KEY-WP-0005-T02-T03: cached tenant_roles claim, close workplan

New internal/adapters/tenantengine package, mirroring
internal/adapters/{lldap,privacyidea,authelia}'s shape: Client.Roles()
calls tenant-engine's cache-read endpoint. Fails open by construction --
unreachable, non-200, malformed body, or a nil *Client all return
(nil, false), never an error to specially handle. Wired into
TokenHandler.TenantEngine (nil by default, existing tests unaffected);
token.go stamps tenant_roles only when ok.

7 adapter tests plus 3 TokenHandler-level tests proving the actual
required behavior end-to-end: present when reachable, token issuance still
200 with every other core claim intact when unreachable (tenant_roles
simply absent -- the literal done-criteria), absent when not configured.

Real bug found and fixed at the source, not worked around: the first live
cross-process check (real flex-auth, real tenant-engine, this adapter)
returned tenant_not_found for a tenant that existed -- tenant-engine's read
endpoint was keyed by its internal tenant_id, but key-cape only ever has
the tenant's profile identifier. Fixed in tenant-engine
(ADHOC-2026-07-24), re-verified with the same live three-process chain --
roles=[IAM] ok=true.

Workplan closed: T01-T03 done. Explicitly still open: client_credentials /
service-token issuance -- no such flow exists in token.go at all, a
materially larger separate piece of work than either task here.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

src/internal/adapters/tenantengine/adapter.go

@@ -0,0 +1,74 @@
+// Package tenantengine calls tenant-engine's cache-read endpoint at
+// token-issuance time to source the optional tenant_roles claim
+// (net-kingdom/canon/standards/iam-profile_v0.3.md, "Tenant Roles").
+//
+// This is the cache-read direction only, and it fails OPEN -- the opposite
+// of flex-auth's live-lookup adapter (flex-auth/internal/adapters/tenantengine),
+// which must fail closed. tenant_roles is documented as a cache callers
+// must never trust for privileged decisions (flex-auth re-validates live
+// before authorizing aal2-class actions); losing this claim at issuance
+// time is a performance regression, not a security one. Blocking login
+// because a cache source is briefly down would be the wrong trade.
+package tenantengine
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+)
+
+// Client fetches cached capability roles for a tenant.
+type Client struct {
+	BaseURL string
+	HTTP    *http.Client
+}
+
+// New returns a tenant-engine cache-read client. A nil httpClient gets a
+// short default timeout -- this call sits on the synchronous token-issuance
+// path and must not turn a cache miss into a slow login.
+func New(baseURL string, httpClient *http.Client) *Client {
+	if httpClient == nil {
+		httpClient = &http.Client{Timeout: 2 * time.Second}
+	}
+	return &Client{BaseURL: strings.TrimRight(baseURL, "/"), HTTP: httpClient}
+}
+
+// Roles fetches GET /tenants/{tenantID}/roles.
+//
+// Returns (nil, false) -- not an error -- on any failure: unreachable
+// tenant-engine, non-200 response, or a malformed body. Callers must treat
+// false as "omit the tenant_roles claim entirely", never as "emit an empty
+// or stale role list".
+func (c *Client) Roles(ctx context.Context, tenantID string) ([]string, bool) {
+	if c == nil || c.BaseURL == "" {
+		return nil, false
+	}
+
+	url := fmt.Sprintf("%s/tenants/%s/roles", c.BaseURL, tenantID)
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return nil, false
+	}
+
+	resp, err := c.HTTP.Do(req)
+	if err != nil {
+		return nil, false
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		return nil, false
+	}
+
+	var body struct {
+		Roles []string `json:"roles"`
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&body); err != nil {
+		return nil, false
+	}
+
+	return body.Roles, true
+}