Unify the Tenant CRD with enterprise-api -provision-tenant (lightweight)

Closes a gap named across CLAUDE.md/docs/architecture.md/deploy/README.md
since early Phase 4: the operator's Tenant CRD and -provision-tenant
were two disconnected mechanisms. The operator's reconciler generated a
K8s Secret with a locally-generated random password that authenticated
against nothing (nothing ever called ClickHouse to create a matching
user), and unconditionally claimed status.phase=Active the moment a
Tenant object existed -- actively misleading, not just incomplete.

Two unification shapes were considered (surfaced to the user via
AskUserQuestion, given the real difference in blast radius): the
operator's reconcile loop becoming a second real actor (new Postgres +
ClickHouse admin credentials flowing into the K8s controller, plus real
reconcile-loop idempotency/retry design for an inherently one-shot
external side effect), or keeping -provision-tenant as the sole real
actor and having it also sync its result into the CRD. Went with the
lighter option.

enterprise/internal/tenantcrd (new): a Syncer using the K8s dynamic
client (unstructured.Unstructured + a GroupVersionResource, not
deploy/operator's typed Tenant struct -- avoids a cross-module Go
dependency between two independently-versioned modules for one type).
Upserts the Tenant object, creates/updates a Secret with the *real*
ClickHouse credentials owned by that Tenant via an OwnerReference, then
patches status.{clickHouseDatabaseName,clickHouseSecretRef,
tantivyIndexPath}. Idempotent and safe to retry: never rotates a
credential across a re-sync, never overwrites a pre-existing
spec.displayName a human/GitOps process set.

cmd/enterprise-api/main.go's runProvisionTenant calls Sync when
TENANT_CRD_NAMESPACE is set (empty = no-op, same shape as every other
optional dependency in this codebase). Its "already active" refusal is
now split: ClickHouse re-provisioning is still refused (rotating a live
credential would break every open connection for no benefit), but CR
sync alone is now retryable using the credentials already on file in
rbacstore -- needed for retrying a previously-failed sync, or
backfilling CR sync for a tenant provisioned before this existed.

deploy/operator's reconciler rewritten to match: it never claims
PhaseActive on its own initiative anymore, only once
status.ClickHouseDatabaseName is non-empty (the field -provision-tenant,
and only -provision-tenant, sets). Phase is now a pure function of
{spec.suspended, status.ClickHouseDatabaseName != ""} recomputed every
reconcile, not toggled in place -- fixes a related bug the old code
would have hit once suspension was involved: un-suspending an
already-provisioned tenant needs to return straight to Active, which
isn't derivable from "last observed phase was Suspended" alone. The
reconciler no longer creates or manages any Secret, dropped its
`secrets` RBAC grant entirely, and gained zero new dependencies.

Helm chart: enterprise-api gets its own ServiceAccount/Role/RoleBinding
(get/list/create tenants, get/update/patch tenants/status, get/create/
update secrets -- least-privilege, scoped to the release namespace, not
a ClusterRole) and a TENANT_CRD_NAMESPACE env var, both gated on
tenantOperator.enabled. tenant-operator's ClusterRole loses the
secrets grant it no longer needs.

Verified in this environment: enterprise/internal/tenantcrd's tests run
against k8s.io/client-go's fake dynamic + typed clientsets (real client
library, fake transport, no cluster needed); deploy/operator's rewritten
tenant_controller_test.go runs against controller-runtime's fake
client, including new regression tests for the "must not claim Active
without confirmation" and "un-suspending returns to Active, not
Provisioning" properties; helm template + parsing the rendered YAML
confirms the RBAC split renders exactly as designed under both
tenantOperator.enabled=true/false. Not verified: an actual
-provision-tenant run against a real cluster with the operator watching
(no live cluster in this environment, same disclosed limitation as the
rest of /deploy). Docs updated in lockstep: CLAUDE.md, docs/architecture.md,
deploy/README.md, deploy/helm/sentry/README.md (including a corrected
"Trying the two-tenant example" walkthrough), phase-4-runbook.md (new
§11), enterprise/README.md. Also fixed two unrelated stale claims found
along the way: docs/architecture.md still said docker-compose.yml ran
plain api unconditionally (fixed in an earlier commit, doc not updated
then), and enterprise-api's own main.go doc comment still said Helm/
docker-compose wiring wasn't built yet.

CLAUDE.md

@@ -213,14 +213,19 @@ stricter still (creator/Admin/Owner only, closing a self-escalation
 path). Verified against a fake store (`api/dashboards/handler_test.go`);
 real integration tests exist but haven't run against a live Postgres,
 same disclosed gap as the rest of this phase's Postgres-backed pieces.
-What still keeps this phase from being done: ingest itself has no
-tenant concept for either storage engine (every record lands in the one
-shared ClickHouse database and Tantivy index no matter what —
-undesigned, not just unbuilt), and the two
-provisioning mechanisms (`deploy/operator`'s `Tenant` CRD and
-`enterprise-api -provision-tenant`) still aren't unified — running both
-for the same tenant ID is two separate operator actions today. Full
-accounting:
+`deploy/operator`'s `Tenant` CRD and `enterprise-api -provision-tenant`
+are now unified too, deliberately lightweight rather than making the
+K8s controller a second real actor: `-provision-tenant` stays the sole
+caller of ClickHouse/`rbacstore`, and (via a new
+`enterprise/internal/tenantcrd`, gated on `TENANT_CRD_NAMESPACE`) syncs
+its real result into the CRD — a real credential Secret, not the
+previous placeholder that authenticated against nothing, and status
+fields the reconciler derives `Phase`/`Ready` from instead of
+independently guessing "Active" the moment a Tenant object exists. What
+still keeps this phase from being done: ingest itself has no tenant
+concept for either storage engine (every record lands in the one shared
+ClickHouse database and Tantivy index no matter what — undesigned, not
+just unbuilt). Full accounting:
 `/docs/security/threat-model.md`; step-by-step verification procedure
 (not yet run against a live cluster in this environment):
 `/docs/phase-4-runbook.md`. The rest of this section describes the exit