fix(api): bound what a forged CF-Connecting-IP can buy

Code review, both findings valid.

The design doc claimed Cloudflare "replaces the header, so a browser
cannot forge it", and that only the X-Forwarded-For fallback was
forgeable. That is true only for traffic that actually passed through
Cloudflare, and nothing in this process can verify that it did. Reaching
the origin directly, both headers are equally attacker-controlled — and
rotating CF-Connecting-IP mints a fresh rate-limit bucket per request,
defeating per-client limits on every endpoint including the
DataFrame-heavy /api/schools. Against abuse that is worse than the
shared bucket it replaced, which at least capped everyone together.

So the ceiling comes back. I dropped it earlier arguing it belonged at
Cloudflare; that argument assumed the keying was sound, and it is not.
GlobalRateLimitMiddleware counts all /api/ traffic in a fixed window
against a total, independent of client identity, outermost so it refuses
before any work happens. Written by hand because slowapi cannot express
a global cap: default_limits and application_limits are both keyed by
key_func, and the latter needs middleware this app does not install.

It does not make the header trustworthy — it makes trusting it
survivable. The real fix is Authenticated Origin Pulls or an origin
firewall, now documented in DEPLOY.md as the open gap it is.

127.0.0.1 is exempt: the healthcheck curls localhost from inside the
container, and starving it would restart the container and turn a load
spike into an outage loop. Keyed on the peer address, never the Host
header, which the caller sets.

Second finding: suggest_schools_typesense promised "never raises" while
the parsing loop sat outside the try, so int(None) on a malformed
document would have made a keystroke a 500. The loop now skips bad rows
rather than dropping the whole list — and a hit with no document no
longer becomes a suggestion pointing at /school/0.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015mWQnpye9F299NVRCCSRvj

backend/app.py

@@ -6,6 +6,7 @@ Uses real data from UK Government Compare School Performance downloads.
 
 import hashlib
 import re
+import time
 from contextlib import asynccontextmanager
 from datetime import datetime, timezone
 from typing import Optional
@@ -15,7 +16,7 @@ import pandas as pd
 from fastapi import FastAPI, HTTPException, Query, Request, Depends, Header
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.middleware.gzip import GZipMiddleware
-from fastapi.responses import FileResponse, Response
+from fastapi.responses import FileResponse, JSONResponse, Response
 from fastapi.staticfiles import StaticFiles
 from slowapi import Limiter, _rate_limit_exceeded_handler
 from slowapi.util import get_remote_address
@@ -321,14 +322,79 @@ def client_key(request: Request) -> str:
     return get_remote_address(request)
 
 
-# Rate limiter. No in-app global ceiling: slowapi's default_limits and
-# application_limits are both keyed by key_func (so per-client, not global)
-# and the latter only applies with SlowAPIMiddleware installed, which this app
-# does not use. A global cap belongs at Cloudflare, which is already in the
-# path. See the spec's §1 for why that is deliberate.
+# Per-client limiter. Paired with the global ceiling below — the two do
+# different jobs and neither substitutes for the other.
 limiter = Limiter(key_func=client_key)
 
 
+# --- The ceiling no header can raise ----------------------------------------
+#
+# client_key trusts CF-Connecting-IP, and nothing in this process can tell an
+# edge-set header from an attacker-set one. That distinction can only be made
+# at Cloudflare, with Authenticated Origin Pulls or an origin firewall. A
+# caller reaching the origin directly could otherwise mint a fresh rate-limit
+# bucket per request and evade per-client limits entirely — which would make
+# correct keying a net regression against abuse, since the single shared bucket
+# it replaced at least capped everyone at 60/minute together.
+#
+# So per-client limits give fairness, and this gives the origin a hard total.
+# It does not make the header trustworthy; it bounds what trusting it can cost.
+# The header problem itself is closed at Cloudflare, not here.
+#
+# [window_start_monotonic, count], or None before the first request. A fixed
+# window is crude, which is right for a backstop: it has to be obviously
+# correct rather than fair.
+_global_window: Optional[list] = None
+
+# The container healthcheck runs `curl http://localhost:80/api/data-info` from
+# inside the container. Starving it would fail the check, restart the
+# container, and turn a load spike into an outage loop — the ceiling exists to
+# protect the origin, not to kill it.
+_LOCAL_HOSTS = frozenset({"127.0.0.1", "::1", "localhost"})
+
+
+def exempt_from_ceiling(request: Request) -> bool:
+    """Whether the ceiling should ignore this request.
+
+    Its own function so the rule is testable without standing up a server —
+    and so the healthcheck exemption is somewhere a reader can find it.
+    """
+    if not request.url.path.startswith("/api/"):
+        return True
+    # The peer address, never the Host header: Host is set by the caller and
+    # would hand every attacker an exemption.
+    return (request.client.host if request.client else "") in _LOCAL_HOSTS
+
+
+class GlobalRateLimitMiddleware(BaseHTTPMiddleware):
+    """A cap on total /api/ traffic, independent of any client identity."""
+
+    async def dispatch(self, request: Request, call_next):
+        global _global_window
+
+        if exempt_from_ceiling(request):
+            return await call_next(request)
+
+        now = time.monotonic()
+        # One event loop, and no await between the read and the write, so this
+        # sequence is atomic without a lock.
+        if _global_window is None or now - _global_window[0] >= 60:
+            _global_window = [now, 0]
+        _global_window[1] += 1
+
+        if _global_window[1] > settings.global_rate_limit_per_minute:
+            return JSONResponse(
+                # Distinguishable from slowapi's per-client 429: an operator
+                # reading logs has to be able to tell "one noisy client" from
+                # "the origin is saturated".
+                {"detail": "The service is at capacity. Please retry shortly."},
+                status_code=429,
+                headers={"Retry-After":
+                         str(max(1, int(60 - (now - _global_window[0]))))},
+            )
+        return await call_next(request)
+
+
 class SecurityHeadersMiddleware(BaseHTTPMiddleware):
     """Add security headers to all responses."""
 
@@ -532,6 +598,10 @@ app.add_middleware(CacheAndETagMiddleware)
 app.add_middleware(SecurityHeadersMiddleware)
 app.add_middleware(RequestSizeLimitMiddleware)
 app.add_middleware(GZipMiddleware, minimum_size=512)
+# Added last, so it is outermost and refuses before anything downstream does
+# work. A ceiling that only applies after the expensive part has run is not a
+# ceiling.
+app.add_middleware(GlobalRateLimitMiddleware)
 
 # CORS middleware - restricted for production
 app.add_middleware(