diff --git a/.claude/agents/backend-engineer.md b/.claude/agents/backend-engineer.md
index a4bcb03..0221a83 100644
--- a/.claude/agents/backend-engineer.md
+++ b/.claude/agents/backend-engineer.md
@@ -16,6 +16,8 @@ HTTP API + services + business logic + DB schema/migrations + event producers/co
 
 **Pagination is mandatory on every unbounded collection endpoint** (baseline §4.1 / `governance/pagination-standard.md`, enforced by `gate-pagination`). Any LIST/collection GET you implement MUST: accept `limit` (apply the contract's default + **enforce the max server-side**, clamping over-max requests) and `cursor` (preferred — opaque, server-issued, encoding sort-key + tiebreaker) or `offset`; return the envelope `{ items, page: { nextCursor|null, hasMore, total? } }`; and walk the full set deterministically (no gaps/dupes under concurrent writes). **Your unit tests assert** the limit clamp, the envelope shape, and that the cursor pages through correctly. An endpoint is exempt only if inherently bounded/singleton and so annotated in the contract (`x-pagination: exempt`).
 
+**Identifiers are server-minted** (baseline §4.2 / `governance/identifier-standard.md`, enforced by `gate-identifier`). Never accept an `id` for a resource you are creating — mint it with `mintId()`/`mint_id()`, the only sanctioned constructor; never call `randomUUID()`/`uuid4()` for an entity id. Validate every incoming reference with `assertRef(type, id)` before use, and key polymorphic lookups on the `(type, id)` pair — never on a bare id. Type repository signatures with the branded `EntityId<T>` so a raw string off `req.body` cannot compile; store via `toUuid()` into a native `uuid` column and render the prefixed form at the serialization boundary. **An id is never a capability** — authorization still comes from the token and the policy engine. Graph create (`lid`/`idMap`) is provided by the shared middleware: mount it and implement nothing per-route.
+
 ## NOT your scope — never implement these (name them for the orchestrator)
 - **UI / frontend** (incl. any change to `design-system/` — `frontend-engineer` is its sole owner) → that's the `frontend-engineer`.
 - The **independent acceptance/contract test suite** → that's the `test-engineer` (API/contract) or `frontend-test-engineer` (UI e2e). You write your own unit tests, but you do NOT grade your own feature.
diff --git a/.claude/agents/contract-designer.md b/.claude/agents/contract-designer.md
index e6c32af..ccae156 100644
--- a/.claude/agents/contract-designer.md
+++ b/.claude/agents/contract-designer.md
@@ -11,7 +11,7 @@ You are the **contract designer** — the **API/event-contract lifecycle owner**
 
 ## Your scope (and ONLY this)
 You are the single owner of the API/event contracts — authoring, **versioning**, linting, and the generated client — not merely their initial design. From the user story / requirements (and the locked product decisions), design, freeze, and thereafter steward:
-- the **HTTP API contract** — an OpenAPI/Swagger spec (resources, paths, request/response schemas, error shapes, auth scopes, **pagination per the standard**, **explicit versioning** — bump the spec version on every change and keep a changelog). **Pagination (baseline §4.1 / `governance/pagination-standard.md`, enforced by `gate-pagination`):** every unbounded collection GET in the spec MUST declare `limit` (with default + max) + `cursor` (preferred, opaque) or `offset`, and the `{ items, page: { nextCursor|null, hasMore, total? } }` response envelope; mark a genuinely bounded/singleton endpoint `x-pagination: exempt` (+ `x-pagination-reason`). The contract is the single place these params/envelopes are defined so backend/test/UI all derive from one source;
+- the **HTTP API contract** — an OpenAPI/Swagger spec (resources, paths, request/response schemas, error shapes, auth scopes, **pagination per the standard**, **explicit versioning** — bump the spec version on every change and keep a changelog). **Pagination (baseline §4.1 / `governance/pagination-standard.md`, enforced by `gate-pagination`):** every unbounded collection GET in the spec MUST declare `limit` (with default + max) + `cursor` (preferred, opaque) or `offset`, and the `{ items, page: { nextCursor|null, hasMore, total? } }` response envelope; mark a genuinely bounded/singleton endpoint `x-pagination: exempt` (+ `x-pagination-reason`). The contract is the single place these params/envelopes are defined so backend/test/UI all derive from one source; **Identifiers (baseline §4.2 / `governance/identifier-standard.md`, enforced by `gate-identifier`):** a create body MUST NOT declare an `id`/`uuid` for the resource being created and MUST set `additionalProperties: false` — the owning service mints ids; every polymorphic reference (`entityId`, `ownerId`, `subjectId`, …) MUST carry a sibling type discriminator so no lookup resolves a bare id; a create that legitimately needs client-assigned ids is marked `x-client-assigned-id: allowed` (+ `x-client-assigned-id-reason`). Where a client must create linked entities in one request, model it as `lid` in / `idMap` out, scoped to this service's aggregate;
 - the **event contract** — the Kafka/AsyncAPI **Zod** event schemas + topic names/keys in the shared package, following the topic-prefix convention;
 - the **generated typed client** — run `openapi-typescript` to emit the `@<scope>/<svc>-client` package (private `publishConfig` + repository field), so UI, backend, and tests import the SAME types and drift becomes a compile error.
 **Lint the spec (Spectral)** on every revision, validate the event schemas, **version** the artifacts, regenerate the client, and **open/refresh the contract PR**. That PR — merged/frozen — is the dependency gate for the whole fan-out, and any later contract change re-enters through you, never around you.
diff --git a/.claude/agents/test-engineer.md b/.claude/agents/test-engineer.md
index 53212bb..10da6c1 100644
--- a/.claude/agents/test-engineer.md
+++ b/.claude/agents/test-engineer.md
@@ -14,6 +14,8 @@ Author the **API/service verification suite against the frozen spec** — contra
 
 **Pagination verification (mandatory).** For **every paginated endpoint in the frozen contract** (baseline §4.1 / `governance/pagination-standard.md`), your suite independently asserts: the endpoint accepts `limit` + `cursor|offset`; the response matches the `{ items, page: { nextCursor|null, hasMore, total? } }` envelope; **`limit` is enforced** (a request over the declared max is clamped, never returns more); and **the cursor walks the whole set** — paging with the returned `nextCursor` visits every item exactly once with no gaps/dupes and terminates (`nextCursor: null` / `hasMore: false`) at the end. An endpoint marked `x-pagination: exempt` is skipped (and you confirm it is genuinely bounded/singleton).
 
+**Identifier verification (mandatory).** For every create in the frozen contract (baseline §4.2 / `governance/identifier-standard.md`), your suite independently asserts: a body carrying an `id` is **rejected** (422), not silently accepted or echoed; an id minted for one entity type is **rejected** where another type is expected (the cross-type confusion this standard exists to stop); a polymorphic reference without its type discriminator is rejected; and — the case implementers most often miss — that **knowing an id grants nothing**: a caller authorized for entity A presenting a valid id for entity B is denied. Where graph create is used, assert `idMap` covers every first-class entity created and that a `lid` naming an entity this service does not own is rejected.
+
 ## File bugs in Jira when a test reveals a real defect
 A failing test against a real bug is a *valuable deliverable* — but the deliverable isn't just the red test, it's a **tracked ticket**. When your suite uncovers a genuine product defect, **file a bug in Jira** through `agile-manager`'s ticket standards: use the `ticket-creator` skill's **bug template** (and the Atlassian MCP) to create a well-formed bug — repro steps, expected vs actual, the failing test that proves it, severity, and a link back to the contract/acceptance criterion it violates. This routes the defect to the implementer (`backend-engineer` / `frontend-engineer`) instead of silently fixing it yourself. Keep the failing test in the suite so the bug stays provable until closed.
 
diff --git a/agent-templates/sync/role_loader.py b/agent-templates/sync/role_loader.py
index 9cf3685..d71f87f 100644
--- a/agent-templates/sync/role_loader.py
+++ b/agent-templates/sync/role_loader.py
@@ -4,9 +4,12 @@
 followed by the base guardrail block and any role-specific `system_append`.
 """
 import json
+import logging
 import os
 import re
 
+log = logging.getLogger(__name__)
+
 HERE = os.path.dirname(os.path.abspath(__file__))
 TEMPLATES_ROOT = os.path.dirname(HERE)                # agent-templates/
 REPO_ROOT = os.path.dirname(TEMPLATES_ROOT)           # repo root (personas live under .claude/agents)
@@ -76,14 +79,29 @@ def agent_payload(manifest):
     # or set-but-empty -> "") — the API rejects an empty/invalid url. Also drop the
     # matching mcp_toolset so the agent creates cleanly with only its configured servers;
     # re-provision after setting the URL to add the server + tool back.
+    #
+    # A drop is NEVER silent (FA-13): a required server logs at WARNING (its tools are
+    # missing until the URL is set), an `optional: true` server logs at INFO. A silent
+    # strip was the worst failure mode — an agent came up tool-less with no signal at all.
     servers = expand_env(manifest.get("mcp_servers", []))
     valid, dropped = [], set()
     for s in servers:
+        # `optional` is our own hint, never part of the API payload — strip it either way.
+        optional = bool(s.pop("optional", False))
         url = s.get("url", "")
         if url and "${" not in url:
             valid.append(s)
+            continue
+        name = s.get("name")
+        dropped.add(name)
+        reason = "url unset/empty" if not url else f"url unresolved ({url!r})"
+        if optional:
+            log.info("MCP server %r on agent %r dropped (optional): %s — continuing without it.",
+                     name, manifest.get("name"), reason)
         else:
-            dropped.add(s.get("name"))
+            log.warning("MCP server %r on agent %r dropped: %s — its tools will be MISSING until "
+                        "the URL is configured; re-provision after setting it.",
+                        name, manifest.get("name"), reason)
     tools = expand_env(manifest.get("tools", []))
     if dropped:
         tools = [t for t in tools
diff --git a/deploy/argocd/README.md b/deploy/argocd/README.md
index 66dcd41..86737ba 100644
--- a/deploy/argocd/README.md
+++ b/deploy/argocd/README.md
@@ -1,20 +1,62 @@
-# Argo CD wiring for FuzeAgent — a FuzeInfra delegation
+# Argo CD wiring for FuzeAgent — owned by this repo
 
-**FuzeInfra owns Argo `Application` and `AppProject` resources.** A product repo
-does not author them; two repos independently declaring an Application for the
-same workload is the competing-unrestricted-app failure FuzeMarket PR #61
-removed. This file is the **handoff spec**, not a manifest.
+`deploy/argocd/` holds FuzeAgent's app-of-apps and its three child Applications,
+and this repo owns them. FuzeInfra registers this directory once; ArgoCD
+self-syncs it thereafter.
 
-> **The manifests already in this directory are live wiring. This change
-> neither adds, edits nor removes any of them.**
+## Who owns what — correcting an earlier claim in this file
 
-| Existing manifest | Deploys |
+An earlier revision opened *"FuzeInfra owns Argo `Application` and `AppProject`
+resources. A product repo does not author them ... This file is the handoff spec,
+not a manifest."* **That is wrong**, and this directory already contradicted it —
+it has carried four Application manifests the whole time, and FuzeInfra's
+`argocd/applications/fuzeagent.yaml` is a *root* app-of-apps that points **at**
+them, i.e. it is built on the assumption that this repo authors them.
+
+FuzeInfra's own onboarding contract
+([`docs/CONSUMER_ONBOARDING_SHARED_CLUSTER.md`](https://github.com/izzywdev/FuzeInfra/blob/main/docs/CONSUMER_ONBOARDING_SHARED_CLUSTER.md)):
+
+> Boundary: the consumer owns its `deploy/**` (Helm/kustomize + Argo Applications
+> + sealed secrets). FuzeInfra owns the cluster, Argo, the tunnel, and the shared
+> datastores.
+
+FuzeInfra's `argocd-register` workflow is built around it: it takes a *"Path in
+the consumer repo holding the Argo Application/AppProject manifests"*,
+`kubectl apply`s the directory **once**, and then
+
+> After the first registration, ArgoCD polls and self-syncs the consumer's
+> `deploy/argocd` manifests.
+
+Registration is a **one-time owner action** — never `kubectl apply` from CI or by
+hand from here.
+
+That mistaken claim was propagated to several sibling repos in the same session
+and used to delete their Applications outright (FuzeService #32, FuzeSales #44),
+leaving them with no path to prod. Those deletions are being reverted. **None of
+FuzeAgent's manifests were deleted**, so nothing here changes but the wording.
+
+## The invariant that IS real
+
+**One Application per workload.** FuzeContact #33 and FuzeMarket #61 removed
+*duplicates*: two Applications with `prune: true` + `selfHeal: true` on the same
+namespace with disagreeing values, each pruning what the other did not create.
+
+FuzeAgent's manifests are **not** that — the app-of-apps recurses into
+`applications/`, and each child owns a **different** chart:
+
+| Manifest | Deploys |
 |---|---|
 | `app-of-apps.yaml` | recursive discovery of `applications/` |
-| `applications/fuzeagent.yaml` | `deploy/helm/fuzeagent` — the product chart this change extends |
+| `applications/fuzeagent.yaml` | `deploy/helm/fuzeagent` — the product chart |
 | `applications/fuzeagent-sealed.yaml` | the sealed-secret bundle |
 | `applications/a2a-shared.yaml` | `deploy/helm/a2a-shared` — **the family's only A2A server** |
 
+All four use `project: fuzeagent`, the restricted AppProject FuzeInfra owns at
+`argocd/projects/fuzeagent.yaml`. That project **does** exist — verified.
+
+> **These four manifests are live wiring. Do not delete them, and do not add a
+> second Application for any chart already listed above.**
+
 ## The A2A finding — read this first
 
 **`a2a-shared` is wired for deployment, not merely implemented.** This was the