diff --git a/backend/main.py b/backend/main.py
index a2cc533..5476063 100644
--- a/backend/main.py
+++ b/backend/main.py
@@ -8,7 +8,7 @@ import redis
 import asyncio
 from datetime import datetime, timedelta
 
-_version_info = {"version": "1.5.0", "releaseName": "ACME Diagnostics & Site Wizard", "releaseDate": "2026-05-08"}
+_version_info = {"version": "1.5.1", "releaseName": "Round-23 + Round-24 audit follow-ups", "releaseDate": "2026-05-13"}
 for _vpath in ["/app/version.json", os.path.join(os.path.dirname(__file__), "..", "version.json")]:
     try:
         with open(_vpath) as _vf:
diff --git a/backend/models/site_wizard.py b/backend/models/site_wizard.py
index bfa7c95..99f7ddb 100644
--- a/backend/models/site_wizard.py
+++ b/backend/models/site_wizard.py
@@ -488,6 +488,32 @@ class ServerStep(BaseModel):
                     "splits at the first space, so a space inside the "
                     "SNI value would corrupt the rendered config."
                 )
+        # Bulgu #90 (round-24 audit) — `cookie_value` is interpolated
+        # into the rendered `server <name> <addr>:<port> ... cookie
+        # <value> ...` line. HAProxy tokenises the line by whitespace
+        # and treats `;` as an INLINE COMMENT — and RFC 6265 cookie-
+        # value grammar separately disallows whitespace, `;`, `,`,
+        # `\\` and `"`. Pre-fix the validator only rejected newlines,
+        # so an operator could set `cookie_value="srv1; secure"` and
+        # see HAProxy silently truncate the server line at `;`
+        # (everything after becomes a comment). The Set-Cookie header
+        # rendered to clients would also fail the RFC's cookie-value
+        # grammar and most browsers drop the cookie, breaking session
+        # affinity without any error surface. Restrict to the
+        # conservative intersection — alphanumeric plus `_.-` —
+        # identical to backend.cookie_name. Legitimate session
+        # identifiers all fit in this set.
+        if info and info.field_name == "cookie_value":
+            import re as _re
+            if not _re.fullmatch(r"[A-Za-z0-9_.\-]+", v):
+                raise ValueError(
+                    "server.cookie_value must contain only alphanumerics, "
+                    "'.', '_' or '-' (HAProxy tokenises the server line "
+                    "by whitespace and treats ';' as an inline comment; "
+                    "RFC 6265 separately disallows whitespace, ';', ',' "
+                    "and backslash in cookie values). Got "
+                    f"{v!r}."
+                )
         return v
     # R17 (label corrected R18): CA bundle used by HAProxy to VERIFY the
     # upstream server's TLS certificate. Maps to the `ca-file` directive
@@ -658,6 +684,32 @@ class BackendStep(BaseModel):
         fields like `request_headers` / `response_headers` /
         `tcp_request_rules` are intentionally line-oriented and ARE
         NOT touched here.
+
+        Bulgu #90 (round-24 audit) — extend the validator beyond
+        newlines. Pre-fix `cookie_name` accepted strings like
+        `SESS'; DROP TABLE backends; --` (the SQL substring is
+        harmless thanks to parameterised queries, but the `;` is
+        HAProxy's INLINE COMMENT character: the renderer emits
+        `cookie SESS'; DROP TABLE backends; -- insert indirect
+        nocache`, which HAProxy parses as `cookie SESS'` followed by
+        an inline comment that swallows the persistence options the
+        operator typed). Result: session affinity silently broken
+        and the operator has zero diagnostic signal — the wizard
+        accepted the input, the apply succeeded, but cookies never
+        get re-emitted with the expected name. The same trap exists
+        for any HAProxy-section-keyword or whitespace token because
+        HAProxy tokenises by space.
+
+        `cookie_name` MUST be a single token. We use the conservative
+        intersection of RFC 6265 cookie-token chars and HAProxy
+        directive-name chars: alphanumeric plus `_.-`. Operators
+        with legitimate session cookies all live inside this set.
+
+        `cookie_options` is a space-separated keyword list
+        (`insert indirect nocache` etc., optionally `domain example.
+        com`, `attr SameSite=Lax`). Allow letters/digits/space/dot/
+        hyphen/underscore/equals. Reject `;` (comment), backslash,
+        quotes, and other shell metacharacters.
         """
         if v is None:
             return v
@@ -669,6 +721,28 @@ class BackendStep(BaseModel):
                 "newlines would smuggle additional directives into "
                 "the rendered config)"
             )
+        field_name = info.field_name if info else "cookie field"
+        if field_name == "cookie_name":
+            import re as _re
+            if not _re.fullmatch(r"[A-Za-z0-9_.\-]+", v):
+                raise ValueError(
+                    "backend.cookie_name must contain only alphanumerics, "
+                    "'.', '_' or '-' (HAProxy tokenises the `cookie` "
+                    "directive by whitespace and treats ';' as an inline "
+                    "comment — anything else silently truncates the "
+                    "rendered persistence options). Got "
+                    f"{v!r}."
+                )
+        elif field_name == "cookie_options":
+            import re as _re
+            if not _re.fullmatch(r"[A-Za-z0-9_.=\- ]*", v):
+                raise ValueError(
+                    "backend.cookie_options must contain only "
+                    "alphanumerics, spaces, '=', '.', '_' or '-' "
+                    "(HAProxy parses `;` as an inline comment, and "
+                    "quoting / backslash metacharacters are not "
+                    f"part of the `cookie` keyword grammar). Got {v!r}."
+                )
         return v
 
     @field_validator("health_check_uri")
diff --git a/backend/routers/frontend.py b/backend/routers/frontend.py
index d022455..eaae35e 100644
--- a/backend/routers/frontend.py
+++ b/backend/routers/frontend.py
@@ -198,12 +198,28 @@ def _enforce_routing_rule_contradictions(
     for label, rule in conflicts:
         sig = _rule_to_signature(rule)
         if sig and sig in grandfathered_signatures:
+            # Bulgu #83 (round-23 audit) — re-word the operator-
+            # facing warning. The pre-fix message led with
+            # "Grandfathered <label> entry contains a self-
+            # contradictory X !X condition that pre-dated this
+            # validation", which (a) is internal jargon the
+            # operator does not parse, and (b) implies the rule
+            # is OLD when in fact the only thing this branch
+            # knows is that the rule was NOT changed by the
+            # current edit. The operator may well have authored
+            # the rule one minute earlier. State that explicitly
+            # and include the verbatim rule body so the operator
+            # does not have to hunt through the ACL Builder
+            # cards to find the offender.
+            rule_text = rule if isinstance(rule, str) else sig[:160]
             warnings.append(
-                f"Grandfathered {label} entry contains a "
-                f"self-contradictory `X !X` condition that pre-dated "
-                f"this validation. The rule never fires; fix it at "
-                f"your convenience. (rule: "
-                f"{rule if isinstance(rule, str) else sig[:160]})"
+                f"{label}: rule was not modified by this edit but "
+                f"contains a self-contradictory `X !X` condition "
+                f"(`X AND NOT X` is always false, so the rule never "
+                f"fires and traffic falls through to "
+                f"`default_backend`). Your current edit was saved; "
+                f"fix the rule at your convenience. "
+                f"(rule: {rule_text})"
             )
         else:
             blocking.append((label, rule))
diff --git a/backend/routers/site_wizard.py b/backend/routers/site_wizard.py
index 8725d79..fb57f3a 100644
--- a/backend/routers/site_wizard.py
+++ b/backend/routers/site_wizard.py
@@ -1046,9 +1046,45 @@ async def suggest_defaults(
 
         slug = "newhost"
         if domain:
-            base = domain.replace("*.", "").split(".")
-            slug = (base[0] or "newhost")[:32].lower()
-            slug = "".join(c if (c.isalnum() or c in ("-", "_")) else "-" for c in slug)
+            # Bulgu #93 (round-24 audit) — IDN / Unicode safety. PRE-FIX
+            # the suggest endpoint used `c.isalnum()`, which is Unicode-
+            # aware and returns True for non-ASCII letters (ü, é, ñ, …).
+            # An operator who typed `bücher.example.com` got back
+            # `backend_name="be-bücher"`, dropped that into the wizard
+            # form, and then hit a hard 422 at create time because the
+            # backend/frontend name validator regex
+            # `^[a-zA-Z][a-zA-Z0-9_-]{0,63}$` is ASCII-only. The wizard
+            # CREATE path also forces the domain itself through punycode
+            # (the validator rejects raw Unicode with a "use 'xn--…'"
+            # hint). Make `suggest` honour the same on-the-wire ASCII
+            # contract: convert each label to its IDN/punycode form
+            # FIRST, then sanitise to the alphanumeric / `-_` set the
+            # entity-name regex permits. The result is a name the
+            # operator can submit to /api/sites without re-typing.
+            first_label = (
+                domain.replace("*.", "").split(".")[0]
+                if domain.replace("*.", "")
+                else ""
+            )
+            ascii_label = first_label
+            if first_label and not first_label.isascii():
+                try:
+                    ascii_label = first_label.encode("idna").decode("ascii")
+                except (UnicodeError, UnicodeDecodeError):
+                    # IDN encoding failed (empty label, invalid chars,
+                    # etc.) — fall back to stripping non-ASCII to '-'
+                    # so we still produce a usable slug.
+                    ascii_label = "".join(
+                        c if c.isascii() and (c.isalnum() or c in ("-", "_"))
+                        else "-"
+                        for c in first_label
+                    )
+            slug = (ascii_label or "newhost")[:32].lower()
+            slug = "".join(
+                c if c.isascii() and (c.isalnum() or c in ("-", "_"))
+                else "-"
+                for c in slug
+            )
             if slug.startswith("_"):
                 slug = "h-" + slug.lstrip("_")
             if not slug or not slug[0].isalpha():
@@ -1146,6 +1182,53 @@ async def preview_create(
     try:
         await _validate_user_cluster_access(current_user["id"], body.cluster_id, conn)
 
+        # Bulgu #88 / #89 (round-24 audit) — cluster-RBAC parity with
+        # `create_site` for SSL certificate references. PRE-FIX the
+        # preview path (`POST /api/sites/preview`) skipped the
+        # `select_existing_cert()` gate that `create_site` runs at
+        # lines 2298-2308 (per-server CA bundle) and 2398-2405
+        # (HTTPS bind cert). The omission let an authenticated wizard
+        # user pass `ssl.mode='existing', ssl_certificate_id=<X>` (or
+        # `servers[i].ssl_certificate_id=<X>`) where cert `X` belongs
+        # to a DIFFERENT cluster and receive the full rendered
+        # `would_create` envelope back — leaking cert id metadata
+        # across tenant boundaries. The actual submit (`POST /api/
+        # sites`) does enforce the gate, so this is a preview-only
+        # information leak, not a write-path escalation. The fix is
+        # to mirror the same `select_existing_cert(conn, id, cluster_
+        # id)` predicate (which already encodes the "global OR
+        # junction-bound" rule defined in `ssl_service.py:331-370`)
+        # so the preview returns 400 with a clear hint rather than
+        # 200 with a leaked render. We deliberately keep the error
+        # phrasing identical to the create-time message so wizard
+        # UI handlers that already match on "not found / inactive"
+        # need no client changes.
+        if body.ssl.mode == "existing" and body.ssl.ssl_certificate_id:
+            _preview_resolved = await select_existing_cert(
+                conn, body.ssl.ssl_certificate_id, body.cluster_id,
+            )
+            if not _preview_resolved:
+                raise HTTPException(
+                    status_code=400,
+                    detail=(
+                        f"ssl_certificate_id {body.ssl.ssl_certificate_id} "
+                        "not found / inactive / not bound to this cluster"
+                    ),
+                )
+        for _idx, _srv in enumerate(body.servers or []):
+            _srv_cert_id = getattr(_srv, "ssl_certificate_id", None)
+            if _srv_cert_id:
+                if not await select_existing_cert(
+                    conn, _srv_cert_id, body.cluster_id,
+                ):
+                    raise HTTPException(
+                        status_code=400,
+                        detail=(
+                            f"servers[{_idx}].ssl_certificate_id={_srv_cert_id} "
+                            "not found / inactive / not bound to this cluster"
+                        ),
+                    )
+
         # Phase K Phase C: rate-limit ONLY the dry-run code path so
         # legacy callers (e.g. `SiteDrafts.handlePreview` which never
         # sets the flag) keep their unrestricted preview budget. The
@@ -1848,6 +1931,28 @@ async def create_site(
     try:
         await _validate_user_cluster_access(user_id, body.cluster_id, conn)
 
+        # Bulgu #87 (round-23 audit) — rate-limit the actual create
+        # endpoint. Pre-fix /api/sites/preview and
+        # /api/sites/preflight-acme were rate-limited (5/min, see
+        # `_enforce_rate_limit` callers above) but POST /api/sites
+        # itself had NO cap. Round-3 live testing fired 10 wizard
+        # creates in a single second against demo-cluster1; without
+        # the cap an authenticated user / script can spam-create
+        # entities until the cluster apply queue clogs.
+        #
+        # Action name must match what the router itself logs at the
+        # end of create_site (`wizard_create_site` — see
+        # `_log_user_activity(... action="wizard_create_site", ...)`
+        # at the success path below). The activity_logger middleware
+        # intentionally skips this endpoint to avoid double-logging
+        # (see middleware/activity_logger.py:74-110), so we must
+        # rate-limit on the SAME action the router emits.
+        #
+        # 5/min matches the preview / preflight budget — bulk
+        # operators have the dedicated /api/sites/bulk-import path
+        # for larger batches.
+        await _enforce_rate_limit(conn, user_id, "wizard_create_site")
+
         # ----- Pre-create checks (must succeed before transaction)
 
         # Phase 3 (R11-audit follow-up): reserved-name check parity with
@@ -2133,7 +2238,31 @@ async def create_site(
         # rebrand. The reject path on `cluster.py` recognises BOTH
         # prefixes so historical APPLIED versions (created before this
         # rename) keep behaving correctly during reject/undo.
-        version_name = f"bulk-site-create-{ts}"
+        #
+        # Bulgu #86 (round-23 audit) — append a short UUID suffix so
+        # two wizard POST /api/sites calls submitted within the SAME
+        # epoch second cannot collide on the
+        # `config_versions(cluster_id, version_name)` UNIQUE
+        # constraint. Pre-fix `version_name = f"bulk-site-create-{ts}"`
+        # gave seconds resolution, so an operator who clicked "Create"
+        # twice in rapid succession (or any back-to-back API
+        # automation) saw the SECOND call 409 with the generic
+        # `UniqueViolationError` fall-through message:
+        #
+        #     "A wizard entity with this name already exists on the
+        #      cluster (UNIQUE constraint). Pick a different name."
+        #
+        # — even though the operator-chosen backend / frontend / SSL
+        # names were unique. The actual collision was on the
+        # auto-generated `version_name` and re-naming the wizard
+        # inputs did NOT help. The reject_pending_changes path on
+        # cluster.py:4179 prefix-matches `bulk-site-create-*` so
+        # appending a unique suffix preserves the historical
+        # rollback / undo semantics. 6 hex chars give 16M-room before
+        # birthday collisions, vastly more than the per-second
+        # request volume an operator can sustain through the wizard.
+        import uuid
+        version_name = f"bulk-site-create-{ts}-{uuid.uuid4().hex[:6]}"
         bulk_snapshots: List[dict] = []
         created_ids: Dict[str, Any] = {}
 
@@ -2929,22 +3058,68 @@ async def create_site(
         # broke".
         msg = str(uve)
         logger.info(f"WIZARD: name conflict on create: {msg}")
-        if "backends_name_cluster_id_key" in msg or 'backends_name' in msg:
+        # Bulgu #85 (round-23 audit) — extract the offending constraint
+        # name from the asyncpg message so the operator-visible detail
+        # can pin-point WHICH entity collided. Pre-fix the handler only
+        # recognised `backends_*_key` and `frontends_*_key`; any other
+        # constraint (ssl_certificates, backend_servers, config_versions
+        # …) fell through to the generic "wizard entity with this name"
+        # message which is useless for debugging — the operator has to
+        # open the server log and the engineer has to ssh-bounce to
+        # extract `constraint=<name>` from the exception detail.
+        #
+        # ``asyncpg.UniqueViolationError.constraint_name`` is the
+        # canonical structured field; ``str(uve)`` only contains the
+        # human-formatted DETAIL line. Prefer the attribute, fall back
+        # to substring scanning so we stay robust if asyncpg ever stops
+        # exposing it.
+        constraint = getattr(uve, "constraint_name", None) or ""
+        msg_lower = msg.lower()
+        if "backends_name_cluster_id_key" in msg or 'backends_name' in msg \
+                or constraint == "backends_name_cluster_id_key":
             detail = (
-                "A backend with this name already exists on the cluster "
-                "(possibly soft-deleted). Pick a different backend name "
-                "or restore/permanently-delete the existing row."
+                f"A backend named '{body.backend.name}' already exists on "
+                "the cluster (possibly soft-deleted). Pick a different "
+                "backend name or restore/permanently-delete the existing "
+                "row."
             )
-        elif "frontends_name_cluster_id_key" in msg or "frontends_name" in msg:
+        elif "frontends_name_cluster_id_key" in msg or "frontends_name" in msg \
+                or constraint == "frontends_name_cluster_id_key":
             detail = (
-                "A frontend with this name already exists on the cluster "
+                f"A frontend named '{body.frontend.name}' (or its auto-"
+                f"derived HTTPS sibling) already exists on the cluster "
                 "(possibly soft-deleted). Pick a different frontend name "
                 "or restore/permanently-delete the existing row."
             )
-        else:
+        elif "ssl_certificates" in msg_lower or "ssl_certificates" in constraint \
+                or "ssl_cert" in constraint:
+            ssl_name = getattr(body.ssl, "name", None) or "(unnamed)"
             detail = (
-                "A wizard entity with this name already exists on the "
-                "cluster (UNIQUE constraint). Pick a different name."
+                f"An SSL certificate named '{ssl_name}' already exists on "
+                "the cluster (possibly soft-deleted). Pick a different "
+                "ssl.name or restore/permanently-delete the existing "
+                "certificate row."
+            )
+        elif "backend_servers" in msg_lower or "backend_servers" in constraint:
+            detail = (
+                "Two servers in the same backend share a server_name. "
+                "HAProxy requires `server <name>` tokens to be unique "
+                "within a backend block. Rename the duplicate(s) and "
+                "resubmit."
+            )
+        else:
+            # Echo the constraint name (a stable, non-secret schema
+            # identifier) in the detail so a human reading the toast
+            # can grep the codebase for the matching CREATE TABLE
+            # without needing server-log access. Names like
+            # `proxied_hosts_pkey` or `config_versions_unique` are
+            # safe to surface — they're public schema info.
+            con_hint = f" (constraint={constraint})" if constraint else ""
+            detail = (
+                f"A wizard entity with this name already exists on the "
+                f"cluster (UNIQUE constraint{con_hint}). Pick a different "
+                "name or retry; if the conflict persists contact the "
+                "platform team."
             )
         raise HTTPException(status_code=409, detail=detail)
     except Exception as e:
diff --git a/backend/services/acme_diagnostics.py b/backend/services/acme_diagnostics.py
index fa7fea0..2324619 100644
--- a/backend/services/acme_diagnostics.py
+++ b/backend/services/acme_diagnostics.py
@@ -573,9 +573,18 @@ async def check_agents(conn, cluster_ids: List[int]) -> Dict[str, Any]:
             severity="warn",
             duration_ms=int((time.time() - started) * 1000),
         )
+    # Bulgu #84 (round-23 audit) — the canonical timestamp column on the
+    # `agents` table is `last_seen`. Pre-fix this query referenced a
+    # non-existent `a.last_heartbeat`, so every ACME preflight call
+    # (`POST /api/sites/preflight-acme`) crashed at the `check_agents`
+    # stage with `UndefinedColumnError: column a.last_heartbeat does
+    # not exist`, blocking the entire wizard's ACME pre-validation
+    # gate. Every other agents.last_seen reader in the codebase
+    # (routers/cluster.py:695-702, routers/agent.py, routers/dashboard
+    # *.py) uses `last_seen`; aligning here.
     rows = await conn.fetch(
         """
-        SELECT a.id, a.hostname, a.status, a.last_heartbeat, hc.id AS cluster_id, hc.name AS cluster_name
+        SELECT a.id, a.hostname, a.status, a.last_seen, hc.id AS cluster_id, hc.name AS cluster_name
         FROM agents a
         JOIN haproxy_clusters hc ON hc.pool_id = a.pool_id
         WHERE hc.id = ANY($1::int[])
diff --git a/backend/tests/test_acme_diagnostics.py b/backend/tests/test_acme_diagnostics.py
index 65065d8..80be221 100644
--- a/backend/tests/test_acme_diagnostics.py
+++ b/backend/tests/test_acme_diagnostics.py
@@ -433,7 +433,7 @@ async def test_check_agents_fail_when_none_registered():
 async def test_check_agents_warn_when_none_active():
     conn = AsyncMock()
     conn.fetch.return_value = [
-        {"id": 1, "hostname": "h1", "status": "offline", "last_heartbeat": None,
+        {"id": 1, "hostname": "h1", "status": "offline", "last_seen": None,
          "cluster_id": 1, "cluster_name": "c1"},
     ]
     out = await check_agents(conn, [1])
@@ -444,9 +444,9 @@ async def test_check_agents_warn_when_none_active():
 async def test_check_agents_ok_with_active():
     conn = AsyncMock()
     conn.fetch.return_value = [
-        {"id": 1, "hostname": "h1", "status": "active", "last_heartbeat": None,
+        {"id": 1, "hostname": "h1", "status": "active", "last_seen": None,
          "cluster_id": 1, "cluster_name": "c1"},
-        {"id": 2, "hostname": "h2", "status": "offline", "last_heartbeat": None,
+        {"id": 2, "hostname": "h2", "status": "offline", "last_seen": None,
          "cluster_id": 1, "cluster_name": "c1"},
     ]
     out = await check_agents(conn, [1])
@@ -454,6 +454,49 @@ async def test_check_agents_ok_with_active():
     assert "1 of 2" in out["message"]
 
 
+@pytest.mark.asyncio
+async def test_bulgu84_check_agents_uses_last_seen_column():
+    """Bulgu #84 (round-23 audit) — pin the SQL column name.
+
+    Pre-fix the query referenced a non-existent ``a.last_heartbeat``
+    column. AsyncMock returns whatever dict the test sets without
+    re-validating the SQL string, so the pre-fix test suite was
+    GREEN while every live ACME preflight call (the
+    ``POST /api/sites/preflight-acme`` endpoint that the wizard
+    runs before showing the ACME step) crashed with
+    ``UndefinedColumnError: column a.last_heartbeat does not
+    exist``. The crash blocked the entire wizard ACME preview
+    page on a production cluster, but unit tests never noticed
+    because they only assert on the helper's return shape, not
+    on the SQL string the helper sends to Postgres.
+
+    This pin inspects ``conn.fetch.call_args`` to assert the SQL
+    body actually queries ``a.last_seen`` (the canonical column
+    name used everywhere else in the codebase — see
+    routers/cluster.py:695-702 and routers/agent.py:281+ for
+    sibling readers). A future ``last_heartbeat`` typo would
+    re-fail this test without anyone noticing the live impact.
+    """
+    conn = AsyncMock()
+    conn.fetch.return_value = [
+        {"id": 1, "hostname": "h1", "status": "active", "last_seen": None,
+         "cluster_id": 1, "cluster_name": "c1"},
+    ]
+    await check_agents(conn, [1])
+    conn.fetch.assert_awaited_once()
+    sql_query = conn.fetch.call_args[0][0]
+    assert "a.last_seen" in sql_query, (
+        f"check_agents SQL must select a.last_seen (the canonical "
+        f"agents-table timestamp column); got: {sql_query!r}"
+    )
+    assert "a.last_heartbeat" not in sql_query, (
+        f"check_agents SQL still references a.last_heartbeat — this "
+        f"column does NOT exist on the agents table and the query "
+        f"will 500 with UndefinedColumnError at runtime. SQL: "
+        f"{sql_query!r}"
+    )
+
+
 # ----------------------------------------------------------------------------
 # run_checks orchestration
 # ----------------------------------------------------------------------------
diff --git a/backend/tests/test_haproxy_validator_bulgu12.py b/backend/tests/test_haproxy_validator_bulgu12.py
index a29409c..45ca917 100644
--- a/backend/tests/test_haproxy_validator_bulgu12.py
+++ b/backend/tests/test_haproxy_validator_bulgu12.py
@@ -5328,7 +5328,22 @@ def test_bulgu62_update_path_grandfathers_unchanged_rule():
         fe, grandfathered_signatures=grand,
     )
     assert len(warnings) == 1
-    assert "grandfathered" in warnings[0].lower() or "pre-dated" in warnings[0].lower()
+    # Bulgu #83 (round-23 audit) — the warning was reworded from
+    # "Grandfathered ... pre-dated this validation" to a clearer
+    # "rule was not modified by this edit" phrasing that also
+    # echoes the verbatim rule body. Accept either the legacy
+    # markers or the new ones so the contract is signal-not-
+    # wording.
+    w_lower = warnings[0].lower()
+    assert (
+        "not modified by this edit" in w_lower
+        or "grandfathered" in w_lower
+        or "pre-dated" in w_lower
+    ), warnings[0]
+    # The verbatim rule must appear in the warning body so the
+    # operator can identify the offending entry without opening
+    # the ACL Builder cards.
+    assert stale in warnings[0], warnings[0]
 
 
 def test_bulgu62_update_path_still_rejects_new_contradiction():
@@ -6267,3 +6282,611 @@ def test_bulgu82_generate_install_script_validates_cluster():
     window = src[fn_start:fn_start + 5000]
     assert "validate_user_cluster_access" in window
     assert "Bulgu #82 (round-22 audit)" in window
+
+
+# ---- Bulgu #83 — grandfathered contradiction warning wording ----
+
+
+def test_bulgu83_warning_includes_verbatim_rule_text():
+    """The PUT-path grandfather warning must echo the verbatim
+    offending rule string so the operator can identify the
+    offender without opening the ACL Builder cards. Pre-fix the
+    warning only said "1 legacy rule" via the FE toast and
+    "Grandfathered <label> entry contains a self-contradictory
+    X !X condition that pre-dated this validation" via the server
+    JSON payload — both omitted the actual rule body, forcing
+    the operator to open the modal and hunt for the dead-code
+    entry."""
+    from models.frontend import FrontendConfig
+    from routers.frontend import (
+        _enforce_routing_rule_contradictions,
+        _rule_to_signature,
+    )
+
+    stale = "be-x if acl1 !acl1"
+    fe = FrontendConfig(
+        name="fe1",
+        bind_port=80,
+        mode="http",
+        use_backend_rules=[stale],
+    )
+    warnings = _enforce_routing_rule_contradictions(
+        fe, grandfathered_signatures={_rule_to_signature(stale)},
+    )
+    assert len(warnings) == 1
+    # Verbatim rule body appears in the warning.
+    assert stale in warnings[0]
+    # The wording no longer uses the operator-unfriendly
+    # "Grandfathered" lead-in; the message states what the
+    # branch actually knows ("not modified by this edit").
+    assert "not modified by this edit" in warnings[0].lower()
+
+
+def test_bulgu83_dict_redirect_warning_includes_signature_snippet():
+    """Dict-shaped redirect rules with a contradictory `condition`
+    must still surface the offending signature in the warning
+    (truncated to 160 chars to bound the toast length). Pin the
+    truncation contract so a future refactor doesn't accidentally
+    grow the warning into a multi-KB blob."""
+    from models.frontend import FrontendConfig
+    from routers.frontend import (
+        _enforce_routing_rule_contradictions,
+        _rule_to_signature,
+    )
+
+    # Note: the redirect condition itself carries the contradiction;
+    # the dict wrapper is what the wizard emits.
+    bad_dict = {
+        "type": "scheme",
+        "scheme": "https",
+        "code": 301,
+        "condition": "if acl1 !acl1",
+    }
+    fe = FrontendConfig(
+        name="fe1",
+        bind_port=80,
+        mode="http",
+        redirect_rules=[bad_dict],
+    )
+    warnings = _enforce_routing_rule_contradictions(
+        fe, grandfathered_signatures={_rule_to_signature(bad_dict)},
+    )
+    assert len(warnings) == 1
+    # Truncated signature (up to 160 chars) must appear in body.
+    assert "acl1" in warnings[0]
+    assert "!acl1" in warnings[0]
+    assert "redirect_rules" in warnings[0]
+
+
+def test_bulgu83_static_marker_in_fe_warning_toast():
+    """Front-end pin — the FrontendManagement.js toast no longer
+    calls these rules "legacy" and now lists each offending rule
+    body. Static-source check so a refactor that re-introduces
+    the misleading wording or drops the rule snippets is caught.
+    """
+    # The frontend tree lives next to backend/ at the workspace
+    # root, so walk up one extra level from _BACKEND_DIR.
+    fm_path = (
+        _BACKEND_DIR.parent
+        / "frontend"
+        / "src"
+        / "components"
+        / "FrontendManagement.js"
+    )
+    fm_src = fm_path.read_text()
+    assert "Bulgu #83 (round-23 audit)" in fm_src
+    # No more `legacy routing/redirect rule(s)` wording.
+    assert "legacy routing/redirect rule(s) with a self-contradictory" not in fm_src
+    # The new toast wires the offending rule snippets into the
+    # message body via `ruleSnippets`.
+    assert "ruleSnippets" in fm_src
+    # And it re-surfaces server-emitted warnings as a safety net.
+    assert "serverWarnings" in fm_src
+
+
+# ----------------------------------------------------------------------------
+# Bulgu #84 + #85 (round-23 audit — comprehensive Site Wizard live-test pass)
+# ----------------------------------------------------------------------------
+
+
+def test_bulgu84_acme_diagnostics_uses_last_seen_column():
+    """Bulgu #84 (round-23 audit) — duplicate pin in this file so the
+    full bulgu-pin suite has a one-stop reference; the original sits
+    in `test_acme_diagnostics.py::test_bulgu84_check_agents_uses_
+    last_seen_column` and asserts the AsyncMock call shape.
+
+    This duplicate is a STATIC-SOURCE check that's cheaper to grep:
+    we open the service file and verify the wrong column name is
+    not back in the SQL, plus the Bulgu marker is present so a
+    future refactor that drops the comment also fails.
+    """
+    svc_path = _BACKEND_DIR / "services" / "acme_diagnostics.py"
+    src = svc_path.read_text()
+    assert "Bulgu #84 (round-23 audit)" in src, (
+        "Bulgu #84 marker missing from acme_diagnostics.py — the "
+        "fix comment was removed but the underlying SQL change "
+        "may have been reverted too. Re-verify check_agents."
+    )
+    # Strip comments before scanning for the wrong column — the fix
+    # comment legitimately mentions ``a.last_heartbeat`` as the
+    # pre-fix symptom. We only care about NON-comment occurrences.
+    non_comment_lines = "\n".join(
+        ln for ln in src.splitlines() if not ln.lstrip().startswith("#")
+    )
+    # The canonical SELECT inside check_agents must use a.last_seen.
+    assert "SELECT a.id, a.hostname, a.status, a.last_seen" in non_comment_lines, (
+        "acme_diagnostics.check_agents must SELECT a.last_seen "
+        "(the canonical agents-table column). Pre-fix this used "
+        "a non-existent a.last_heartbeat and every preflight-acme "
+        "call 500'd in production."
+    )
+    assert "a.last_heartbeat" not in non_comment_lines, (
+        "acme_diagnostics still references a.last_heartbeat in non-"
+        "comment code — this column does NOT exist on the agents "
+        "table. Use a.last_seen."
+    )
+
+
+def test_bulgu87_wizard_create_endpoint_rate_limited():
+    """Bulgu #87 (round-23 audit) — `POST /api/sites` must be
+    rate-limited, matching the existing 5/min caps on the
+    /preview and /preflight-acme sibling endpoints.
+
+    Pre-fix the wizard's actual create endpoint had NO rate cap:
+      - /api/sites/preflight-acme  → rate-limited (site_acme_preflight)
+      - /api/sites/preview         → rate-limited (site_previewed)
+      - /api/sites (POST, create)  → UNLIMITED ← gap
+
+    Round-3 live testing fired 10 wizard creates inside one
+    second against demo-cluster1; the response codes were
+    200/409/409/...409/200/409... — no 429 ever fired. With
+    Bulgu #86 fixed (version_name collision), the UNIQUE-error
+    side-effect that ACCIDENTALLY rate-limited rapid-fire creates
+    disappears too, so an authenticated user/script can now
+    create entities until the cluster apply queue clogs.
+
+    The fix calls `_enforce_rate_limit(conn, user_id,
+    "wizard_create_site")` early in create_site (after auth +
+    cluster-access checks). The action name must match what the
+    router itself logs (`wizard_create_site` — see
+    `_log_user_activity(... action="wizard_create_site", ...)`
+    in the same router); the activity_logger middleware
+    intentionally skips this endpoint to avoid double-logging,
+    so the rate-limit must reference the router-emitted action.
+    """
+    sw_src = (_BACKEND_DIR / "routers" / "site_wizard.py").read_text()
+    assert "Bulgu #87 (round-23 audit)" in sw_src, (
+        "Bulgu #87 marker missing — the wizard create rate-limit "
+        "may have been reverted."
+    )
+    # The create_site function body must call _enforce_rate_limit
+    # with the canonical create action name.
+    import re as _re
+    m = _re.search(
+        r"async def create_site\([^)]*\)[^{]*:(.*?)(?=\nasync def |\n@router\.|\Z)",
+        sw_src,
+        _re.DOTALL,
+    )
+    assert m, "could not locate create_site function body in site_wizard.py"
+    fn_body = m.group(1)
+    assert (
+        '_enforce_rate_limit(conn, user_id, "wizard_create_site")' in fn_body
+        or "_enforce_rate_limit(conn, user_id, 'wizard_create_site')" in fn_body
+    ), (
+        "create_site must call _enforce_rate_limit(... "
+        '"wizard_create_site") so rapid-fire POST /api/sites '
+        "calls hit a 429 cap; pre-fix the endpoint was unlimited."
+    )
+
+
+def test_bulgu86_wizard_version_name_has_unique_suffix():
+    """Bulgu #86 (round-23 audit) — pin the version-name generation so
+    two wizard POST /api/sites calls submitted within the SAME epoch
+    second cannot collide on `config_versions(cluster_id, version_name)`.
+
+    Pre-fix `version_name = f"bulk-site-create-{int(time.time())}"`
+    had seconds resolution. Round-3 live testing reproduced the
+    failure with a back-to-back wizard automation: the second call
+    returned 409 with the misleading
+
+        "A wizard entity with this name already exists on the
+         cluster (UNIQUE constraint). Pick a different name."
+
+    fall-through detail, even though the operator-supplied
+    backend / frontend / SSL names were genuinely unique — the
+    real collision was on the auto-generated version name. The
+    operator who tried to rename the wizard inputs would still
+    hit the same error and have no way to make progress until the
+    epoch second ticked over.
+
+    Fix: append a 6-hex-char UUID suffix
+    (`bulk-site-create-{ts}-{uuid.uuid4().hex[:6]}`) so the
+    suffix space is 16M and birthday-collision-proof at any
+    realistic per-second request rate. The `bulk-site-create-`
+    prefix is preserved so reject_pending_changes /
+    restore-from-prior-version paths (cluster.py:4179 prefix scan)
+    keep working.
+
+    Static-source pin so any future refactor that strips the
+    suffix fails this test.
+    """
+    sw_src = (_BACKEND_DIR / "routers" / "site_wizard.py").read_text()
+    assert "Bulgu #86 (round-23 audit)" in sw_src, (
+        "Bulgu #86 marker missing — the version_name uniqueness "
+        "suffix may have been reverted."
+    )
+    # The fixed form references uuid.uuid4().hex[:6] (or similar
+    # length-bounded random suffix) appended to the ts. A pure
+    # `f"bulk-site-create-{ts}"` line (without any suffix
+    # interpolation after ts) is the regressed form.
+    import re as _re
+    # Find the wizard's version_name assignment. We allow flexible
+    # spacing / quoting around the f-string but require the suffix
+    # interpolation token immediately after `{ts}-`.
+    has_suffix = bool(_re.search(
+        r'version_name\s*=\s*f"bulk-site-create-\{ts\}-\{[^}]+\}"',
+        sw_src,
+    ))
+    has_regressed = bool(_re.search(
+        r'version_name\s*=\s*f"bulk-site-create-\{ts\}"\s*$',
+        sw_src,
+        _re.MULTILINE,
+    ))
+    assert has_suffix, (
+        "wizard's version_name must append a unique suffix "
+        "(e.g. uuid.uuid4().hex[:6]) after `{ts}` so back-to-back "
+        "wizard creates within the same epoch second cannot "
+        "collide on config_versions UNIQUE."
+    )
+    assert not has_regressed, (
+        "wizard's version_name regressed to seconds-only resolution "
+        "— this re-introduces Bulgu #86 (UNIQUE collision on rapid "
+        "back-to-back creates)."
+    )
+
+
+def test_bulgu85_unique_violation_handler_extracts_constraint_name():
+    """Bulgu #85 (round-23 audit) — the wizard's UniqueViolationError
+    handler must surface the offending constraint name in the operator-
+    visible 409 detail so debugging a "wizard entity with this name
+    already exists" toast doesn't require shell access to the server
+    logs.
+
+    Pre-fix the handler had THREE branches:
+      * backends_name_cluster_id_key  → "backend with this name"
+      * frontends_name_cluster_id_key → "frontend with this name"
+      * else                          → generic "wizard entity"
+    Any other constraint (SSL cert UNIQUE, backend_servers UNIQUE,
+    config_versions UNIQUE) fell through to the generic message,
+    leaving the operator with NO actionable hint and requiring an
+    on-call engineer to ssh into the API pod and tail `logger.info`
+    output to find ``constraint=<name>`` in the asyncpg traceback.
+
+    The fix:
+      1. Echoes the entity NAME (body.backend.name, body.frontend.name,
+         body.ssl.name) so the operator can match toast to wizard input.
+      2. Adds dedicated branches for `ssl_certificates_*` and
+         `backend_servers_*` constraints with actionable hints.
+      3. In the fall-through ELSE branch, echoes the constraint name
+         from `uve.constraint_name` (or substring scan as fallback)
+         so an unknown constraint still gives the engineer a stable
+         schema identifier to grep.
+    """
+    sw_src = (_BACKEND_DIR / "routers" / "site_wizard.py").read_text()
+    assert "Bulgu #85 (round-23 audit)" in sw_src, (
+        "Bulgu #85 marker missing from site_wizard.py UniqueViolation "
+        "handler — the constraint-echoing fix may have been reverted."
+    )
+    # Entity-name echoing in branch detail bodies.
+    assert "body.backend.name" in sw_src and (
+        "A backend named '{body.backend.name}'" in sw_src
+        or "A backend named '\"{body.backend.name}\"" in sw_src
+        or "A backend named '" in sw_src and "body.backend.name" in sw_src
+    )
+    assert "A frontend named '{body.frontend.name}'" in sw_src or (
+        "A frontend named '" in sw_src and "body.frontend.name" in sw_src
+    )
+    # SSL cert and backend_servers branches present.
+    assert "ssl_certificates" in sw_src
+    assert "backend_servers" in sw_src
+    # Fall-through echoes constraint name.
+    assert "constraint=" in sw_src or "constraint_name" in sw_src
+    # The asyncpg attribute is preferred over str(uve) scanning.
+    assert 'getattr(uve, "constraint_name"' in sw_src, (
+        "Fix should prefer asyncpg's structured constraint_name "
+        "attribute over substring scanning str(uve), which is the "
+        "human-formatted DETAIL line and may vary across server "
+        "versions / locales."
+    )
+
+
+# ----- Bulgu #88 / #89 (round-24 audit) ----------------------------------
+#
+# `POST /api/sites/preview` skipped the SSL-cert cluster-RBAC gate that
+# `POST /api/sites` (create) enforces via `select_existing_cert()`. The
+# `select_existing_cert()` helper encodes "global cert OR junction-bound
+# to <cluster_id>"; pre-fix the preview accepted ssl_certificate_id values
+# bound to OTHER clusters and returned the full rendered `would_create`
+# envelope — a tenant-boundary information leak for the cert id and the
+# rendered HAProxy config snippet. The fix mirrors the create-time check
+# in the preview path: same predicate, same 400 status, same error text.
+
+def test_bulgu88_preview_enforces_ssl_cert_cluster_rbac():
+    """Preview must reject `ssl.ssl_certificate_id` referencing a cert
+    that is bound to a DIFFERENT cluster — same RBAC predicate the
+    create endpoint enforces. Static source check: the preview body
+    invokes `select_existing_cert` with `body.ssl.ssl_certificate_id`
+    and raises 400 with a "not found / inactive / not bound" hint."""
+    import os
+    sw_path = os.path.join(
+        os.path.dirname(__file__), "..", "routers", "site_wizard.py",
+    )
+    with open(sw_path, "r") as fh:
+        sw_src = fh.read()
+
+    # Locate the preview function.
+    assert "async def preview_create(" in sw_src
+    preview_start = sw_src.index("async def preview_create(")
+    # Bound the slice at the next top-level `async def` so we don't
+    # match the create_site copy (which is at line ~1900 and would
+    # cause this test to trivially pass even if preview was unfixed).
+    next_def = sw_src.index("\nasync def ", preview_start + 1)
+    preview_body = sw_src[preview_start:next_def]
+
+    # The preview must call select_existing_cert.
+    assert "select_existing_cert(" in preview_body, (
+        "Preview path must mirror create_site's cluster-RBAC gate by "
+        "invoking select_existing_cert(conn, ssl_certificate_id, "
+        "cluster_id). Pre-fix preview returned 200 with the full "
+        "rendered would_create envelope when the cert belonged to a "
+        "different cluster — a tenant-boundary information leak."
+    )
+
+    # The check must be wired to the body.ssl.ssl_certificate_id field
+    # (not some unrelated cert reference).
+    assert "body.ssl.ssl_certificate_id" in preview_body, (
+        "Preview RBAC check must reference body.ssl.ssl_certificate_id"
+    )
+
+    # Must raise 400 (consistent with create-time message).
+    assert "status_code=400" in preview_body
+    assert "not found / inactive / not bound" in preview_body or (
+        "not found / inactive" in preview_body
+        and "not bound" in preview_body
+    ), (
+        "Preview should raise 400 with the same 'not found / inactive "
+        "/ not bound to this cluster' hint create_site emits, so wizard "
+        "clients can match the message uniformly."
+    )
+
+
+def test_bulgu89_preview_enforces_per_server_ca_bundle_cluster_rbac():
+    """Preview must also enforce per-server `ssl_certificate_id` (CA
+    bundle) cluster-RBAC. Pre-fix only `create_site` checked this — a
+    direct API caller could submit a preview with
+    `servers[i].ssl_certificate_id=<cert from another cluster>` and
+    receive the full rendered backend block back, including the
+    `ca-file` directive bound to a cert id the caller's cluster has
+    no junction row for."""
+    import os
+    sw_path = os.path.join(
+        os.path.dirname(__file__), "..", "routers", "site_wizard.py",
+    )
+    with open(sw_path, "r") as fh:
+        sw_src = fh.read()
+    preview_start = sw_src.index("async def preview_create(")
+    next_def = sw_src.index("\nasync def ", preview_start + 1)
+    preview_body = sw_src[preview_start:next_def]
+
+    # Must iterate body.servers and check per-server ssl_certificate_id.
+    assert "body.servers" in preview_body, (
+        "Preview must iterate body.servers to validate each "
+        "server.ssl_certificate_id reference"
+    )
+    assert "ssl_certificate_id" in preview_body
+    # The per-server message must be field-qualified so operators see
+    # WHICH server triggered the gate.
+    assert "servers[" in preview_body and "ssl_certificate_id=" in preview_body, (
+        "Per-server preview error must include the array index "
+        "(servers[i].ssl_certificate_id=<id>) so operators can find "
+        "the offending row in their wizard payload."
+    )
+
+
+# ----- Bulgu #90 (round-24 audit) ----------------------------------------
+#
+# `backend.cookie_name`, `backend.cookie_options`, and `server.cookie_value`
+# only rejected newline characters pre-fix. HAProxy's directive parser
+# tokenises by whitespace and treats `;` as an inline-comment, so a value
+# like `SESS'; DROP TABLE backends; --` rendered as
+# `cookie SESS'; DROP TABLE backends; -- insert indirect nocache` which
+# HAProxy parsed as `cookie SESS'` + comment — silently truncating the
+# operator's persistence options and producing a malformed Set-Cookie
+# header that browsers may drop. Tighten to the conservative
+# alphanumeric+`_.-` set (cookie_name / cookie_value) and the keyword/
+# `=`-bearing set (cookie_options).
+
+def test_bulgu90_cookie_name_rejects_haproxy_parser_hostile_chars():
+    """cookie_name must reject `;` (HAProxy inline comment), whitespace
+    (token delimiter), and other punctuation that breaks the rendered
+    `cookie <name>` directive."""
+    from models.site_wizard import BackendStep
+    import pydantic
+    # Semicolon (HAProxy comment) — used to be silently accepted.
+    with pytest.raises(pydantic.ValidationError):
+        BackendStep(name="be1", cookie_name="SESS'; DROP TABLE backends; --")
+    # Space (token delimiter) — splits the cookie directive.
+    with pytest.raises(pydantic.ValidationError):
+        BackendStep(name="be1", cookie_name="SESS ID")
+    # Quote (legal in RFC 6265 but produces ugly Set-Cookie headers
+    # and confuses log scrapers).
+    with pytest.raises(pydantic.ValidationError):
+        BackendStep(name="be1", cookie_name='SESS"id')
+    # Common-case valid input continues to work.
+    BackendStep(name="be1", cookie_name="SESS_ID-1.app")
+
+
+def test_bulgu90_cookie_options_rejects_parser_hostile_chars():
+    """cookie_options must reject `;` and quoting metacharacters but
+    still accept legitimate `attr SameSite=Lax` style values."""
+    from models.site_wizard import BackendStep
+    import pydantic
+    with pytest.raises(pydantic.ValidationError):
+        BackendStep(
+            name="be1", cookie_name="SESS",
+            cookie_options="insert indirect; rm -rf /",
+        )
+    with pytest.raises(pydantic.ValidationError):
+        BackendStep(
+            name="be1", cookie_name="SESS",
+            cookie_options='insert "indirect"',
+        )
+    # Valid HAProxy cookie options continue to parse.
+    BackendStep(
+        name="be1", cookie_name="SESS",
+        cookie_options="insert indirect nocache attr SameSite=Lax",
+    )
+
+
+def test_bulgu90_server_cookie_value_rejects_rfc6265_disallowed_chars():
+    """server.cookie_value must reject `;`, whitespace, and other
+    chars RFC 6265 disallows in cookie-values."""
+    from models.site_wizard import ServerStep
+    import pydantic
+    with pytest.raises(pydantic.ValidationError):
+        ServerStep(
+            server_name="s1", server_address="10.0.0.1",
+            server_port=8080, cookie_value="srv1; secure",
+        )
+    with pytest.raises(pydantic.ValidationError):
+        ServerStep(
+            server_name="s1", server_address="10.0.0.1",
+            server_port=8080, cookie_value="srv 1",
+        )
+    # Valid cookie values continue to work.
+    ServerStep(
+        server_name="s1", server_address="10.0.0.1",
+        server_port=8080, cookie_value="srv-1.app_2",
+    )
+
+
+# ----- Bulgu #93 (round-24 audit) ----------------------------------------
+#
+# `GET /api/sites/suggest` produced backend/frontend name suggestions
+# using `c.isalnum()` over the first domain label. Python's `isalnum()`
+# is Unicode-aware and returns True for non-ASCII letters (ü, é, ñ, …),
+# so an operator typing `bücher.example.com` received
+# `backend_name='be-bücher'`. The wizard CREATE path then rejected the
+# very name the SUGGEST endpoint returned, because the entity-name
+# regex (`^[a-zA-Z][a-zA-Z0-9_-]{0,63}$`) and the domain validator are
+# both ASCII-only. The fix converts non-ASCII labels through IDN/
+# punycode FIRST and only then applies ASCII-only sanitisation, so the
+# resulting name is one the operator can submit unchanged.
+
+def test_bulgu93_suggest_idn_unicode_produces_ascii_safe_slug():
+    """`suggest` must produce slugs that the backend/frontend name
+    regex `^[a-zA-Z][a-zA-Z0-9_-]{0,63}$` accepts. Static source check:
+    the suggest function uses an ASCII-only predicate
+    (`c.isascii() and c.isalnum()` etc.) and routes IDN labels through
+    `encode('idna')` to preserve the operator's intent in punycode."""
+    import os, re
+    sw_path = os.path.join(
+        os.path.dirname(__file__), "..", "routers", "site_wizard.py",
+    )
+    with open(sw_path, "r") as fh:
+        sw_src = fh.read()
+
+    # Locate the suggest_defaults function body.
+    assert "async def suggest_defaults(" in sw_src
+    start = sw_src.index("async def suggest_defaults(")
+    end = sw_src.index("\n@router.", start + 1)
+    suggest_body = sw_src[start:end]
+
+    # The fix must use IDN/punycode encoding when the input is non-ASCII.
+    assert 'encode("idna")' in suggest_body or "encode('idna')" in suggest_body, (
+        "suggest_defaults must route non-ASCII labels through IDN/"
+        "punycode so the rendered slug matches the ASCII-only entity-"
+        "name regex the create endpoint enforces."
+    )
+
+    # The sanitiser must use an ASCII-only predicate, not Unicode-aware
+    # `isalnum()` alone (which is the pre-fix bug).
+    assert "isascii()" in suggest_body, (
+        "Sanitisation step must explicitly bound to ASCII (via "
+        "`c.isascii()`) so Unicode alphabetics get mapped to '-' rather "
+        "than retained verbatim."
+    )
+
+    # The first occurrence of the Unicode-only `c.isalnum()` (without an
+    # adjacent isascii() guard) must no longer exist in the slug-build
+    # path. We sanity-check by counting the matches of the bare
+    # `c.isalnum()` token within the function body — pre-fix there was
+    # exactly one such occurrence in the slug builder.
+    bare = re.findall(r"\bc\.isalnum\(\)\b", suggest_body)
+    assert all(
+        # Each occurrence must be paired with an `isascii()` guard on
+        # the same line — confirm by inspecting the line context.
+        any("isascii()" in ln for ln in suggest_body.splitlines() if "c.isalnum()" in ln)
+        for _ in bare
+    ), (
+        "Every `c.isalnum()` predicate in the slug builder must be "
+        "AND'd with `c.isascii()` so non-ASCII letters cannot leak "
+        "into the suggested entity names."
+    )
+
+
+def test_bulgu93_suggest_function_behavioural_smoke():
+    """Behavioural check: the same logic, invoked directly, should
+    produce an ASCII-only slug for a Unicode input."""
+    # Replicate the post-fix slug builder inline so the test does not
+    # require an event loop / FastAPI fixture. The point is to confirm
+    # the algorithmic shape matches what the source check above
+    # enforces.
+    def _build_slug(domain: str) -> str:
+        first_label = (
+            domain.replace("*.", "").split(".")[0]
+            if domain.replace("*.", "")
+            else ""
+        )
+        ascii_label = first_label
+        if first_label and not first_label.isascii():
+            try:
+                ascii_label = first_label.encode("idna").decode("ascii")
+            except (UnicodeError, UnicodeDecodeError):
+                ascii_label = "".join(
+                    c if c.isascii() and (c.isalnum() or c in ("-", "_"))
+                    else "-"
+                    for c in first_label
+                )
+        slug = (ascii_label or "newhost")[:32].lower()
+        slug = "".join(
+            c if c.isascii() and (c.isalnum() or c in ("-", "_"))
+            else "-"
+            for c in slug
+        )
+        if slug.startswith("_"):
+            slug = "h-" + slug.lstrip("_")
+        if not slug or not slug[0].isalpha():
+            slug = "h-" + slug
+        return slug
+
+    # Unicode IDN — should round-trip through punycode.
+    s = _build_slug("bücher.example.com")
+    # punycode of 'bücher' is 'xn--bcher-kva'
+    assert s == "xn--bcher-kva", f"expected xn--bcher-kva got {s!r}"
+    # The slug must satisfy the entity-name regex.
+    import re
+    assert re.fullmatch(r"^[a-zA-Z][a-zA-Z0-9_-]{0,63}$", s), s
+
+    # Plain ASCII domain still produces the obvious slug.
+    assert _build_slug("example.com") == "example"
+
+    # Uppercase normalises to lowercase.
+    assert _build_slug("EXAMPLE.COM") == "example"
+
+    # Trailing dot (FQDN absolute) stripped.
+    assert _build_slug("example.com.") == "example"
+
+    # Empty / dots-only input falls back to 'newhost'.
+    assert _build_slug("") == "newhost"
+    assert _build_slug("...") == "newhost"
diff --git a/frontend/package.json b/frontend/package.json
index 73c1fa5..0bf7693 100644
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -1,6 +1,6 @@
 {
   "name": "haproxy-openmanager-frontend",
-  "version": "1.5.0",
+  "version": "1.5.1",
   "description": "HAProxy Load Balancer Management UI",
   "dependencies": {
     "react": "^18.2.0",
diff --git a/frontend/src/components/FrontendManagement.js b/frontend/src/components/FrontendManagement.js
index e0c5e2a..5d86906 100644
--- a/frontend/src/components/FrontendManagement.js
+++ b/frontend/src/components/FrontendManagement.js
@@ -1015,12 +1015,56 @@ const FrontendManagement = () => {
       return;
     }
     if (grandfatheredContradictions.length > 0) {
+      // Bulgu #83 (round-23 audit) — surface the actual offending rule
+      // string(s) instead of just a count. Pre-fix the warning said
+      // "1 legacy rule has X !X" and the operator had to hunt
+      // through the ACL Builder cards to figure out which rule the
+      // gate was complaining about. The unchanged-rule path is the
+      // common case (operator changes port / maxconn on a frontend
+      // that already had a self-contradictory routing rule from a
+      // prior session), so making the rule discoverable from the
+      // toast keeps "Edit and Save" → "fix the dead rule" workflows
+      // single-screen. Also stop calling these rules "legacy" —
+      // the operator may have written them seconds earlier; the
+      // only thing this branch knows is that they weren't modified
+      // by the current edit.
+      const renderGrandfatheredRule = (r) => {
+        if (typeof r === 'string') return r;
+        if (r && typeof r === 'object') {
+          try { return JSON.stringify(r); } catch (_e) { return '[rule]'; }
+        }
+        return '[rule]';
+      };
+      const ruleSnippets = grandfatheredContradictions
+        .slice(0, 5)
+        .map(renderGrandfatheredRule)
+        .map((s) => (s.length > 160 ? `${s.slice(0, 157)}...` : s));
+      const extra = grandfatheredContradictions.length > ruleSnippets.length
+        ? ` (+${grandfatheredContradictions.length - ruleSnippets.length} more)`
+        : '';
       message.warning(
-        `This frontend has ${grandfatheredContradictions.length} legacy ` +
-        `routing/redirect rule(s) with a self-contradictory \`X !X\` ` +
-        `condition. The rule(s) never fire — fix them at your convenience. ` +
-        `Your current edit will still be saved.`,
-        6,
+        <div>
+          <div>
+            <strong>
+              {grandfatheredContradictions.length} routing/redirect rule(s)
+              you didn't modify in this edit contain a self-contradictory
+              `X !X` condition (e.g. `if acl1 !acl1`).
+            </strong>
+          </div>
+          <div style={{ marginTop: 4, fontSize: '12px' }}>
+            HAProxy accepts the syntax but `X AND NOT X` is always false,
+            so the rule never fires and traffic silently falls through to
+            `default_backend`. Your current edit will still be saved; fix
+            the rule(s) at your convenience.
+          </div>
+          <div style={{ marginTop: 6, fontSize: '12px', fontFamily: 'monospace' }}>
+            {ruleSnippets.map((s, i) => (
+              <div key={i}>• {s}</div>
+            ))}
+            {extra && <div>{extra}</div>}
+          </div>
+        </div>,
+        10,
       );
     }
 
@@ -1081,6 +1125,36 @@ const FrontendManagement = () => {
         } else {
           message.success('Frontend updated successfully');
         }
+
+        // Bulgu #83 (round-23 audit) — surface server-emitted
+        // grandfathered-rule warnings (e.g. `X !X` contradictions
+        // in routing/redirect rules that the operator did not
+        // touch this edit). The FE client-side gate ALSO catches
+        // these and fires its own toast above the modal close;
+        // we re-surface the server view here as a safety net in
+        // case the client gate missed an edge shape (different
+        // dict serialization, etc.). Server warnings already
+        // include the verbatim rule text, so the operator sees
+        // exactly which entry to fix.
+        const serverWarnings = Array.isArray(response.data?.warnings)
+          ? response.data.warnings
+          : [];
+        if (serverWarnings.length > 0 && grandfatheredContradictions.length === 0) {
+          message.warning(
+            <div>
+              <div><strong>Frontend saved, but the server flagged {serverWarnings.length} rule warning(s):</strong></div>
+              <div style={{ marginTop: 6, fontSize: '12px', fontFamily: 'monospace' }}>
+                {serverWarnings.slice(0, 5).map((w, i) => (
+                  <div key={i}>• {w.length > 240 ? `${w.slice(0, 237)}...` : w}</div>
+                ))}
+                {serverWarnings.length > 5 && (
+                  <div>(+{serverWarnings.length - 5} more)</div>
+                )}
+              </div>
+            </div>,
+            10,
+          );
+        }
       } else {
         response = await axios.post('/api/frontends', requestData);
         
diff --git a/version.json b/version.json
index 010fce6..b365943 100644
--- a/version.json
+++ b/version.json
@@ -1,5 +1,5 @@
 {
-  "version": "1.5.0",
-  "releaseName": "ACME Diagnostics & Proxied Host Wizard",
-  "releaseDate": "2026-05-08"
+  "version": "1.5.1",
+  "releaseName": "Round-23 + Round-24 audit follow-ups",
+  "releaseDate": "2026-05-13"
 }