revisão e complementação da documentação do projeto

README.en-us.md

@@ -2672,7 +2672,7 @@ The architectural vision is consistent, but a concrete implementation needs to d
    - workload identity?
    - mTLS?
    - cloud IAM?
-   - → **Decision adopted in the RI**: OAuth 2.1 `client_credentials` via Keycloak, with mTLS/workload identity documented as a production extension not implemented (`docs/adr/ADR-019-identity-extension.md`).
+   - → **Decision adopted in the RI**: OAuth 2.1 `client_credentials` via Keycloak, with mTLS/workload identity documented as a production extension not implemented (`docs/adr/ADR-019-extensao-de-identidade.md`).
 
 2. **What will be the granularity of client profiles?**
    - domain?

README.md

@@ -2672,7 +2672,7 @@ A visão arquitetural está consistente, mas uma implementação concreta precis
    - workload identity?
    - mTLS?
    - cloud IAM?
-   - → **Decisão adotada na RI**: OAuth 2.1 `client_credentials` via Keycloak, com mTLS/workload identity documentados como extensão de produção não implementada (`docs/adr/ADR-019-identity-extension.md`).
+   - → **Decisão adotada na RI**: OAuth 2.1 `client_credentials` via Keycloak, com mTLS/workload identity documentados como extensão de produção não implementada (`docs/adr/ADR-019-extensao-de-identidade.md`).
 
 2. **Qual será a granularidade dos client profiles?**
    - domínio?

RI/docs/AS-BUILT.en-us.md

@@ -77,7 +77,7 @@ local JWKS HTTP server as a stand-in for the rest of the e2e suite (valid for va
 not for real protocol integration with the IdP).
 
 **Production extension documented, not implemented:** mTLS/workload identity —
-`docs/adr/ADR-019-identity-extension.md`.
+`docs/adr/ADR-019-extensao-de-identidade.md`.
 
 ### PDP / Policy Engine (EP-03)
 

RI/docs/AS-BUILT.md

@@ -75,7 +75,7 @@ HTTP local mais leve como stand-in para o resto da suíte e2e (válido para a l
 não para a integração de protocolo real com o IdP).
 
 **Extensão de produção documentada, não implementada:** mTLS/workload identity —
-`docs/adr/ADR-019-identity-extension.md`.
+`docs/adr/ADR-019-extensao-de-identidade.md`.
 
 ### PDP / Policy Engine (EP-03)
 

RI/docs/Assumptions.en-us.md

@@ -24,7 +24,7 @@ point being tested.
 and adherent to open standards** (OAuth 2.1/OIDC, Rego), not out of arbitrary preference. This
 resolves two gaps that the reference architecture left open (`README.md` §48, questions
 1 and 3) with a definitive technical decision, not an assumption — they were never revisited
-throughout the implementation. See `docs/adr/ADR-019-identity-extension.md`,
+throughout the implementation. See `docs/adr/ADR-019-extensao-de-identidade.md`,
 `docs/adr/ADR-025-engine-do-pdp.md`.
 
 **We rejected** a custom PDP engine in Python (more control, but reinvents policy evaluation

RI/docs/Assumptions.md

@@ -24,7 +24,7 @@ próprio ponto que estava sendo testado.
 aderentes a padrões abertos** (OAuth 2.1/OIDC, Rego), não por preferência arbitrária. Isso
 resolve duas lacunas que a arquitetura de referência deixava abertas (`README.md` §48, perguntas
 1 e 3) com uma decisão técnica definitiva, não uma suposição — nunca foram revisitadas ao longo
-da implementação. Ver `docs/adr/ADR-019-identity-extension.md`,
+da implementação. Ver `docs/adr/ADR-019-extensao-de-identidade.md`,
 `docs/adr/ADR-025-engine-do-pdp.md`.
 
 **Rejeitamos** um engine de PDP próprio em Python (mais controle, mas reinventa avaliação de

RI/docs/Production-Recommendations.en-us.md

@@ -13,7 +13,7 @@ Target audience: whoever is going to **evaluate or adapt** this RI for a real sc
 A real, highly available IdP (Keycloak cluster, or a managed equivalent), not the single
 instance from `docker-compose.yml`. The extension to mTLS/workload identity — covering cases
 where `client_credentials` is not strong enough (e.g.: workloads in a service mesh) — is
-already documented as an extension point in `docs/adr/ADR-019-identity-extension.md`; production
+already documented as an extension point in `docs/adr/ADR-019-extensao-de-identidade.md`; production
 would implement that extension, not reinvent the authentication middleware from scratch.
 
 ## PDP

RI/docs/Production-Recommendations.md

@@ -13,7 +13,7 @@ Público-alvo: quem vai **avaliar ou adaptar** esta RI para um cenário real.
 Um IdP real em alta disponibilidade (cluster Keycloak, ou equivalente gerenciado), não a
 instância única do `docker-compose.yml`. A extensão para mTLS/workload identity — cobrindo casos
 em que `client_credentials` não é forte o suficiente (ex.: workloads em malha de serviço) — já
-está documentada como ponto de extensão em `docs/adr/ADR-019-identity-extension.md`; produção
+está documentada como ponto de extensão em `docs/adr/ADR-019-extensao-de-identidade.md`; produção
 implementaria essa extensão, não reinventaria o middleware de autenticação do zero.
 
 ## PDP

docs/adr/ADR-001-agent-privilege-ceiling.en-us.md

@@ -0,0 +1,20 @@
+# ADR-001: MaximumEntitlement Boundary at the MCP Client
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Who defines the agent's privilege ceiling.
+
+## Decision
+
+MCP Client (authenticated identity) is the `MaximumEntitlement` boundary, never the agent/prompt.
+
+## Alternatives Considered
+
+Authorization by self-declared role of the agent.
+
+## Consequences
+
+All EP-01/EP-04 design revolves around the client's identity, not the agent's.

docs/adr/ADR-002-enforcement-at-the-gateway.en-us.md

@@ -0,0 +1,20 @@
+# ADR-002: MCP Gateway as the sole PEP
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Where to apply enforcement.
+
+## Decision
+
+MCP Gateway is the sole PEP; it never delegates the ALLOW/DENY decision to the LLM.
+
+## Alternatives Considered
+
+Distributed enforcement in each MCP Server.
+
+## Consequences
+
+Centralizes EP-05; domain servers (EP-06) do not reimplement authz.

docs/adr/ADR-003-reauthorization-of-tools-call.en-us.md

@@ -0,0 +1,20 @@
+# ADR-003: Independent reauthorization on tools/call
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Is filtered `tools/list` sufficient?
+
+## Decision
+
+No — `tools/call` must be reauthorized independently of discovery.
+
+## Alternatives Considered
+
+Relying only on filtered discovery.
+
+## Consequences
+
+EP-05 implements two enforcement points, not one.

docs/adr/ADR-004-context-never-elevates-privilege.en-us.md

@@ -0,0 +1,20 @@
+# ADR-004: Agentic context never elevates privilege
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Can agentic context elevate privilege?
+
+## Decision
+
+Never — context can only subtract (`Context may subtract, must never add`).
+
+## Alternatives Considered
+
+Allowing elevation based on the model's "confidence."
+
+## Consequences
+
+EP-15 tests (Test 3/4) validate the invariant.

docs/adr/ADR-005-private-cache-for-tools-list.en-us.md

@@ -0,0 +1,20 @@
+# ADR-005: Private cache for tools/list
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+`tools/list` cache shared across clients.
+
+## Decision
+
+`cacheScope=private`, key = identity+context+policy_version.
+
+## Alternatives Considered
+
+Shared cache for performance.
+
+## Consequences
+
+EP-05-T03; dedicated isolation test (EP-15).

docs/adr/ADR-006-no-token-passthrough.en-us.md

@@ -0,0 +1,20 @@
+# ADR-006: Prohibition of token passthrough
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Can the agent's token be forwarded to the backend?
+
+## Decision
+
+No — blind token passthrough is prohibited; inbound ≠ outbound identity.
+
+## Alternatives Considered
+
+Direct passthrough of the client's token.
+
+## Consequences
+
+EP-07 implements exchange/service account per backend.

docs/adr/ADR-007-backend-not-directly-exposed.en-us.md

@@ -0,0 +1,20 @@
+# ADR-007: Backend never exposed directly to the client
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Is the backend directly accessible by the client?
+
+## Decision
+
+No — network/identity restrict access to the Gateway/Fabric only.
+
+## Alternatives Considered
+
+Exposing the Fabric publicly with "best effort" authz.
+
+## Consequences
+
+EP-05-T07, EP-15 (bypass test).

docs/adr/ADR-008-segmentation-by-domain-and-risk.en-us.md

@@ -0,0 +1,20 @@
+# ADR-008: Client segmentation by domain and risk tier
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+How to segment clients.
+
+## Decision
+
+By domain + risk tier (R0–R4), not a single global client.
+
+## Alternatives Considered
+
+A single "enterprise admin" client.
+
+## Consequences
+
+Defines P6 (granularity) and EP-04.

docs/adr/ADR-009-risk-tiers-for-capabilities.en-us.md

@@ -0,0 +1,20 @@
+# ADR-009: Risk tiers R0-R4 for capabilities
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+How to classify capabilities.
+
+## Decision
+
+Risk tiers R0–R4 in the Registry, controlling required approval/observability.
+
+## Alternatives Considered
+
+No formal classification.
+
+## Consequences
+
+EP-02-T01 includes mandatory `risk_tier` in the manifest schema.

docs/adr/ADR-010-registry-governance.en-us.md

@@ -0,0 +1,20 @@
+# ADR-010: Governance of the Global Capability Registry
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Governance of the Global Capability Registry (title and number inherited from §49 of the README — "Global Capability Registry Governance").
+
+## Decision
+
+The Registry is a control-plane asset (§6.6) with its own authority and lifecycle, separate from the catalog view delivered to each client. In the RI: dedicated storage and read API (EP-02-T03), consumed by Gateway/Fabric/Offering Filter but never edited by them; formal lifecycle nominate→review→register→publish→observe→(change|deprecate→retire) (§23) implemented as a state machine (EP-02-T04); every change goes through the publishing pipeline (EP-02-T02) — no direct manual registration in production (RF-11). The Registry's organizational owner (who approves schema/policy changes to the control plane itself) remains a corporate governance decision outside the technical scope of the RI (§48, question 4, still open).
+
+## Alternatives Considered
+
+Leaving the Registry as a table editable ad hoc by any service; merging the control view (Registry) with the execution view (catalog filtered per client), which would violate the control plane/data plane separation of §24.
+
+## Consequences
+
+The entire EP-02; reinforces that EP-05 (Gateway) and EP-10 (Offering Filter) only read the Registry, never write to it.

docs/adr/ADR-011-publishing-pipeline.en-us.md

@@ -0,0 +1,20 @@
+# ADR-011: Mandatory pipeline for capability publication
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Capability publication cycle.
+
+## Decision
+
+Mandatory pipeline with schema/risk/policy/health gates before entering the Registry — no manual registration in production.
+
+## Alternatives Considered
+
+Direct registration via administrative API without a pipeline.
+
+## Consequences
+
+EP-02-T02 implements the gate; no capability is "born" already published.

docs/adr/ADR-012-entitlement-manager-as-authority.en-us.md

@@ -0,0 +1,20 @@
+# ADR-012: Entitlement Manager as the sole authority for MaximumEntitlement
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, where applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Authority over the maximum menu.
+
+## Decision
+
+Entitlement Manager is the sole authority for `MaximumEntitlement`; it does not rank nor trust self-declared attributes from the LLM.
+
+## Alternatives Considered
+
+Merging Entitlement Manager and Offering Filter into a single component.
+
+## Consequences
+
+Keeps EP-04 and EP-10 as distinct services/modules with their own contracts.

docs/adr/ADR-013-profile-intelligence-contract.en-us.md

@@ -0,0 +1,20 @@
+# ADR-013: Pluggable Profile Intelligence contract
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Profile Intelligence contract.
+
+## Decision
+
+Generic, replaceable interface (versioned `ProfileView`, with `reasonCodes`), pluggable implementation.
+
+## Alternatives Considered
+
+Coupling to a specific vendor/taxonomy for segmentation.
+
+## Consequences
+
+EP-09 defines only the contract + reference stub (see P7).

docs/adr/ADR-014-profile-never-expands-entitlement.en-us.md

@@ -0,0 +1,20 @@
+# ADR-014: Profile never expands entitlement
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Can profile expand entitlement?
+
+## Decision
+
+Never — `FilteredOfferings ⊆ MaximumEntitlement`, profile only orders/reduces.
+
+## Alternatives Considered
+
+High-engagement profile unlocking new tools.
+
+## Consequences
+
+EP-10-T02 brings a property test validating the invariant for random inputs.

docs/adr/ADR-015-nba-is-not-authorization.en-us.md

@@ -0,0 +1,20 @@
+# ADR-015: NBA as an orchestration decision, not an authorization decision
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Is NBA an authorization decision?
+
+## Decision
+
+No — it is orchestration intent; Gateway/PDP reauthorize every execution.
+
+## Alternatives Considered
+
+Orchestrator dispatching directly to the backend without going through the Gateway.
+
+## Consequences
+
+EP-11-T04 implements the DENY/recalculation branch of §10.1.

docs/adr/ADR-016-dispatch-handoff-contract.en-us.md

@@ -0,0 +1,20 @@
+# ADR-016: Agent Runtime dispatch/handoff contract
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Agent Runtime dispatch/handoff contract.
+
+## Decision
+
+Versioned dispatch object (`decisionId`, `expiresAt`, `reasonCodes`) that does not replace the MCP Client's token.
+
+## Alternatives Considered
+
+Runtime injecting its own credentials into the agent.
+
+## Consequences
+
+EP-12-T01 consumes exclusively the dispatch object + the already-existing MCP Client.

docs/adr/ADR-017-end-to-end-correlation.en-us.md

@@ -0,0 +1,20 @@
+# ADR-017: End-to-end correlation via versioned IDs
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+End-to-end correlation.
+
+## Decision
+
+`profileVersion`+`entitlementVersion`+`decisionId`+`policyDecisionId`+`mcpRequestId` propagated via OTel/trace context across all components.
+
+## Alternatives Considered
+
+Correlation only via free-text logs with no standard.
+
+## Consequences
+
+EP-08-T03 formalizes the fields; basis for EP-15 (auditability).

docs/adr/ADR-018-implementation-stack.en-us.md

@@ -0,0 +1,58 @@
+# ADR-018: RI implementation stack
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+RI implementation stack (closes G2).
+
+## Decision
+
+Python 3.12+, `uv`, Pydantic v2, declarative YAML with JSON Schema, official `mcp` SDK.
+
+## Alternatives Considered
+
+Node/TypeScript (the MCP SDK is also official there); Go.
+
+## Consequences
+
+Defines the entire repository structure (8.5).
+
+## Confirmed Versions (EP-16-T04)
+
+Dependency-free spike, runnable at any time (EP-16-T04): confirm,
+immediately before pinning in `pyproject.toml`/
+`docker-compose.yml`, the exact stable version numbers of each piece of the
+stack — no version cited in code/config without real confirmation
+(`pip index versions`/direct download/actual install), never from memory.
+Consolidated record, with the date each group was verified (see
+`RI/README.md`, "Versões fixadas", for the same list maintained alongside
+`pyproject.toml`):
+
+Confirmed on 2026-09-11:
+
+- `mcp` 2.2.0 (Python SDK v2 — the API changed significantly from v1; see
+  comments in `src/emcp_bus/gateway/server.py`)
+- `pydantic` 2.13.x, `opentelemetry-sdk` 1.44.x, `uvicorn` 0.52.x,
+  `pyjwt[crypto]` 2.14.x, `ruff` 0.16.x, `mypy` 2.3.x, `pytest` 9.1.x
+- `OPA` 1.20.2 (static binary, Rego v1 syntax as default)
+- `pytest-forked` 1.7.x
+
+Confirmed on 2026-09-14 (M3/M4, via actual install/`docker pull`):
+
+- `langgraph` 0.6.x + `langgraph-checkpoint-sqlite` 2.x (EP-11)
+- `hypothesis` 6.1xx.x (dev, EP-10-T02)
+- `openai` 1.109.x (EP-13-T01)
+- `python-dotenv` 1.2.x (dev)
+- `types-jsonschema` 4.26.x (dev, stubs for `mypy --strict`; `jsonschema`
+  itself was already a transitive dependency of `mcp[cli]`, EP-15-T01)
+- Langfuse self-hosted 4.35.0 (`docker.langfuse.com/langfuse/langfuse:4`,
+  official image, EP-13-T02)
+- Keycloak 26.7.3 (`quay.io/keycloak/keycloak:26.7.3`, Quarkus; Standard
+  Token Exchange GA since 26.2 — EP-01-T01/EP-07)
+- `opentelemetry-collector-contrib` 0.114.0 and `jaeger` (all-in-one) 1.62.0
+  (`observability` profile of `docker-compose.yml`, EP-08-T04)
+
+None of the versions above have changed since the original confirmation — this
+record is the single source of truth that `RI/README.md` links back to.

docs/adr/ADR-019-extensao-de-identidade.md

@@ -0,0 +1,106 @@
+# ADR-019: Identidade do MCP Client -- ponto de extensão de produção (mTLS / workload identity)
+
+**Status:** Documentado (spike), não implementado na RI (EP-01-T04, `could`).
+**Fonte:** este ADR é o registro canônico desta decisão. Para contexto de implementação e validação, ver `RI/docs/AS-BUILT.md` e, quando aplicável, o contrato relevante em `RI/docs/interfaces/identity-oidc.md`. Também relevante: README.md §19, RF-02, Axiom 2.
+
+## Contexto
+
+A fronteira de identidade da RI (`src/emcp_bus/identity/authn.py`, EP-01-T03) é
+OAuth 2.1 `client_credentials` contra o Keycloak (EP-01-T01): todo MCP Client
+se autentica com um client_id + secret, obtém um JWT de vida curta, e o
+Gateway valida a assinatura/emissor/audience/expiração desse JWT contra o
+JWKS do IdP antes de confiar na claim `azp` como a identidade do caller.
+
+Esta é uma decisão deliberada e confirmada para a RI (não um placeholder): OAuth
+2.1 satisfaz toda propriedade que o README.md §19 de fato exige de um mecanismo
+de identidade -- vida curta, revogável, vinculado a uma audience, verificável
+sem um segredo de longo prazo compartilhado no caminho da requisição.
+Implantações corporativas reais comumente vão além, usando mTLS (certificados
+de cliente TLS mútuo) ou a federação de workload identity de um provedor de
+nuvem (ex.: SPIFFE/SPIRE, AWS IAM roles for service accounts, GCP Workload
+Identity, Azure Managed Identity) em vez de -- ou além de -- um JWT bearer.
+Este ADR documenta como essa extensão se encaixa na costura já existente da
+RI, sem exigi-la, para que um fork de produção não fique bloqueado por um
+redesenho.
+
+## Decisão
+
+**Não** implementar mTLS/workload identity na RI. Documentar o ponto de
+extensão em vez disso, para que o contrato do `EP-01-T03` seja
+comprovadamente estável sob ele.
+
+## O ponto de extensão
+
+`TokenValidator.validate(token: str) -> AuthenticatedClient` é a *única*
+função na RI que produz identidade de client para fins de autorização
+(`emcp_bus.gateway.server.on_list_tools`/`on_call_tool` chamam ambos
+`get_access_token()`, que é populado pelo protocolo `TokenVerifier` do SDK
+envolvendo esse mesmo validator -- ver
+`emcp_bus.identity.token_verifier.MCPTokenVerifier`). Tudo a jusante dela --
+`EntitlementManager.resolve`, o PDP, a camada de audit/correlação (EP-08) --
+consome apenas o `AuthenticatedClient.client_id` resultante. Nenhum desse
+código se importa com *como* a identidade foi estabelecida.
+
+Uma implantação de produção adicionando mTLS ou workload identity faria:
+
+1. Implementar uma segunda fonte de identidade honrando o mesmo contrato de saída:
+
+   ```python
+   class MTLSClientValidator:
+       def validate(self, request: ...) -> AuthenticatedClient: ...
+   ```
+
+   alimentada pelo certificado de cliente já verificado pelo proxy/load
+   balancer que termina a conexão (um header estilo `X-Client-Cert-CN` de um
+   terminador mTLS confiável, ou o `scope["extensions"]["tls"]` ASGI que um
+   servidor compliant expõe) em vez de um JWT. `AuthenticatedClient.client_id`
+   seria o Subject/SAN do certificado, mapeado 1:1 a um
+   `ClientRegistration.client_id` exatamente da mesma forma que o `azp` de um
+   JWT é hoje -- **nenhum outro modelo da RI muda**: `ClientRegistration`,
+   `ClientProfile`, `EntitlementManager`, o contrato do PDP e o schema de
+   auditoria já são todos agnósticos de mecanismo de identidade.
+
+2. Fornecer um equivalente alternativo de `build_auth()` em
+   `emcp_bus/gateway/server.py` que conecta o novo validator no lugar de
+   `MCPTokenVerifier`, controlado por configuração de deployment -- o mesmo
+   formato do `build_auth()` de hoje retornando `(None, None)` para o modo
+   "sem OIDC configurado". Uma implantação híbrida (mTLS para uma classe de
+   backend service, OAuth para agentes interativos) é possível fazendo a
+   seleção de `TokenVerifier`/equivalente por listener em vez de por
+   processo, o que os parâmetros
+   `streamable_http_app(auth=..., token_verifier=...)` do SDK já suportam
+   por instância de app.
+
+3. Workload identity de nuvem (SPIFFE/SPIRE, IRSA, Workload Identity, Managed
+   Identity) é uma variante da mesma forma: o contrato "verificar uma
+   credencial que o caller não pode forjar, e extrair dela um identificador
+   estável" é idêntico, só o formato da credencial e a biblioteca de
+   verificação diferem. Nada disso exige tocar no `EntitlementManager`, no
+   PDP, no roteador do Fabric (EP-06-T01), ou na camada de audit/correlação
+   (EP-08-T02/T03).
+
+## O que não pode mudar
+
+- **Formato de `AuthenticatedClient`** (`client_id`, `issuer`, `expires_at`):
+  todo consumidor a jusante tipa contra isso, não contra um conjunto de
+  claims específico de JWT.
+- **Contrato fail-closed**: o caminho de falha de qualquer mecanismo de
+  identidade deve levantar uma exceção (no formato `AuthenticationError`) em
+  vez de recair para uma identidade não autenticada ou autodeclarada pelo
+  agente (README.md §5, Axiom 2) -- essa é a propriedade de segurança real
+  que o critério de aceite do EP-01-T03 testa, e ela é independente de
+  mecanismo.
+- **Nenhum caminho privilegiado novo para `clientInfo`**: qualquer que seja o
+  mecanismo de identidade, apenas um valor verificado de forma independente
+  do corpo da requisição MCP pode popular `client_id`.
+
+## Consequências
+
+- Confirmado para a RI: reduz o esforço de implementação do EP-01 sem
+  enfraquecer nenhuma das propriedades de identidade desejáveis do README.md
+  §19 (todas satisfeitas por OAuth 2.1 `client_credentials` contra o
+  Keycloak).
+- Um fork de produção adicionando mTLS/workload identity muda exatamente um
+  módulo (`identity/`) e uma função de wiring (equivalente a `build_auth`);
+  todo outro componente da RI (Entitlement Manager, PDP, Fabric, Registry,
+  audit/correlação, approval) não exige nenhuma mudança.

docs/adr/ADR-020-downstream-identity.en-us.md

@@ -0,0 +1,20 @@
+# ADR-020: Downstream identity: token exchange or isolated service account
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Downstream identity in the RI (closes G5/P5).
+
+## Decision
+
+Token exchange (RFC 8693) when the backend supports OIDC; otherwise a service account isolated per adapter.
+
+## Alternatives Considered
+
+Always reusing the client's token (forbidden by ADR-006).
+
+## Consequences
+
+EP-07; documented per backend in the manifest.

docs/adr/ADR-021-langgraph-langfuse-scope.en-us.md

@@ -0,0 +1,20 @@
+# ADR-021: Scope of LangGraph and Langfuse in the RI
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Scope of LangGraph/Langfuse (closes G8/P8).
+
+## Decision
+
+LangGraph only in the Service Orchestrator (EP-11) and the Agent Runtime handoff (EP-12); Langfuse only where there is an actual LLM call (EP-13). Checkpointer: SQLite in dev.
+
+## Alternatives Considered
+
+Also using LangGraph in the Gateway/Fabric for "stack consistency."
+
+## Consequences
+
+Avoids the over-engineering vetoed by section 6 of the superprompt; Gateway/Fabric remain plain Python.

docs/adr/ADR-022-reference-hoshikawa-agent-platform-oci.en-us.md

@@ -0,0 +1,20 @@
+# ADR-022: Operational reference: Agent Platform OCI (Hoshikawa)
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Concrete Python/LangGraph/Langfuse implementation reference (closes G2 — **decision confirmed by the author on 2026-09-11**).
+
+## Decision
+
+Adopt as an operational reference the "Agent Platform OCI" by Christiano Hoshikawa (`github.com/hoshikawa2/agent_platform_oci`): a StateGraph with a router/supervisor per backend + an upper routing layer across backends (our EP-11/EP-12); pluggable checkpointing (memory/sqlite/mongodb/production — we adopt SQLite in dev, see ADR-021); **IC/NOC/GRL** observability taxonomy via Langfuse (business/operational/guardrail event), with hybrid instrumentation (automatic span per node + manual fail-open emission); `identity.yaml`/`BusinessContext` to normalize business identity across channels; FastMCP as the standard MCP server. Full analysis recorded in `docs/research/hoshikawa-agent-platform-oci.md`.
+
+## Alternatives Considered
+
+Continuing with the generic pattern from section 6 of the superprompt without a concrete third-party reference (slower, more ad hoc decisions); adopting a market agent framework (CrewAI, AutoGen) instead of plain LangGraph.
+
+## Consequences
+
+Changes the form (not the security) of EP-08, EP-11, EP-12, EP-13 — see ADR-023 for the explicit boundary of what is NOT adopted.

docs/adr/ADR-023-authorization-never-comes-from-payload.en-us.md

@@ -0,0 +1,20 @@
+# ADR-023: Tool authorization never comes from a field declared in the payload
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+Tool authorization NEVER comes from a field declared in the payload/conversation (a critical finding from the Hoshikawa analysis, reinforcing ADR-001/002/003).
+
+## Decision
+
+The Agent Platform OCI's MCP Gateway authorizes tools via a static allowlist (`tools.yaml`/`authorization.agents.<agent_id>.allowed_tools`) evaluated over `agent_id`, a field that arrives as **data from the conversation's input payload**, not as a claim from an independently verified authenticated identity — their own SPEC-004 admits that this does not replace real authentication/authorization. Our RI reuses the proxy *form* of this MCP Gateway (`ToolInvocation`/`ToolResult` contracts, idempotent per-tool cache, `/v1/tools` catalog) **exclusively at the Fabric layer (EP-06)/conversational UX**, but the ALLOW/DENY decision remains 100% in our PEP (EP-05) consulting the PDP (EP-03) with the authenticated identity of the MCP Client (EP-01) — never with a self-declared `agent_id`/`tenant_id` in the payload.
+
+## Alternatives Considered
+
+Also reusing their `authorization` block as an implementation shortcut.
+
+## Consequences
+
+EP-15-T03 gains an explicit adversarial "payload-declared agent_id spoofing" scenario (a variation of Test 4, §45) to prove that this specific vector is closed in our RI.

docs/adr/ADR-024-multiagent-positioning.en-us.md

@@ -0,0 +1,20 @@
+# ADR-024: E-MCP-BUS positioning relative to external multi-agent platforms
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+E-MCP-BUS positioning relative to external multi-agent orchestration platforms, such as Agent Platform OCI (**decision confirmed and expanded by the author on 2026-09-11** — resolves G11).
+
+## Decision
+
+E-MCP-BUS **does not compete** with orchestration platforms such as Agent Platform OCI — but it also **is not limited to coexisting under them**. Explicit design principle: **modularity + compatibility = choice**. Our Service Orchestrator (EP-11) is not a thin pass-through: it computes NBA **and NBO** (Next Best Action/Next Best Offer) with its own decision model (rules + guardrails +, optionally, a ranking/recommendation AI model — see EP-11-T03) over the universe of `FilteredOfferings` that the MCP Fabric exposes (services, offers, actions, campaigns — not just technical "tools"). This enables **two equally valid integration modes, chosen by the adopter according to the desired topology**: **(A) Governance-only** — an external orchestration platform (e.g., Agent Platform OCI, with its own Global Supervisor and StateGraph deciding NBA) connects to E-MCP-BUS solely for governed access to the Fabric; our Service Orchestrator does not participate, or participates only as a source of `FilteredOfferings` (as already described in the rest of this ADR and in EP-13-T05). **(B) Full replacement** — our own Service Orchestrator (EP-11) fully takes on the role of the decision "brain" (NBA/NBO), dispensing with the external platform's decision-making Global Supervisor/StateGraph; the Agent Runtime (EP-12) and the external platform's channel adapters (e.g., Hoshikawa's agent backends) become mere executors of the dispatch our Orchestrator produces. In either mode, the security core (PEP/PDP, EP-05/EP-03) never changes place — only who computes the NBA/NBO changes. In both modes, each domain backend/agent remains registered as a **distinct MCP Client** (EP-01), bound to its own client profile/entitlement per domain+risk tier (EP-04, P6), and any consuming platform's internal "MCP Gateway" (ADR-023) is always superseded by our Gateway/PEP as the single point of authorization. This is the "Enterprise elevation": centrally governed catalog/entitlement/audit, with the NBA/NBO decision able to come from inside (mode B) or outside (mode A) the bus, by the adopter's modular choice.
+
+## Alternatives Considered
+
+Forcing a single integration mode (e.g., only "governance-only"), which would underuse the Service Orchestrator and always require an external decision platform; treating Agent Platform OCI as a substitute for E-MCP-BUS (incorrect — it does not solve entitlement); reimplementing from scratch the orchestration engineering already validated in ADR-022 even when mode B is chosen (the StateGraph/checkpointing/handoff form remains reusable within our EP-11/EP-12).
+
+## Consequences
+
+Reshapes the goal of EP-11 (Service Orchestrator explicitly described as computing NBA **and NBO** over Fabric services/offers/actions, with a pluggable decision model) and of EP-13 (demonstrates the two modes — see EP-13-T05 for mode A and EP-13-T06, new, for mode B). Reinforces that EP-06/EP-05 treat any consuming platform's "MCP Gateway" as untrusted for authorization, in both modes.

docs/adr/ADR-025-pdp-engine.en-us.md

@@ -0,0 +1,20 @@
+# ADR-025: OPA/Rego as the PDP engine
+
+**Status:** Accepted.
+**Source:** this ADR is the canonical record of this decision. For implementation and validation context, see `RI/docs/AS-BUILT.md` and, when applicable, the relevant contract in `RI/docs/interfaces/`.
+
+## Context
+
+PDP engine (closes G3 — **decision confirmed by the author on 2026-09-11**; renumbered from ADR-010 to ADR-025 on 2026-09-11 to avoid colliding with the README's ADR-010, "Global Capability Registry Governance").
+
+## Decision
+
+OPA/Rego, self-hosted, OSS (Apache-2.0). Final decision, not subject to review. **Verified on 2026-09-11 via web search**: OPA is at v1.0+, with Rego v1 syntax as the default (a breaking change vs. v0) — EP-03 must write policies in Rego v1 syntax from the start.
+
+## Alternatives Considered
+
+Cedar (AWS, cited in §43.2); a custom engine in Python.
+
+## Consequences
+
+The entire EP-03 becomes coupled to OPA's REST API; policies versioned as code in Rego v1.

docs/research/hoshikawa-agent-platform-oci.en-us.md

@@ -0,0 +1,135 @@
+# Research: Agent Platform OCI (Christiano Hoshikawa) as a Python/LangGraph/Langfuse reference
+
+> Source: https://github.com/hoshikawa2/agent_platform_oci (README + SPEC-001, 002, 003, 004, 007, 012, 018). Analyzed on 2026-09-11 to underpin the decisions recorded in [`ADR-022`](../adr/ADR-022-reference-hoshikawa-agent-platform-oci.en-us.md)/[`ADR-023`](../adr/ADR-023-authorization-never-comes-from-payload.en-us.md) (see also [`../../RI/docs/Assumptions.md`](../../RI/docs/Assumptions.md)). Referenced by ADR-022/ADR-023.
+>
+> **Scope of this research**: extract the *operational* LangGraph/Langfuse/MCP pattern from that project for our RI's Python implementation layer. The security core (PEP/PDP/Client-Bound Entitlement) of our architecture remains 100% defined by the `README.md` at the root of this repository — never by the project researched here.
+
+---
+
+## 1. LangGraph pattern
+
+**StateGraph — nodes and state shape.** The per-backend graph is fixed and corporate:
+
+```text
+START → input_guardrails → routing_decision → [domain agent(s)] → output_supervisor
+      → output_guardrails → judge → supervisor_review → persist → END
+routing_decision → handoff → routing_decision   (conditional return edge)
+```
+
+`AgentState` is a `TypedDict(total=False)` with: `user_text, sanitized_input, response_text, tenant_id, agent_id, channel, session_id, conversation_key, message_id, route, intent, context, business_context, tool_arguments, mcp_tools, mcp_results, rag_context, rag_metadata, guardrails, judges, metadata, errors`, plus `active_transaction`/`last_transaction` in the transactional extension (a durable state contract that survives checkpoint/resume).
+
+Registering a new agent is mechanical: import the class → instantiate it in the workflow's `__init__` (receiving shared dependencies: telemetry, tool_router, rag_service, cache, observer, memory) → async wrapper with a telemetry span → `builder.add_node(...)` → entry in `add_conditional_edges("routing_decision", ...)` → wiring to `output_supervisor`.
+
+**Two routing layers, explicitly separated:**
+- **Backend-local supervisor/router** — decides an agent's internal flow (`routing.yaml`, `ROUTING_MODE=router|supervisor`).
+- **Global Supervisor / Agent Gateway** — a separate deployable service that decides **which whole backend** handles the conversation (`GLOBAL_ROUTING_MODE=router|supervisor|hybrid`, `backends.yaml`). HTTP routing between backends, not a second LangGraph.
+
+**Checkpointing**: `memory` (tests), `sqlite` (local dev), `mongodb` (distributed), `autonomous` (OCI production) providers. Contract: `{conversation_key, checkpoint_id, state, pending_writes, created_at}`.
+
+**Handoff**: intra-backend via the conditional edge `routing_decision → handoff → routing_decision`; inter-backend via `metadata.handover_backend` in the response, detected by the Agent Gateway, which keeps `global_session_id` separate from `backend_session_id`.
+
+---
+
+## 2. Langfuse pattern
+
+**IC / NOC / GRL taxonomy** (not "Informational Context" — it's a **business/journey**, **operational**, and **guardrail** event, respectively):
+
+| Family | Meaning | Examples |
+|---|---|---|
+| IC | agent business/journey event | `IC.GATEWAY_RECEIVED`, `IC.AGENT_STARTED/COMPLETED`, `IC.MCP_CACHE_HIT/MISS/SET/BYPASS/NOT_STORED`, `IC.MCP_TOOL_DEDUPED`, `IC.GLOBAL_BACKEND_SELECTED` |
+| NOC | error/unavailability/timeout/degradation | `NOC.RUNTIME_FAILED`, `NOC.MCP_TIMEOUT`, `NOC.LLM_FAILED` |
+| GRL | governance/guardrail/block/sanitization | `GRL.INPUT_BLOCKED`, `GRL.OUTPUT_BLOCKED`, `GRL.MASK_APPLIED` |
+
+Practical rule: don't create an event per line of code — only for decisions that matter.
+
+**Hybrid instrumentation**: automatic technical spans per graph node; LLM calls captured automatically via `ENABLE_LANGFUSE_OPENAI_AUTO_INSTRUMENTATION=true` (OpenAI-compatible providers); business events (IC/NOC/GRL) always manual via `_emit_ic/_emit_noc/_emit_grl`, **fail-open** (a missing or erroring observer never breaks the journey). Trace mode `LANGFUSE_TRACE_MODE=verbose|compact`.
+
+**Prompt management**: **does not use** Langfuse's prompt CMS — versions the prompt locally in YAML (`prompt_policy.yaml`, `id`/`version` fields). Langfuse comes in as a data source for batch evaluation (`EvaluationRun.source=langfuse`, `evals/offline`/`evals/certification` apps), not as the synchronous judging engine (that is a node of the graph itself, using the `judge` LLM profile).
+
+**Local setup (docker-compose)**: `langfuse-web`, `langfuse-worker`, PostgreSQL, ClickHouse, Redis, MinIO, plus MongoDB (the framework's own memory, not Langfuse's). Ports: Web/API 3005, MinIO S3 9090/Console 9091, MongoDB 27017, Postgres 5433, Redis 6379, ClickHouse 8124/9002. Flow: `docker compose up -d` → create org/project → generate a `pk-lf-.../sk-lf-...` key pair under Settings > API Keys → configure the backend's `.env`.
+
+---
+
+## 3. MCP integration pattern
+
+**MCP Gateway**: a single deployable app — `/v1/tools`, `/v1/tools/{name}`, `/v1/tools/{name}/invoke`, `/v1/servers`, `/health`, `/ready`. `ToolInvocation`/`ToolResult` contracts carrying `tenant_id`, `agent_id`, `business_context`, `metadata.request_id/trace_id`.
+
+**MCP Parameter Mapping** (`mcp_parameter_mapping.yaml`): translates canonical business keys (`customer_key`, `contract_key`, `session_key`) into the parameter name each tool expects.
+
+**MCP cache with Langfuse events**: key `tenant_id:agent_id:tool_name:hash(arguments)`, idempotent tools only, per-tool TTL. Event flow: miss → `IC.MCP_CACHE_MISS → IC.MCP_TOOL_EXECUTING → IC.MCP_TOOL_EXECUTED → IC.MCP_CACHE_SET`; hit → just `IC.MCP_CACHE_HIT`.
+
+**FastMCP** as the official MCP server default:
+```python
+from mcp.server.fastmcp import FastMCP
+mcp = FastMCP("telecom_mcp_server")
+
+@mcp.tool()
+def consultar_fatura(msisdn: str | None = None) -> dict: ...
+
+mcp.run(transport="streamable-http")
+```
+
+### ⚠️ Critical finding: shallow and potentially insecure authorization
+
+Their MCP Gateway's `authorization` block is a **static allowlist keyed by `agent_id`**:
+
+```yaml
+authorization:
+  default_policy: deny
+  agents:
+    telecom_contas:
+      allowed_tools: [consultar_fatura, consultar_pagamentos]
+```
+
+The `agent_id` field that governs this allowlist **arrives as data from the conversation's input payload** (`GatewayRequest`/`ToolInvocation`), not as a claim of an authenticated, independently verified MCP Client identity. Their own SPEC-004 admits it: *"conversational policy does not replace authentication, authorization, idempotency, or atomicity in the MCP Server."* There is no PDP/OPA/Cedar, no `policy_decision_id`/`entitlement_version`, no discovery/execution separation as two deliberately independent enforcement points.
+
+**This is structurally the anti-pattern our RI's ADR-001 forbids** ("authorization by the agent's self-declared role"). See ADR-023 for the decision on how to reuse the proxy engineering without inheriting this problem.
+
+---
+
+## 4. YAML configuration conventions
+
+| File | Purpose |
+|---|---|
+| `agents.yaml` | Agent registry (agent_id, isolated config paths, domain metadata) |
+| `routing.yaml` | Intents, keywords/examples/priority, `mcp_tools` per intent, `state_policies` |
+| `tools.yaml` | Tool catalog: target server, args schema, cache, `allowed_agents` |
+| `guardrails.yaml` | Input/output guardrails, enable/disable by code |
+| `judges.yaml` | Post-response evaluation criteria (deterministic/llm), threshold, sample_rate |
+| `llm_profiles.yaml` | Model/temperature profiles per component (router, judge, supervisor, RAG) |
+| `identity.yaml` | Channel field name aliases → canonical business keys (Identity Resolver) |
+| `mcp_parameter_mapping.yaml` | Canonical key → parameter name expected by each tool |
+| `mcp_servers.yaml` (+ `.docker.yaml`) | MCP server addresses/transport (local vs. Docker) |
+| `backends.yaml` | Agent backend registry + Agent Gateway's global routing rules |
+| `prompt_policy.yaml` | Base/persistent agent prompt, versioned (`id`, `version`) |
+| `tool_policies.yaml` | Classifies a tool as `read_only`/`transactional`, requires confirmation |
+| `authorization` (block in the MCP Gateway's config) | Tool allowlist by `agent_id` — **do not reuse as an authorization source** (see section 3) |
+
+---
+
+## 5. Security/identity model (SPEC-018) — compatibility
+
+- **Workload authentication** (5 modes: config_file/instance_principal/workload_identity/resource_principal/api_key) — addresses how the *framework* itself authenticates against OCI resources (Vault, GenAI, ADB). **Orthogonal** to our tool-entitlement problem.
+- **Authorization** — listed as 5 questions to answer, with no formal PDP; in practice resolved by the shallow allowlist described in section 3.
+- **Secrets, data protection, channel security** — compatible/complementary, do not touch entitlement.
+- **Auditing** — a good audit-trail model (user/channel, agent_id, tenant_id, tool, model, guardrail, judge score, trace_id), but with no equivalent to `decisionId`/`policyDecisionId`/`entitlementVersion` (our ADR-017), because there is no separate PDP generating those decisions.
+
+**Where it is clearly compatible/orthogonal**: `identity.yaml`/`BusinessContext` — normalization of **business** identity (end customer, contract), never a tool-authorization decision. Adoptable without tension with the security core.
+
+**Verdict**: SPEC-018 should not be used as a normative security spec for our RI — it mixes workload authentication with tool authorization without separating the two problems or proposing a PDP.
+
+---
+
+## 6. Deployment
+
+Not tied to OCI at the application layer — runs 100% locally via Docker Compose with generic, self-hostable services (Redis, MongoDB, PostgreSQL, ClickHouse, MinIO, Langfuse). OCI integrations (GenAI, ADB, Vault, workload identity) are pluggable options among several, not a development requirement.
+
+---
+
+## Reuse recommendations (applied in the RI via ADR-022/ADR-023)
+
+**Adopt as-is** (do not touch the security core): `identity.yaml`/BusinessContext · IC/NOC/GRL taxonomy + EventEnvelope · local Langfuse setup via docker-compose · `ConversationSummaryMemory` · RAG-vs-MCP heuristic (Hoshikawa §30.3-30.4: systems → MCP, documents → RAG) · `mcp_parameter_mapping.yaml` · FastMCP as the default MCP server.
+
+**Adopt with adaptation**: the *shape* of the MCP Gateway's proxying (catalog, `ToolInvocation`/`ToolResult`, cache with events, retry/timeout/circuit breaker) — but the ALLOW/DENY decision is always delegated to our own PEP/PDP, never to their `authorization` block. The Agent Gateway/Global Supervisor layer (two session levels, handoff) informs the design of EP-11/EP-12, keeping it strictly as UX/business routing, never as an authorization authority. `tool_policies.yaml` (read_only/transactional + confirmation) is conversational convenience, not a security control.
+
+**Do not copy as a source of security truth**: the `authorization` block keyed by a self-declared `agent_id` in the payload; SPEC-018 as authorization doctrine.