@wairon/cli 5.1.1-dev.102 → 5.1.1-dev.104

package/docs/standards/architecture.md

@@ -0,0 +1,953 @@
+# wairon — Architecture Standard
+
+> The canonical, language-neutral definition of wairon's architecture model: the
+> building blocks, the patterns, the entity/behavior rules, the spec-tree layout,
+> and the rules a conformant design must obey. The schema, the validator, the MCP
+> tools, and the SDD skills all reconcile to this document.
+>
+> This standard describes *how to structure any system designed with wairon*. It
+> is language-agnostic — the structural rules apply to any OOP or structured
+> language. Language-specific realization (concurrency, memory) lives in
+> [language-bindings.md](language-bindings.md).
+
+---
+
+## 1. Two orthogonal axes: containment and refinement
+
+Everything is positioned on **two independent axes**. Conflating them is the most
+common source of confusion, so they come first.
+
+- **Containment (the tree → folders):** what contains what —
+  `System ⊃ Subsystem ⊃ Component ⊃ owned blocks`.
+- **Refinement (level of detail → files):** for *one* component, increasing
+  detail — `structure → interface → implementation + narrative`.
+
+The labels L0–L5 walk down containment first, then refinement:
+
+| Label | Axis | Artifact |
+|---|---|---|
+| **L0 System** | containment | `.index.yaml` (specs root) |
+| **L1 Subsystem** | containment | `<subsystem>/.index.yaml` |
+| **L2 Component** | containment leaf **and** refinement root (structure) | `<component>/.index.yaml` |
+| **L3 Interface** | refinement | `<component>/.interface.yaml` |
+| **L4 Implementation** | refinement | `<component>/.implementation.yaml` |
+| **L5 Narrative** | refinement (a *sub-field* of L4, not a separate file) | steps inside `.implementation.yaml` |
+
+---
+
+## 2. The spec tree on disk
+
+```
+.wai/specs/
+  .index.yaml                      ← L0  system: vision, boundaries, global requirements
+  types/                           ← shared value objects (Money, Address, Email …)
+    money.yaml
+  <subsystem>/                     ← L1  a bounded context / service
+    .index.yaml                    ← L1  public interface(s) the service exposes
+    types/                         ← entities/aggregates owned by THIS subsystem (e.g. order.yaml)
+    <component>/                   ← L2  a standalone block OR a pattern facade
+      .index.yaml                  ← L2  structure: role/stereotype, responsibility, owns, dependsOn
+      .interface.yaml              ← L3  contract: method signatures (referencing types by id)
+      .implementation.yaml         ← L4+L5  sourcePath, concurrency variant, per-method narrative
+      <owned-block>/               ← (only if this component is a pattern) a private owned block
+        .index.yaml
+        .interface.yaml
+        .implementation.yaml
+```
+
+**File contents:**
+- `.index.yaml` (L2) — identity, **role/stereotype**, **responsibility description**, `owns` (member blocks; patterns only), `dependsOn` (collaborators). No methods, no narratives.
+- `.interface.yaml` (L3) — **method signatures** (name, params/returns referencing types, optional HTTP/gRPC/event binding). Interfaces are **method contracts only** — fields are implementation, never part of the contract.
+- `.implementation.yaml` (L4+L5) — `sourcePath`, the **concurrency variant**, and the **per-method narrative** (L5 steps).
+
+Different agents own different refinement files: the **architect** owns
+`.index.yaml` + `.interface.yaml`; the **implementer** owns `.implementation.yaml`. (Legacy undotted names — `system.yaml`, `subsystem.yaml`, `component.yaml`, `interface.yaml`, `implementation.yaml` — still load; `wairon doctor --fix` migrates them.)
+
+---
+
+## 3. Building blocks (the primitives)
+
+These ten are the **only** atomic component types.
+
+| Block | Role | Owns | Default interface |
+|---|---|---|---|
+| **Portal** | The service's **inbound** front door — receives external commands and dispatches them | the inbound transport binding + dependencies | `dispatch(command) → result` |
+| **Orchestrator** | **Logic as a flowchart** — ordered steps, conditions, branches, loops and parallel dispatch — calling its collaborators, other Orchestrators among them, nested down to leaf logic. One or more cohesive methods sharing those collaborators, realized as a class or a module of functions. Its `dependencyClass` (`pure` or `read`) bounds what it may depend on; unset, it is a **workflow** | dependencies only | domain-specific methods |
+| **Supervisor** | Owns the **set** of live Actors and their lifecycle; may supervise other Supervisors | runtime state (the set of Actors) | `start/stop`, find-or-start by id |
+| **Actor** | Owns **one** live thing — a session, a connection, a timer, or an entity instance — and handles its events one at a time. Its methods are full flowcharts, not shims | that thing's runtime state | `start/stop` + its messages |
+| **Store** | Authoritative state boundary for **one aggregate** | the aggregate's entities | `get`, `list`, `write`, `remove` |
+| **Index** | Read path; optimized derived lookup maps over a Store | derived lookup state | `get_by_<key>`, `list_by_<key>` |
+| **Query** | A **Repository member** for computed reads over its Store — results the backend computes per call (an aggregate, a ranking, a report) | dependencies only | one method per computed read |
+| **Registry** | Write path (CUD) for one aggregate | dependencies only | `create`, `update`, `delete` |
+| **Adapter** | A **client to an external system/protocol** (DB, FS, HTTP, gRPC, WS, message bus). The **only** block doing raw external I/O | a connection/resource handle | protocol-shaped (`query`/`execute`, `send`/`receive`, `publish`…) |
+| **Observer** | Subscribes to events and forwards them to one workflow | a subscription | `on_event(event)` |
+
+### State vs dependencies vs workflow
+
+Every component holds **dependencies** (its injected collaborators). What differs
+is whether it also owns **state** or **workflow**:
+
+| Owns domain/runtime state | Owns workflow | Owns only dependencies (stateless paths and boundaries) |
+|---|---|---|
+| Store (authoritative), Index (derived); Supervisor, Actor (runtime) | Orchestrator | Registry, Query, Portal, Observer, Adapter* |
+
+\*The Adapter holds a *resource handle* (a connection), not domain state.
+"Stateless" means **no domain state** — not "no fields."
+
+An **Orchestrator holds only its collaborators**: whatever one call computes lives
+on that call's stack. Runtime state — a live session, a timer, the set of running
+instances — belongs to an **Actor** or a **Supervisor**, whose own methods run
+logic over it; domain state belongs to a **Store**. State an Orchestrator keeps
+for itself between calls is hidden state, invisible to every rule in this
+standard.
+
+### Dependency rules
+
+- **Pure logic** — an Orchestrator with `dependencyClass: pure` — may be used by
+  **every** block: a Store, Registry, Index or Query validating or deriving
+  through it, an Adapter translating with it. It depends only on other pure
+  Orchestrators, so it never drags data, I/O or held state into its caller.
+- **Portal** is the inbound boundary; no *local* component may depend on it. It
+  dispatches to Orchestrators and may consult Repository and Index **read** faces
+  for passthrough reads; a write reached from a Portal's narrative or dispatch
+  table — or a lifecycle change — routes through an Orchestrator
+  (`PORTAL_WRITE_SHORTCUT`). It never depends
+  on a Store, Registry, Adapter or Query. A Portal may **message a Supervisor by
+  id** to reach a live Actor (below).
+  The **one** exception to "no component depends on a Portal" is a
+  **cross-subsystem client Adapter** (see below): from its side the remote
+  subsystem's Portal is an external front door, so it depends on that Portal —
+  never on the remote subsystem's internals.
+- **Observer** may depend on exactly **one** Orchestrator *or* one Supervisor
+  entry; it may use a message-bus **Adapter** to subscribe. No component depends on it.
+- **Orchestrator** — its `dependencyClass` bounds its dependencies. The class is
+  checked on every hop, so by induction nothing below pure logic reaches data or I/O:
+  - `pure` depends only on pure Orchestrators;
+  - `read` also depends on read Orchestrators, Repositories, Indexes and Adapters,
+    and never calls their write methods (judged once facade methods carry effect tags);
+  - unset, it is a **workflow**: it may depend on whatever the other rules allow —
+    Repositories, other Orchestrators, Adapters, and Supervisors to reach live Actors.
+- **Supervisor** depends on its Actors, Orchestrators, Adapters and other
+  Supervisors, and never on a presentation block
+  (`ARCHITECTURE_VIOLATION_SUPERVISOR_DEP`). Its held state has two homes:
+  - its **supervision state** — restart counts, live sets, run brackets — is a
+    Store or Registry it **owns** (`owns`, one hop): private, read and written in
+    full. A Supervisor owns nothing else (`SUPERVISOR_CONTAINMENT`), and nobody
+    else depends on that state: a component that does is the intruder, and the
+    finding is reported on it, naming the Supervisor and the state
+    (`SUPERVISION_STATE_INTRUSION`);
+  - **shared data** — a data component it does not own — it reaches only through
+    `read`- and `lifecycle`-effect methods. A write to an entity's fields goes
+    through a workflow: the Supervisor depends on the Orchestrator that does it
+    (`SUPERVISOR_WRITE_SHORTCUT`, judged from its narrative call steps against the
+    callee's declared effect; a callee that declares no effect is refused too).
+    A read made on behalf of a request still passes; in practice the paired write
+    forces the workflow out, and the read moves with it.
+- **Actor** — a live Actor is **reached through its supervision**. A caller finds
+  it by id through a **Registry its Supervisor maintains** — owns, or keeps
+  through lifecycle calls (register and unregister a live handle) — so the
+  caller's `dependsOn` lists that Registry and the Actor: the lookup hop the code
+  really takes. Depending on a Supervisor that supervises the Actor (messaging it
+  by id) is reach too; anything else is `ACTOR_REACHED_WITHOUT_SUPERVISOR`. A
+  Supervisor depending on the Actor supervises it. The Actor itself may use
+  Repository facades, Indexes, Adapters and Orchestrators, and never depends on
+  its own Supervisor.
+- **Method effects.** A contract method's `effect` says what it does to held
+  state: `read` observes it; `write` modifies an entity's domain fields;
+  `lifecycle` creates, destroys, or (un)registers an entity's existence or
+  membership and never modifies its domain fields. The lifecycle effect is
+  **closed under composition**: a lifecycle method calls only read and lifecycle
+  methods besides its construction and local steps (`LIFECYCLE_CALLS_WRITE`).
+  Both `write` and `lifecycle` are mutations: a Portal's read shortcut covers
+  neither, and a durable Store needs a hydration read-back for both.
+- **Store** may depend on an **Adapter** (its backend), another Store, or pure
+  logic. It is depended *upon* by Registries, Indexes and Queries — never the reverse.
+- **Index** is a read-only **projection over a Store**: it shares the Store's
+  per-entry references. It may depend on its Store, an Adapter for cold reads, and
+  pure logic. It **never** depends on a Registry — read and write paths are decoupled.
+  As an **exceptional case**, when one Index's projection is itself worth
+  re-presenting another way, a derived Index may depend on another Index — only
+  one **owned by the same Repository**, and never in a cycle of Index edges
+  (`ARCHITECTURE_VIOLATION_INDEX_DEP` otherwise). It is not a way to chain lookups:
+  reach for it only when a second parse of the Store would duplicate state the
+  first Index already holds (wairon's export tables over its scanned specs).
+  An Index that **scans** a tree of project roots may read the configuration of
+  the root it is walking, through that root's own binding — the same per-root
+  binding it already resolves that root's files through — and nothing else of
+  it (wairon's spec scan reads each root's `project.yaml` for the project graph).
+- **Query** computes reads over its own Repository's Store: it depends **only on
+  its Store, a backend Adapter or pure logic**, and it lives only inside a
+  Repository (`UNOWNED_QUERY`).
+- **Registry** is the write path to a Store; it may depend on its Store, Adapters
+  (e.g. a message-bus Adapter to publish), and pure logic for validation. It does
+  **not** depend on Indexes, and it drives no workflow.
+- **Adapter** is a dependency-sink toward the system (its other side is external).
+  It may use pure logic, but never a Store or any other Orchestrator. Workflows,
+  read logic, Actors, Supervisors and the data blocks may depend on an Adapter to
+  reach an external system. **Cross-subsystem client Adapter** (the one
+  Adapter that depends *outward* on a Portal): a sibling subsystem is "potentially
+  external" — designed to be swappable for a network service — so a local client
+  Adapter is its only reachable surface, and it depends on the **remote subsystem's
+  Portal** (its front door), which dispatches inward. It still may not depend on the
+  remote subsystem's Orchestrators or Stores directly.
+- The narrative `call` graph and the `dependsOn` graph must both be **DAGs** (no
+  cycles), and every narrative `call` must target a method on a declared dependency.
+
+### The Orchestrator, precisely
+
+An Orchestrator is **logic as a flowchart**: ordered steps, conditions, branches,
+loops, parallel dispatch and calls, where a call to another Orchestrator nests its
+flowchart, down to leaf Orchestrators of native logic. It holds its collaborators
+(constructor-injected, or closed over by a module of functions — see
+[language-bindings.md](language-bindings.md)) and **no domain state**: each
+call's state lives on the stack, so a shared instance is concurrency-safe and its
+instantiation lifecycle is irrelevant to correctness.
+
+- **All logic is a flowchart.** It is an Orchestrator's, an Actor's or
+  Supervisor's own method over its own runtime state, or a type method for
+  arithmetic over one value's own fields (§5). A verdict, a translation, a derived
+  view and a multi-step workflow are all Orchestrators; what tells them apart is
+  what they may depend on. **Local control flow** (guard clauses, a Registry's
+  `exists?` check) is fine anywhere.
+- **The dependency class** says what an Orchestrator may reach:
+
+  | `dependencyClass` | May depend on | Typical shapes |
+  |---|---|---|
+  | `pure` | pure Orchestrators only | a verdict over supplied facts, a format translation, a view or text derived from values it is handed |
+  | `read` | pure and read Orchestrators, Repositories, Indexes, Adapters — reading, never writing | a derivation that loads its own source, a lookup that gathers the facts for a verdict |
+  | unset (a workflow) | whatever the other rules allow | writes, sequences, transactions, reaching live Actors |
+
+  Choose the narrowest class that holds, and split to reach it: loading in the
+  workflow and deriving in pure logic makes the derivation replayable and
+  testable without fakes.
+- **One or more cohesive methods.** An Orchestrator is the unit that holds one
+  set of collaborators. Its methods belong together when they serve one purpose
+  and share those collaborators; groups of methods that call no common component
+  are two Orchestrators in one.
+- Dependencies are injected at construction; method arguments carry only the
+  **workflow input**. For "same workflow over many repositories," prefer one
+  instance per concrete dependency behind a shared interface; use dynamic dispatch
+  only when the choice is genuinely per-call.
+
+### Self-initiated and long-running processes
+
+There is **no separate "Process/Daemon" block** — owning one live thing *is* the
+**Actor's** defining job. An Actor holds the runtime state of one live thing — a
+session, a connection, a timer, or an entity instance — and handles its events one
+at a time, so that state has a single writer. Its **methods are full
+flowcharts**: an Actor method may call Repository facades, Indexes, Adapters and
+Orchestrators itself. A part moves into an Orchestrator only when something else
+also triggers it, or when it is too big to read in place. An Actor never passes
+itself to an Orchestrator; it passes values from its state and applies the result.
+The **Supervisor** owns the *set* of Actors and their lifecycle (find-or-start by
+id, stop, restart with backoff, unload idle instances) and may supervise other
+Supervisors, so a runtime is a **tree** under one root Supervisor. The state that
+bookkeeping needs is the Supervisor's own supervision state — a Store or Registry
+it owns — and the live handles callers look Actors up by are a Registry it
+maintains.
+
+| Need | Modeled as |
+|---|---|
+| cron / scheduler | an Actor owning the timer; each tick is its own method, calling the Orchestrator that does the work |
+| batch job | an Actor owning the iteration, handling each item itself or through an Orchestrator |
+| stream / websocket consumer | an Actor owning the read-loop + an Adapter owning the connection |
+| message-bus consumer | an Observer subscribing via a bus Adapter, or an Actor owning a poll-loop |
+| daemon | an Actor owning a long-running loop |
+
+**Entity Actors.** An Actor may be the live instance of an entity — an auction
+lot, a match, a device twin. The entity type stays data with pure methods; the
+Actor holds one instance's working state and handles its messages; the **Store
+stays authoritative**, loaded when the Actor starts and written through the
+Repository. Use an entity Actor when live in-memory interaction is the point; use
+persisted status plus event-driven workflows when steps are minutes or days
+apart. A live entity Actor is the **only writer of its aggregate** in its
+process: every other writer messages it through its Supervisor (a cluster needs a
+lease or sharded routing to keep one writer).
+
+**Large decisions: the Decider shape.** When an Actor's handling of a message
+holds real decision logic, split it: *decide* — a pure Orchestrator, given the
+state and the message, returns a decision; *persist* — the Actor's method writes
+it in the transaction it owns (§10); *apply* — only after the commit, the Actor
+updates its own state; *act* — it notifies or emits through its collaborators.
+The decision logic is then replayable and testable without the Actor.
+
+**Timers.**
+
+| Timer | Modeled as |
+|---|---|
+| per-instance and loss-safe (a heartbeat, an idle timeout) | state of the Actor that owns the instance |
+| a deadline on an aggregate (a lot's end, an order's payment due) | a **field of the aggregate** (`lot.endsAt`), read by its Index; a sweep Actor handles what falls due |
+| a timer that spans several aggregates | a schedule aggregate with its own Repository, fired by an Actor |
+| a process-wide schedule (a backup every minute) | its own Actor |
+| restart backoff, passivation | the Supervisor |
+| an external scheduler | an Adapter |
+
+A deadline kept as a field survives a restart with nothing preloaded: the sweep
+finds whatever fell due while the process was down.
+
+---
+
+## 4. External I/O and emission — the Adapter
+
+The **Adapter** is the single boundary to the outside world, in two shapes:
+1. **Behind a Store** — a DB/FS Adapter the Store translates domain calls into
+   (`store.get` → `adapter.query(...)` → deserialize → entity).
+2. **Standalone outbound** — an Orchestrator, Actor or Registry calling an
+   HTTP/gRPC/email/message-bus Adapter directly for an external effect. An
+   external tool — a linter run as a process, a policy engine behind an API — has
+   the same shape: an Adapter runs it, and an Orchestrator interprets its output.
+
+**Inbound vs outbound:** the **Portal** is *our* front door (others call us); the
+**Adapter** is *our* client to external systems (we call/connect out, including
+publishing). They are different roles, not duplicates.
+
+**Emission has no special block, and is optional.** Events are a pattern for
+**decoupled reactions** — used when a producer should not know its consumers.
+Realizations, smallest-first:
+- *No bus (in-process):* "emit" is a **fire-and-forget** call — spawn a task/thread
+  that invokes the handler/Observer; the producer does not await it.
+- *Message bus:* emit by calling a **message-bus Adapter** (`publish(topic, event)`)
+  — a normal `dependsOn → Adapter` edge; an Observer subscribes via a bus Adapter.
+- *Simplest systems* may skip events entirely and call the next step directly.
+
+There is no "EventPublisher" block in any case. Event topology is declared, not
+inferred: a component lists the topics it `emits` and `subscribesTo` (L2
+fields), and a MessageBus public interface's endpoints carry a
+`direction: publish | subscribe`. The validator pairs the two by topic —
+an emitted topic nobody consumes is `UNCONSUMED_TOPIC`, a subscription with no
+source is `UNSOURCED_SUBSCRIPTION` (topics with external ends are acknowledged
+via `lint.allow`). Trees that use no events see neither check.
+
+**Ports + implementations (the Carrier rule).** When a capability has multiple
+external providers, define an **interface (a port)** and one Adapter per provider:
+`Carrier` interface ← `DhlAdapter`, `UpsAdapter`. Callers depend on the `Carrier`
+interface; the concrete Adapter is injected at bootstrap. This is composition +
+polymorphism via interfaces — **not** inheritance.
+
+---
+
+## 5. Object modeling: where behavior lives
+
+Entities are **passive state**; behavior belongs to the object that can perform it.
+The test: **can the object do this autonomously, using only its own state?**
+
+- **Yes → an intrinsic method on the entity.** `order.total()`, `order.is_valid()`,
+  `dog.bark()` — pure logic over the entity's own fields (computed values,
+  invariants, self-contained state transitions).
+- **No (it needs an external actor or system) → a method on the acting *block*,
+  taking the entity as an argument.** An order cannot ship itself → a `Carrier`
+  (interface) with `DhlAdapter.ship(order)`. A dog cannot take itself out → a
+  `Caretaker` (an Orchestrator) with `take_out(dog)`.
+
+This is the rich-vs-anemic balance, and it maps onto the blocks: **the Store holds
+the entity; Registries/Orchestrators/Actors/Adapters act *on* it.** It keeps
+domain logic out of passive data and prevents "god entities" that secretly
+orchestrate. Behavior that belongs to **one live instance**, message by message,
+lives on an entity Actor (§3), which still keeps the entity itself passive.
+
+**Writes follow the aggregate.** An owner's value objects (a customer's address)
+are saved with it in one Store call. A separate aggregate (the customer's
+credentials) is written by the workflow, in one transaction when both must land
+together (§10), with the ids generated in the workflow and the dependent
+referencing its owner.
+
+**Composition over inheritance.** The model has no class hierarchies. Variation is
+expressed by **interfaces (ports) with multiple implementations** (the Carrier
+example) and by **composition** (a pattern owns blocks). Inheritance is avoided;
+"is-a" relationships that tempt a hierarchy are almost always "implements-an-
+interface" or "is-composed-of."
+
+---
+
+## 6. Entities and types
+
+Entities and value objects are **first-class type specs** (not one of the ten
+blocks — they are the *data* the blocks operate on). Each is defined **exactly
+once**, scoped by ownership:
+
+- **Shared value objects** with no single owner (Money, Address, Email) → the
+  **system-level `types/`** (sibling to subsystems, under L0).
+- **Entities / aggregates** (Order, Customer) → defined in the **subsystem that
+  owns their lifecycle** (the bounded context whose Store is authoritative):
+  `<subsystem>/types/`.
+
+A type is **never** redefined in multiple places and is **not** a sibling to all
+subsystems. When other subsystems process an Order, they **reference the owner's
+contract** (its L1 public interface) or hold a **local projection/DTO** (an
+anti-corruption view) — they do not share or redefine the owning aggregate. An L3
+interface references types by id, so signatures stay structured (which also makes
+data-model/ER diagrams derivable — see §14).
+
+### Type expressions: one neutral grammar
+
+Every structured type position — a param's `type`, a method's `returns`, a
+field's `type`, a type method's and a signature type's params and returns — is
+written in one language-neutral grammar, so a contract means the same thing to
+every implementation language:
+
+| Form | Meaning |
+|---|---|
+| `string`, `int`, `float`, `bool`, `bytes` | the scalar primitives (`int` and `float` are distinct; width is an L4 concern) |
+| `date`, `datetime`, `duration` | a calendar day, an instant, an elapsed time |
+| `void`, `any` | no value (a whole returns only; `()` is another spelling), and an unconstrained value |
+| `list<T>`, `set<T>`, `map<K, V>` | ordered, unique, keyed; a map key is `string`, `int` or an enum |
+| `T?` | T, or no value |
+| `A \| B` | exactly one of the named types (a primitive or collection never joins a union) |
+| `async T` | completes later with T — at the top of a returns only (`async void`) |
+| `result<T, E>` | completes with T or fails with E — a whole returns only, or under its `async`; T may be `void` |
+| `Invoice`, `billing::invoice`, `Page<T>` | a named type: an entity, value-object, enum or signature type |
+
+"No value" (`T?`) belongs to a type; "may be left out" belongs to a position
+and stays the `optional` flag on a param or field. Both together mean "may be
+left out, and may be explicitly none".
+
+An **enum** is a type of its own kind: a closed, ordered set of named values
+(`values: [{ name, description? }]`), optionally with pure methods. A field
+typed by an enum is a value domain, not a relation, so the ER diagram draws no
+edge to it, and OpenAPI renders it as a string `enum`.
+
+A **named scalar** is a value-object that declares `holds: <primitive>` in
+place of fields: one value under a name (`order_id` holds `string`), a newtype
+or type alias in every language (`type OrderId = string`). It holds one of
+`string`, `int`, `float`, `bool`, `bytes`, `date`, `datetime` or `duration` —
+not a union, not a named type, not `any` — and carries no fields beside it
+(`NAMED_SCALAR_MEMBERS`). It is how a value that needs its own meaning gets a
+name instead of staying a bare `string`, and how a choice that mixed in a
+primitive (`PackSelection | string`) becomes a union of named types
+(`PackSelection | PackPath`). Like an enum it is a value domain: the ER diagram
+draws it compactly with no edge to it, OpenAPI renders it as its primitive's
+schema under its name, and type-shape conformance compares it with the code's
+alias through the language's dialect (`TYPE_HOLDS_MISMATCH`).
+
+`result<T, E>` is the success-or-failure type the mainstream languages with
+typed failures share (Rust's and Swift's `Result`, Kotlin's `Result`, F#'s
+`Result`). In a language without typed failures it is written as its success
+type: in TypeScript the function returns T and throws E, so code conformance
+compares the annotation with T and never with E.
+
+TypeScript spellings (`string[]`, `boolean`, `T | null`, `Promise<T>`,
+`Record<K, V>`) and Rust's `Result<T, E>` and `()` are accepted as input and
+stored in the canonical spelling; the write answers with each respelling.
+`number` is not accepted ("int or float?"). Inline object shapes, inline
+function types, string-literal unions, intersections, utility types and unions
+mixing in a primitive are not in the grammar: each has a named replacement (a
+value-object, a signature type, an enum). The writer refuses each of them; a
+tree that already holds one still loads and stays lockable, with a warning per
+position, and `wairon doctor --fix` rewrites every alias and proposes an `int`
+or an enum where it can. How each form is written in a given language is in
+[Language bindings](language-bindings.md#type-mapping).
+
+---
+
+## 7. Patterns (named compositions)
+
+A **pattern** is a *named, rule-bound composition of building blocks, exposed as
+one component (a facade)*. A pattern owns **only building blocks** — never another
+pattern (§9).
+
+### Repository
+> The data-access-layer component for **one aggregate**. It **owns** exactly one
+> **Store** — the aggregate's entity and everything that must stay consistent with
+> it, across as many tables as that takes — and the faces over it: a **Registry**
+> (write face), one-or-more **Indexes** (read faces), any **Queries** (computed
+> reads), and optionally one **Adapter** (the backend). It contains **no logic of
+> its own**. The Store and Adapter are **private behind** the Registry, Indexes and
+> Queries. Its facade **forwards 1:1** to the Registry (writes) and to the Indexes
+> and Queries (reads); **consumers depend on the Repository facade only** — never
+> the inner blocks.
+
+The Registry's write is `validate → store.write (→ optionally publish)`; it does
+**not** touch the Indexes. Indexes are **projections over the Store** that share its
+per-entry references — a value-content update is seen through the shared reference
+with no index change; only create/delete, or an update that changes an *indexed
+field*, propagates structurally from the Store to its Indexes. Write-time
+constraints (uniqueness, validation) are checked against the **Store** (the
+authoritative source) with pure validation logic — not via an Index.
+
+**Index or Query.** An Index shares the Store's references, so it is never stale.
+A read the backend computes on every call — an aggregate, a ranking, a report the
+database assembles — is not a projection: it is a **Query** member, depending on
+its Store or the backend Adapter (and pure logic). A custom computed query is a
+Query, never an Adapter-backed Index method, and never a consumer→Adapter shortcut.
+
+**Reads across aggregates.** A Repository stays a self-contained unit: a Query
+never depends on another Repository's facade (two such Queries would also form a
+cycle). A join falls into one of three cases:
+- **within one aggregate** — a Store or Query method whose join is its implementation;
+- **across aggregates** with small data or in different databases — a **read
+  Orchestrator** joining the facades' results;
+- **across aggregates in one database, as one statement** — a **read-model
+  Repository** whose Query runs the join through the shared database Adapter.
+  Declaring which aggregates a read model reads is not modelled yet; name them in
+  its description.
+
+**Mapping an ORM.** An object-relational mapper fits the model: the database
+context (Entity Framework's `DbContext`, for one) is the database **Adapter** —
+the connection, change tracking and the transaction context §10 has it issue;
+entity mappings, navigation properties and `Include` joins implement a **Store**;
+cross-aggregate projections and view-mapped keyless entities are read-model
+**Queries**. The one habit the model pushes against is querying the context
+straight from a workflow.
+
+**When the full pattern is overkill — the two sanctioned shapes for held state.**
+The Repository is the RECOMMENDED shape wherever read and write consumers diverge
+or write-time constraints exist. For genuinely simple held state, a **deliberately
+standalone Store** is the sanctioned lightweight form: the state stays visible as a
+component, reachable from the workflow layer (a workflow Orchestrator or an Actor),
+acknowledged with a `lint.allow` reason on the `UNOWNED_STORE` warning — and it
+declares its `durability` like every Store (`durable` | `read-through` |
+`ram-projection` | `cache`). The durability axis is **orthogonal** to the shape:
+an in-memory Repository (`ram-projection` Store inside) and a persisted bare Store
+(`read-through` file backing) are both legal quadrants. What is NEVER sanctioned is
+the third path — folding held state into an Orchestrator's fields; hidden state
+is invisible to every rule in this standard. And a standalone **Registry** is not a
+shape at all: a Registry is the write path to a Store (`REGISTRY_WITHOUT_STORE`) —
+a component that itself holds persisted state is a *Store*, whatever its file I/O
+looks like. A **Query** has no standalone shape either: it lives only inside the
+Repository whose Store it reads (`UNOWNED_QUERY`).
+
+### gateway: a Portal variant
+> A gateway is a **Portal** with the built-in `gateway` variant: a Portal that
+> **authenticates, authorizes, validates or rate-limits before it dispatches**, by
+> calling that logic, and declares its inbound authentication in `auth`. The
+> checks are Orchestrators it depends on — typically read logic that gathers the
+> caller's grants and pure logic that rules on them — never members it owns. A
+> Portal with no ingress checks needs no variant.
+
+The variant carries the implementation guidance (§8). The Gateway *pattern*
+(`Portal + ingress Orchestrator + interceptor Specialists`) is retired
+(`STEREOTYPE_RETIRED`); §8 lists its migration.
+
+### The facade rule (mechanically enforced: `FACADE_FORWARDING`)
+A pattern's facade does **pure 1:1 forwarding with no logic**: **every facade
+method's authored narrative is exactly one `call` step targeting an owned
+member.** More than one step, a `local` step, or a call that leaves the pattern
+means the facade contains logic — a violation. The validator enforces this on
+Repository facades as the `FACADE_FORWARDING` warning
+(lint.allow-suppressible for deliberate exceptions); methods without an
+authored narrative are governed by the detail dial, not this rule.
+
+### No wildcard block
+The model has no residual block for what fits nowhere else. A need that seems to
+fit no block is a decomposition not yet made: held or derived state → a Store,
+Index or Repository; behavior over a value's own fields → a type method; external
+I/O → an Adapter; logic → an Orchestrator with the narrowest `dependencyClass`
+that holds; one live thing → an Actor.
+
+---
+
+## 8. Roles (naming vocabulary)
+
+Below patterns are **roles** — conventional names for Orchestrators (and some
+Indexes) that share a block's rules but carry a documented name + recommended
+interface. They are vocabulary, **not** distinct types.
+
+| Role | Recommended interface | Underlying block |
+|---|---|---|
+| Router | `route(input) → destination` | Index (flat key→target) or a pure Orchestrator (rule-based) |
+| Validator | `validate(x) → Result` | a pure Orchestrator |
+| Scanner | `scan(scope) → findings` | a pure Orchestrator handed the scope, or a read Orchestrator that loads it |
+| Mapper / Compiler / Evaluator | one transform method | a pure Orchestrator |
+
+### The built-in variants
+
+One step above roles sit **variants** — recurring shapes with a machine-recognized
+identity (`variant:` on the component, resolved against the variant registry)
+whose guidance travels into every agent brief. wairon ships a **built-in layer**.
+The registry loads it first, then the global variants (`WAIRON_VARIANTS_DIR`, else
+`~/.wairon/variants`), then the project's `.wai/variants/`; a later layer
+**overrides a variant with the same id**, so a project can reword the built-in
+guidance for its own team. A component's stereotype must equal its variant's
+`base` (`VARIANT_BASE_MISMATCH`).
+
+Variants carry **identity and guidance, not rules**. What an Orchestrator may
+depend on is its `dependencyClass` field, which the validator enforces; one shape
+can sit in either class, which is why the class is a field and not a variant.
+
+| Variant | Base | Class | Shape | Discipline (guidance-enforced) |
+|---|---|---|---|---|
+| `arbiter` | Orchestrator | `pure` | subject + supplied world → deterministic verdict + reasons | NO I/O, no state deps — the caller gathers the world (companion shape: read logic or a workflow gathers, the arbiter rules); no clock/randomness; `idempotent` where it holds |
+| `projector` | Orchestrator | `pure` when handed its source, `read` when it loads it | source model → self-contained derived view (snapshot, graph, artifact, digest) | ≤1 read facade or parameters-only; recomputed per call, owns nothing, writes nothing. NOT an Index: an Index is a maintained read model over an owned Store (or, exceptionally, over another Index of the same Repository) |
+| `composer` | Orchestrator | `pure` when handed its values, `read` when it loads them | templates + values → authored text/file map | returns content, never writes or executes it; degrades gracefully on missing optional inputs |
+| `codec` | Orchestrator | `pure` | format ↔ format, bidirectional | pure whole-value translation; inbound half validates + safety-checks; both directions in one component so the round-trip stays testable |
+| `gateway` | Portal | — | a Portal that authenticates, authorizes, validates or rate-limits before dispatching | calls that logic before it dispatches; declares its inbound auth in `auth`; writes still route through Orchestrators |
+
+A variant is **promoted to a first-class stereotype** only when independent
+projects/packs keep re-registering it, or when its edge rules exceed what guidance
+can express.
+
+### Retired stereotypes
+
+`Specialist` and `Gateway` are retired. A component typed with either is
+`STEREOTYPE_RETIRED` (error) until it is migrated; meanwhile the rules that judge a
+stereotype's shape skip it, so it reports once.
+
+| Retired | Becomes | Migration |
+|---|---|---|
+| `Specialist` | an Orchestrator with the `dependencyClass` its dependencies give it | `wairon doctor --fix` retypes it with that class, and rebases each Specialist-based variant onto Orchestrator |
+| `Gateway` | the Portal it owned, with the `gateway` variant | by hand: (1) the owned Portal becomes the front door, with `variant: gateway`; (2) the Gateway's other members become that Portal's dependencies; (3) its consumers depend on the Portal; (4) delete the Gateway spec — `sdd_rename_component` can then give the Portal the Gateway's id |
+
+**Router vs Portal vs Facade:** a **Router** selects a destination from *dynamic
+input values*; a **Portal** is the *transport* boundary that *uses* a Router/Index
+to dispatch; a **Facade** does *no* routing (the caller already chose the method).
+
+---
+
+## 9. Relationships, composition, and layering
+
+- **`owns` (composition):** a pattern owns its member blocks — **exactly one hop**;
+  patterns never own patterns, blocks own nothing. This makes depth finite (no
+  "L2.5").
+- **`dependsOn` (collaboration):** any component uses others.
+- **Visibility rule:** a component may depend on (a) blocks within its own group,
+  (b) the *facade* of any other group, or (c) any standalone block — **never** a
+  block private to another group.
+- **"Leaf" is graph-specific:** an *ownership-leaf* owns no sub-components (every
+  block); a *dependency-sink* depends on nothing (typically a Store or Adapter). A
+  read Orchestrator is an ownership-leaf that may still depend on a Repository —
+  so it is not a dependency-sink. The two are orthogonal.
+
+**Layering inside a subsystem:**
+```
+Ingress:   Portal  (with the gateway variant when it authenticates, authorizes, validates or rate-limits first)
+Workflow:  Orchestrators: workflows, over read and pure logic
+Data:      Repository = Store (one aggregate) + Registry (write+constraints) + Indexes/Queries (read) + Adapter (private)
+Process:   a Supervisor tree + Actors (sessions, connections, timers, live entity instances)
+```
+
+Workflow Orchestrators **use** Repositories; Repositories never contain a Portal or
+an Orchestrator. **When you are tempted to nest a pattern inside a
+component, promote it to a subsystem (L1)** — composition of *patterns* is an L1
+concern. A **Saga** is therefore an L1-level arrangement (an Orchestrator + an
+Observer + a sibling Repository for its persisted progress), not an L2 pattern.
+
+### Crossing a subsystem boundary: published surfaces and their consumers
+A component in another subsystem may depend only on what that subsystem
+**publishes** in its `publicInterfaces` (`CROSS_SUBSYSTEM_PRIVATE_ACCESS`
+otherwise), and the hop enters through a Portal (`CROSS_SUBSYSTEM_TARGET_NON_PORTAL`).
+An entry may narrow who may use it with **`consumers`**: a list of subsystem ids.
+Without it, any subsystem may depend on the entry; with it, only those listed
+may — a dependency from any other subsystem is `CROSS_SUBSYSTEM_UNLISTED_CONSUMER`
+(an error). When several entries publish the same component, each naming
+consumers, the union of their lists is the whole set.
+
+### Listeners mount Portals
+A Portal's routes are its methods' endpoint bindings; the **listener** that
+hands the Portal its requests declares that with **`mounts`** on its own L2
+spec — each mount names the `portal`, the path `prefixes` it routes there, and
+optionally `via`, the router entry the Portal's file exports for it. An HTTP
+Portal that no listener mounts, and that declares no `mounts` itself (which
+marks it as a listener, even with `mounts: []`), is `UNMOUNTED_PORTAL`: nothing
+would ever reach it. `mounts: []` is honest only on a Portal that IS the whole
+HTTP server and serves just its own routes; a service with several HTTP
+Portals has one listener that mounts each of them. A prefix matches whole path segments (`/web` covers
+`/web/admin`, not `/webhooks`). Only HTTP endpoints are judged.
+
+### Forwarding a contract: `signatureFrom`
+A method that takes its params and returns from somewhere else names the source
+in **`signatureFrom`** instead of restating them: a signature type
+(`billing.change_listener`, or `alias::name` across projects), or a contract
+method `component.method` that its component reaches along a `dependsOn` or
+`owns` edge — the shape of a facade or a forwarding Portal. The loader resolves
+the source on every read, so the forwarding method can never drift from it.
+
+---
+
+## 10. Implementation variants (L4) — language-neutral rule
+
+Structure (L2/L3) is *what it is*; the concurrency strategy (L4) is *how it is
+realized*. **Concurrency is only a concern when shared mutable state is actually
+accessed concurrently** — a single-threaded or simple system uses plain ownership
+and skips this section entirely; do not add concurrency machinery a system does not
+need. When it *is* needed, the rule, stated neutrally:
+
+| Updating… | Strategy |
+|---|---|
+| one value, simple read-modify-write (counter, flag, single reference) | a single **atomic operation** — wait-free, ordered, cheaper than a lock |
+| one location, compound update with no atomic equivalent (rare) | a compare-and-set retry (race-free; avoid over large/variable values) |
+| multiple locations that must stay mutually consistent (a map; or two fields) | **serialize writes (in order) + wait-free snapshot reads** (the default) |
+
+**Default = wait-free reads / in-order serialized writes:** readers read a snapshot
+reference without blocking; writers are serialized so they apply in order; on
+commit the snapshot reference is swapped for all readers. Values are held by
+reference so the snapshot swap is pointer-only (zero-copy).
+
+The concrete primitives per language (atomics, swappable references, sharded maps,
+persistent structures, transactional backends) live in
+[language-bindings.md](language-bindings.md). The standard itself prescribes only
+the *strategy*, so it remains language-agnostic.
+
+**Exceptions to the default:** write-heavy large collections → sharded locks or a
+persistent structure; cross-store atomic transactions → a transactional backend via
+the Adapter (see the unit-of-work rule below); single-writer (Actor) state →
+wait-free reads, no write lock; non-shared / write-once state → plain ownership.
+
+### Transactions & the unit of work
+
+A transaction is a **value, not a place**. The doctrine, in five rules:
+
+1. **The backend Adapter issues the transaction context.** The Adapter over the
+   transactional technology (declared via `technologies` on its L4) exposes
+   begin/commit/rollback as ordinary contract methods; `begin` returns an opaque
+   transaction-context value. No other block ever creates one.
+2. **The method that owns the workflow owns the scope** — an Orchestrator's
+   method, or an Actor's own method when the Actor runs the workflow. Begin,
+   commit, and rollback are narrative steps of that method — begin, then a
+   `try` region whose calls carry the context, commit as the body's last step,
+   rollback in the `catch` handler. The transaction boundary is thereby VISIBLE
+   in the L5 narrative, reviewable like any other flow.
+3. **The context travels as an explicit method argument.** Repository facades
+   (and their inner Registry/Store) MAY accept an optional transaction-context
+   parameter on write-face methods and pass it through to the backend Adapter.
+   Never store the context in component state — a held transaction is hidden
+   state and a concurrency hazard in one.
+4. **In-memory state changes only after commit.** An Actor applies a decision to
+   its own state after the commit step, never before: a rollback leaves nothing in
+   memory to undo, and no reader sees a state the Store does not hold.
+5. **Guarantee tags stay honest.** A method declaring `transactional`/`atomic`
+   should either bind a transactional backend through its technology Adapter or
+   accept the context parameter. The validator checks claim↔declaration
+   consistency (a narrative step asserting a guarantee must call a method
+   declaring it) — whether the transaction actually holds is implementation
+   correctness, proven by tests, not by the gate.
+
+**Publish-and-persist atomically = the outbox pattern.** When a workflow must
+persist domain state AND emit an event without a gap, do not call the bus in the
+same breath as the write: append an outbox entry in the SAME transaction as the
+domain write, through a **sibling outbox Repository** (the outbox is an aggregate
+of its own, so it has its own Store), and let an Actor (or Observer on the
+backend's change feed) drain the outbox, publish through the bus Adapter, and mark
+entries delivered. A relay that restarts republishes, so consumers drop duplicates
+by event id. The Saga arrangement (§9) composes with this: the outbox is its
+delivery half, and the saga's progress Repository its memory.
+
+---
+
+## 11. Zero-copy / reference semantics (language-neutral)
+
+The **Store** owns the authoritative per-entry reference. **Indexes** hold the
+**same** references (shared, not value-copied), keyed differently — so reads are
+zero-copy and **value-content updates are transparent**: updating the entity behind
+its shared reference is seen by the Store and every Index at once, with no index
+change. Only **structural** changes — create, delete, or an update that changes an
+indexed field — propagate from the Store to its Indexes (an internal change-
+propagation mechanism, *not* the domain Observer block). The **Registry** writes
+only the Store; it never updates Indexes. (Where a language can't share references
+safely, the binding appendix gives the equivalent.)
+
+A derived Index — the exceptional Index over another Index of the same
+Repository (§3) — projects that Index's references rather than the Store's,
+and is dropped and rebuilt with it, so the same holds for it.
+
+A coherent Index defined this way is **never stale**. A deliberately **evicting /
+TTL cache** is a different thing — it is for *external or expensive-to-compute*
+data, not a projection of an in-process Store. An external cache service (e.g.
+Redis) is a caching **Adapter**; an in-process memo/TTL cache is a **Store with
+`durability: cache`** — evictable, loss-safe, hydration-exempt, and VISIBLE as
+held state. It is never a memoizing Orchestrator (an Orchestrator holds only its
+collaborators) and never a private field inside a logic component (hidden state).
+
+---
+
+## 12. Infrastructure (not building blocks)
+
+- **Message broker** — external infra (like a database). *Our client to it is an
+  Adapter* (so publish/subscribe edges are visible in the graph — see §4).
+- **Composition root / wiring (DI)** — the bootstrap that constructs the graph and
+  injects dependencies. Infrastructure, not a domain block.
+- **Cross-cutting** (authz, logging/metrics/tracing, config): coarse checks run at
+  the ingress — a Portal with the `gateway` variant calls authentication,
+  authorization, validation and rate-limit logic before it dispatches — plus infra
+  facilities; fine-grained authz lives in Registries/Orchestrators. Config loaded
+  at boot is infra injected at the composition root.
+
+---
+
+## 13. Right-sizing the model
+
+The standard upholds good structure at **any scale** — it must not force
+microservice / distributed / concurrent complexity on a system that doesn't need it.
+
+- **L1 is optional.** A small system (a restaurant reservation app, a single
+  webapp) can be **one subsystem ≈ the whole system** — the L1 layer collapses into
+  L0. Use multiple subsystems only when there are real bounded-context or deployment
+  boundaries.
+- **Concurrency is optional** (§10) — only when shared mutable state is actually
+  accessed concurrently. A simple single-threaded app uses plain ownership.
+- **Events are optional** (§4) — a pattern for decoupled reactions, not a mandate.
+  Without a message bus, in-process events are fire-and-forget calls; the simplest
+  systems call the next step directly.
+- **Use the smallest set of blocks/patterns the system needs.** A tiny app might be
+  one Portal + a couple of Orchestrators + one Repository. Don't add blocks,
+  patterns, or layers a system doesn't require.
+
+The *discipline* is the same everywhere (right block for the job, correct
+dependencies, behavior on the right object); only the *amount* of structure scales.
+
+---
+
+## 14. Diagram generation
+
+The spec tree is a typed graph, so it renders to diagrams with no extra
+modeling. These surfaces are SHIPPED (interactive canvas, Mermaid, draw.io,
+Excalidraw exports, and the flow modal):
+- component + `dependsOn` → C4 / component diagrams
+- `owns` tree → containment / module diagrams
+- typed entities (§6) → data-model / ER diagrams
+- L3 interfaces → contract views; L1 public interfaces → API maps
+- **L5 narratives → sequence diagrams and flowcharts** (each `call` is a
+  message; `parallel` fans out, `detach` renders fire-and-forget)
+
+Designs should stay faithful to the typed model so the diagrams stay free.
+
+---
+
+## 15. Worked example: a live auction
+
+Sellers list lots. Bidders watch a lot live and place bids; a bid in the last
+seconds extends the lot's end. An open lot closes at its deadline, a payment feed
+marks sold lots paid, and each seller has a dashboard of their lots, bids and
+bidders.
+
+| Component | Block | Responsible for |
+|---|---|---|
+| `bidding_portal` | Portal (HTTP and WebSocket) | bids, watch subscriptions, admin commands |
+| `bidding` | Orchestrator (workflow) | `place(lotId, bid)`: load the bidder, hand the bid to the live lot |
+| `lot_lifecycle` | Orchestrator (workflow) | `open`, `close`, `cancel`, `markPaid` — shared by the deadline sweep, admins and the payment feed |
+| `bid_rules` | Orchestrator, `pure` | `decide(lot, bid, bidder, policy, now)`, including the last-second extension |
+| `seller_dashboard` | Orchestrator, `read` | a seller's lots with their bids and bidders, joined across two Repositories |
+| `live_lot` | Actor | one open lot's working state, its watchers, its heartbeat |
+| `live_lots` | Supervisor | find-or-start a live lot by id, stop it, unload idle ones |
+| `deadline_sweep` | Actor | the process-wide tick that closes lots whose deadline has passed |
+| `outbox_relay` | Actor | publishing outbox entries through the event bus and marking them delivered |
+| `auction_runtime` | Supervisor (root) | starting and stopping `live_lots`, `deadline_sweep` and `outbox_relay` |
+| `lot_repository` | Repository | the lot aggregate: a Store over the tables `lots` and `bids`, a Registry, an Index by id and by `endsAt`, and the Query `bid_history` |
+| `bidder_repository` | Repository | the bidder aggregate |
+| `outbox_repository` | Repository | the outbox aggregate |
+| `auction_database` | Adapter | the database connection; issues transactions |
+| `event_bus` | Adapter | publishing and subscribing to events |
+| `watcher_sockets` | Adapter | pushing updates to watching clients over WebSocket |
+| `payment_feed` | Observer | forwarding payment events to `lot_lifecycle.markPaid` |
+
+```
+bidding_portal    → bidding, lot_lifecycle, seller_dashboard, live_lots, live_lot
+bidding           → bidder_repository, live_lots, live_lot
+lot_lifecycle     → live_lots, live_lot
+seller_dashboard  → lot_repository, bidder_repository
+payment_feed      → lot_lifecycle, event_bus
+auction_runtime   → live_lots, deadline_sweep, outbox_relay
+live_lots         → live_lot
+live_lot          → bid_rules, lot_repository, outbox_repository, auction_database, watcher_sockets
+deadline_sweep    → lot_repository, lot_lifecycle
+outbox_relay      → outbox_repository, event_bus
+```
+
+**Placing a bid.** `bidding_portal` hands the bid to `bidding.place`, which loads
+the bidder through `bidder_repository`, asks `live_lots` for the live lot by id
+(starting it, with its state loaded through `lot_repository`, when it is not
+running) and calls `live_lot.placeBid(bid, bidder)`. That method is the Decider
+shape (§3):
+1. **decide** — `bid_rules.decide(lot, bid, bidder, policy, now)` accepts or
+   rejects the bid, and moves `endsAt` when the bid lands in the last seconds; a
+   rejection returns at once;
+2. **persist** — begin a transaction through `auction_database`; in a `try`
+   region, write the bid and the lot's `endsAt` through `lot_repository` and
+   append a `bid_placed` entry through `outbox_repository`, both carrying the
+   context; commit; on failure, roll back and rethrow;
+3. **apply** — only after the commit, update its own leading bid and `endsAt`;
+4. **act** — push the new standing to the lot's watchers through `watcher_sockets`.
+
+`outbox_relay` later publishes `bid_placed` through `event_bus` and marks the
+entry delivered.
+
+**Closing.** The deadline is the field `lot.endsAt`, which the lot Index keeps in
+order. On each tick `deadline_sweep` reads the lots that are due through
+`lot_repository` and calls `lot_lifecycle.close` for each; `close` reaches the live
+lot through `live_lots`, the same single writer a bid goes through. `live_lot`
+checks `endsAt` against its own state before it closes, because a last-second bid
+it handled first may have moved the deadline.
+
+**After a restart.** Nothing is preloaded. A lot's Actor starts again on its next
+bid, watch or close; lots that fell due while the process was down close within one
+tick; the relay republishes entries it had not marked delivered, and consumers drop
+duplicates by event id.
+
+**The names.** Each component is named for what it is responsible for: a process
+(`bidding`, `lot_lifecycle`), a live thing (`live_lot`), a set of them
+(`live_lots`), an aggregate's data (`lot_repository`), an external system
+(`auction_database`, `event_bus`), an audience or a feed (`seller_dashboard`,
+`payment_feed`). Logic without collaborators is named for the rules it holds
+(`bid_rules`). A method is a verb phrase that does not repeat its component's
+words — `bidding.place`, not `bidding.placeBid` — and arithmetic over one value's
+own fields is a method on its type (`lot.isOpen(now)`). A role word appears only to
+tell apart components that share a noun, and it matches the block:
+`lot_repository` is a Repository.
+
+**The rules it demonstrates.**
+1. **A live Actor is addressed through its Supervisor.** `bidding`,
+   `lot_lifecycle` and `bidding_portal` reach a lot by id through `live_lots` and
+   never hold it; each lists `live_lot` together with `live_lots` in `dependsOn`,
+   since a call must target a declared dependency.
+2. **One writer per live aggregate.** `live_lot` is the only writer of its lot in
+   the process: a close from the sweep or an admin goes through `live_lots` just as
+   a bid does. A cluster needs a lease or sharded routing to keep it so.
+3. **The method that owns the workflow owns the transaction.** `live_lot.placeBid`
+   begins, commits and rolls back, and changes its in-memory state only after the
+   commit.
+4. **Supervisors form a tree.** `auction_runtime` supervises `live_lots` and the
+   two process-wide Actors.
+5. **A Portal messages a Supervisor by id**, as an Observer may: `bidding_portal`
+   adds a watcher to a live lot through `live_lots`.
+6. **Deadlines are aggregate fields read by an Index.** A schedule aggregate is
+   only for timers that span aggregates.
+7. **Reads across aggregates go through read logic.** `seller_dashboard` joins
+   `lot_repository` (its `bid_history` Query computes a lot's bids per call) with
+   `bidder_repository`; the outbox is a sibling Repository, written in the bid's
+   transaction.
+
+---
+
+## 16. Summary of binding rules
+
+1. Two axes: containment (folders) and refinement (the three files per component).
+2. Ten building blocks only; dependency rules fixed; the **Adapter** is the single
+   external-I/O boundary.
+3. A pattern owns **only blocks**, **one hop**, **never a pattern**; compose
+   patterns at L1. The one block that owns is a Supervisor, owning its
+   supervision state (its own Stores and Registries), private to it.
+4. `owns` ≠ `dependsOn`; cross-group access is via facades only.
+5. A Repository facade forwards 1:1 with no logic (single-`call`-to-owned-member
+   narratives; enforced as `FACADE_FORWARDING`).
+6. State-owners (Store/Index for domain state; Supervisor/Actor for runtime state) ≠
+   components holding only collaborators (Orchestrator/Registry/Query/Portal/Observer).
+7. All logic is a flowchart: an Orchestrator's methods, or an Actor's or
+   Supervisor's own methods over its runtime state. An Orchestrator's
+   `dependencyClass` bounds its dependencies — `pure` on pure logic only, `read` also
+   on read logic, Repositories, Indexes and Adapters, unset a workflow — and pure
+   logic may be used by every block. Local control flow → anywhere.
+8. Behavior lives where it can be performed autonomously; coordinated/external
+   behavior lives on the acting block, taking the entity as an argument.
+   Composition + interfaces over inheritance.
+9. Entities are first-class types, defined once by their owner; referenced, never
+   redefined, elsewhere. Interfaces are method contracts; fields are implementation.
+10. The Actor owns one live thing and its methods are full flowcharts; the
+    Supervisor owns the set of Actors, may supervise Supervisors, owns its
+    supervision state, and reaches shared data only through read and lifecycle
+    methods — writes go through a workflow. A live Actor is reached through its
+    supervision — looked up by id in a Registry its Supervisor maintains — and is
+    the only writer of its aggregate; the method that owns a workflow owns its
+    transaction, and in-memory state changes only after commit. A lifecycle
+    method changes what exists, never domain fields, and calls only read and
+    lifecycle methods.
+11. Structure is L2/L3; the concurrency variant is L4 (neutral strategy here,
+    primitives in the language-bindings appendix). The bus client, composition root,
+    and cross-cutting concerns are infrastructure, not blocks.
+12. A Store covers one aggregate. Registry (write), Index and Query (read) are
+    decoupled — all work on the Store; Indexes are reference-sharing projections,
+    Queries compute per call inside their Repository, and the Registry updates
+    neither. Reads across aggregates go through read logic or a read-model Repository.
+13. Right-size: L1, concurrency, and events are all optional; use the smallest set
+    of blocks a system needs, while upholding the same discipline at any scale.
+14. A gateway is a Portal with the `gateway` variant. Variants load built-in →
+    global → project and carry guidance, not rules. `Specialist` and `Gateway` are
+    retired (`STEREOTYPE_RETIRED`) until migrated.