M39: harden application and acceptance gates

backend/app/core/config.py

@@ -64,12 +64,52 @@ class Settings(BaseSettings):
     oidc_auto_provision: bool = True
     oidc_default_role: str = "rental_employee"
     log_level: str = "INFO"
+    # Failed password logins per client IP before a temporary 429 (0 disables).
+    login_max_failures: int = 10
+    login_failure_window_seconds: int = 900
     metrics_bearer_token: str = ""
     privacy_minimum_booking_retention_days: int = 30
     privacy_audit_retention_days: int = 2555
     privacy_audit_export_max_rows: int = 10000
 
 
+# Secrets that guard *inbound* trust (session cookies, service callbacks). Running
+# production with any of these at their placeholder value means forged sessions or
+# unauthenticated writes, so startup refuses.
+INSECURE_DEFAULT_SECRETS: tuple[tuple[str, str], ...] = (
+    ("app_secret", "replace-in-production"),
+    ("n8n_callback_token", "replace-me-n8n-callback-token"),
+    ("mcp_hub_service_token", "replace-me-mcp-hub-token"),
+)
+
+
+def insecure_default_secrets(settings: "Settings") -> list[str]:
+    """Return the names of secret settings that still carry their placeholder value.
+
+    Only secrets that actually guard something in the given deployment are reported:
+    ``mcp_hub_service_token`` is irrelevant while MCP Hub registration is disabled.
+    """
+    insecure: list[str] = []
+    for name, placeholder in INSECURE_DEFAULT_SECRETS:
+        if name == "mcp_hub_service_token" and not settings.mcp_hub_registration_enabled:
+            continue
+        value = getattr(settings, name)
+        if not value or value == placeholder or value.startswith("replace-me"):
+            insecure.append(name)
+    return insecure
+
+
 @lru_cache
 def get_settings() -> Settings:
-    return Settings()
+    settings = Settings()
+    if settings.mobilityops_env.lower() == "production":
+        insecure = insecure_default_secrets(settings)
+        if insecure:
+            # Refuse to boot rather than run production with forgeable session cookies
+            # or guessable service tokens. Development/test/demo keep the defaults.
+            raise RuntimeError(
+                "Refusing to start in production with placeholder secrets: "
+                + ", ".join(insecure)
+                + ". Set real values in the environment (see .env.example)."
+            )
+    return settings

backend/app/core/observability.py

@@ -74,10 +74,19 @@ def correlation_id_for(request: Request) -> str:
         return str(uuid.uuid4())
 
 
+UNMATCHED_ROUTE_LABEL = "<unmatched>"
+
+
 def route_label(request: Request) -> str:
+    """Return the route *template* for metrics labels.
+
+    Unmatched paths (404 probes, scanners) must not become their own label value:
+    every distinct URL would otherwise create a new Prometheus time series and the
+    metric cardinality would grow without bound.
+    """
     route = request.scope.get("route")
     path = getattr(route, "path", None)
-    return str(path or request.url.path)
+    return str(path) if path else UNMATCHED_ROUTE_LABEL
 
 
 def request_started() -> float:

backend/app/core/ratelimit.py

@@ -0,0 +1,49 @@
+"""Small in-process failed-attempt limiter for credential endpoints.
+
+Fleet Ops runs as a single API process per deployment, so an in-memory sliding window
+is sufficient to blunt online password guessing (and the scrypt CPU amplification that
+comes with it) without adding Redis. Only *failed* attempts count, so legitimate users
+and the automated test suite are never throttled.
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from collections import deque
+
+
+class FailedAttemptLimiter:
+    def __init__(self, *, max_failures: int, window_seconds: float) -> None:
+        self.max_failures = max_failures
+        self.window_seconds = window_seconds
+        self._failures: dict[str, deque[float]] = {}
+        self._lock = threading.Lock()
+
+    def _prune(self, key: str, now: float) -> deque[float]:
+        bucket = self._failures.setdefault(key, deque())
+        cutoff = now - self.window_seconds
+        while bucket and bucket[0] <= cutoff:
+            bucket.popleft()
+        if not bucket:
+            self._failures.pop(key, None)
+        return bucket
+
+    def retry_after_seconds(self, key: str) -> int:
+        """Return >0 seconds to wait when the key is currently blocked, else 0."""
+        now = time.monotonic()
+        with self._lock:
+            bucket = self._prune(key, now)
+            if len(bucket) < self.max_failures:
+                return 0
+            return max(1, int(bucket[0] + self.window_seconds - now + 0.999))
+
+    def record_failure(self, key: str) -> None:
+        now = time.monotonic()
+        with self._lock:
+            self._prune(key, now)
+            self._failures.setdefault(key, deque()).append(now)
+
+    def reset(self, key: str) -> None:
+        with self._lock:
+            self._failures.pop(key, None)