then the emitting
+        # aggregate's own reference key; saga_interpreter/correlation.rb), so
+        # an absent field is not this check's business. Only a field that
+        # resolves, and resolves to something other than a scalar, is.
+        def self.validate_correlation_keys!(process_managers, aggregates)
+          process_managers.each do |pm|
+            next unless pm.correlates_by
+
+            reason = correlation_key_violation(pm, aggregates)
+            next unless reason
+
+            raise ProcessManagerBuilder::InvalidProcessManager,
+                  "#{pm.name} correlates_by #{pm.correlates_by.inspect}, but #{reason}"
+          end
+        end
+
+        def self.correlation_key_violation(pm, aggregates)
+          head, *rest = pm.correlates_by.to_s.split(".")
+          events = reacted_events(pm)
+
+          emitting_commands(events, aggregates).each do |owner, command|
+            attribute = command.attributes.find { |a| a.name == head.to_sym }
+            next unless attribute
+
+            reason = list_or_scalar_violation(owner, attribute, rest)
+            return reason if reason
+          end
+
+          nil
+        end
+
+        def self.reacted_events(pm)
+          ([pm.starts_on, pm.ends_on] + pm.handlers.map(&:event_type))
+            .compact
+            .reject { |event| event == ProcessManager::REFUSED }
+            .map { |event| event.to_s.split("::").last }
+            .uniq
+        end
+
+        def self.emitting_commands(events, aggregates)
+          aggregates.flat_map do |aggregate|
+            commands = aggregate.commands + aggregate.entities.flat_map(&:commands)
+            commands.select { |command| (command.emits.map(&:to_s) & events).any? }
+                    .map { |command| [aggregate, command] }
+          end
+        end
+
+        def self.list_or_scalar_violation(owner, attribute, segments)
+          return "#{attribute.name} is a list — a correlation key must name one instance's own field, " \
+                 "not a whole collection" if attribute.list?
+
+          walk_scalar(owner, attribute.type.to_s, segments)
+        end
+
+        # Walks the remaining dotted segments through nested value objects.
+        # `type_name` starts as the head attribute's own declared type ; each
+        # step either bottoms out at a real scalar (nil — no violation) or
+        # names why it cannot: still a value object with no more path left,
+        # a value object this domain never declared, a field that value
+        # object does not have, or a segment left over after already
+        # reaching a scalar.
+        def self.walk_scalar(owner, type_name, segments)
+          if segments.empty?
+            return nil if Attribute::PRIMITIVES.include?(type_name)
+
+            return "#{type_name} is a value object, not a scalar — name one of its own fields, " \
+                   "e.g. #{type_name.downcase}.value"
+          end
+
+          return "#{type_name} is already a scalar — #{segments.join('.')} has nothing left to reach" if Attribute::PRIMITIVES.include?(type_name)
+
+          shape = owner.value_object(type_name)
+          return "#{type_name} is not a value object this domain declares" unless shape
+
+          segment, *rest = segments
+          attribute = shape.attributes.find { |a| a.name == segment.to_sym }
+          return "#{type_name} has no field #{segment.inspect}" unless attribute
+          return "#{type_name}.#{segment} is a list — a correlation key must name one instance's own field, " \
+                 "not a whole collection" if attribute.list?
+
+          walk_scalar(owner, attribute.type.to_s, rest)
+        end
+
+        # A CHAPTER MAY BE DECLARED IN SEVERAL FILES, meant to merge into ONE
+        # domain — `lib/hecks/language/bluebook/*.bluebook` all open
+        # `Hecks.bluebook "Bluebook" do ... end`. Each `Hecks.bluebook` call used to
+        # mint a fresh builder, so a second file with the same chapter name
+        # silently replaced the first's aggregates instead of adding to them.
+        #
+        # The registry now holds the builder OPEN across calls : the first file
+        # for a name creates it, every later file for the same name reuses the
+        # same instance, so `@aggregates`/`@read_models` accumulate. `#build` is
+        # safe to call once per file on the same builder — it constructs a fresh
+        # `Bluebook` from whatever is currently held and re-`Namespace.install`s
+        # over the previous one, so the LAST file's call leaves every aggregate
+        # seen so far reachable, and each call's IR is a strict superset of the
+        # one before. `Registry#add_bluebook` still simply stores by name — with
+        # this in place, "last write wins" is the cumulative, correct write.
+        def self.build(name, version: nil, &block)
+          registry = Hecks.current_registry
+          builder  = registry ? registry.bluebook_builder(name) { new(name, version: version) } : new(name, version: version)
+          builder.__send__(:adopt_version, version)
+          # A bare constant in a bluebook — `attribute :name, PizzaName` — is a NAME,
+          # not a reference to something Ruby has heard of. `const_missing` hands
+          # over a `ConstShim::ScopedConstant` (S0b, const_shim.rb's own comment),
+          # and that is still the whole answer for a bare name: `Attribute` spells
+          # it with `to_s`, so the `TypeName` wrapper this used to build existed
+          # only long enough to be stringified. The concept still has a home — the
+          # language declares `value_object "TypeName"` — it just needed no Ruby
+          # class of its own. A Module rather than a Symbol is what also lets
+          # `Account::Debit`/`admits: Account::LedgerDirection` answer their OWN
+          # `::` — a plain Symbol cannot.
+          resolver = ->(const) { ConstShim::ScopedConstant.for(const) }
+          ConstShim.with(resolver) { builder.instance_eval(&block) } if block
+          builder.build
+        end
+      end
+    end
+  end
+end