feat(tier2): smarter fraud detection — Block G

Per-event fraud tuning. Hosts can now dial the medium / high / block
boundaries, allowlist trusted networks, and feed verdicts back on
flagged accesses — the seed corpus for a future ML model.

Schema (migration 0011)
- events.fraud_{medium,high,block}_threshold default 30/60/85 so
  existing events behave identically until a host changes them
- access_logs.geo_{country,city,lat,lon} for future enrichment
- fraud_feedback table — verdict ('legitimate' | 'suspicious') + note,
  PK on access_log_id so re-mark is an upsert
- event_allowlists table — (event_id, ip_cidr) primary key, inet column
  so containment checks use the native >>= operator (indexed lookup)

Domain
- FraudThresholds with Valid() + Band() helpers; Default trio echoed
  through GET responses so the frontend doesn't duplicate constants
- ParseAllowlistCIDR accepts bare IPs (auto-widens to /32 or /128) and
  canonicalises the output (203.0.113.42 → 203.0.113.42/32)
- Event.Thresholds() falls back to defaults if columns weren't
  populated yet, so the API never wedges every score into "low"

Storage
- AllowlistRepo: List / Add / Remove + Matches() — the latter pushes
  CIDR containment into Postgres rather than streaming rows back
- FeedbackRepo: Record (upserts) + ListForEvent (joined through guests)
- EventRepo.GetThresholds + UpdateThresholds, plus the threshold
  columns baked into scanEvent so every event load carries them
- AccessLogRepo.BelongsToEvent — stops a hostile editor on event A
  from marking event B's access logs

API
- GET/PUT /events/{id}/security/thresholds (viewer/editor)
- GET/POST/DELETE /events/{id}/security/allowlist
- POST /events/{id}/access-logs/{log_id}/feedback (editor)
- GET /events/{id}/security/feedback
- RSVP scoring path: allowlist short-circuit fires before the fraud
  engine; the engine's score is then re-banded against the event's
  thresholds (engine.Risk becomes advisory — API is the source of
  truth for "what counts as block here")
- CORS Allow-Methods already includes PUT (Block D fix)

Fraud engine
- Single-signal cap: it now takes ≥2 sub-scores of ≥70 to push the
  final into HIGH. Fixes the well-known "second visit with a slightly
  shifted fingerprint scores 60+" false positive
- Engine band remains advisory; API re-bands using per-event
  thresholds before deciding to block

Frontend
- SecurityCard.vue: visual band ribbon (proportional to thresholds),
  three sliders with mutual clamping so dragging medium past high
  pushes high (not an invalid ordering), reset-to-defaults button,
  CIDR allowlist with inline add + per-row remove, verdict-history
  inbox. Toast feedback on save/add/remove
- "Security" tab added to the event-detail tab nav (5th tab,
  right of Analytics)
- Viewer role hides write affordances; server enforces too

Tests
- Domain: ThresholdsBand, ThresholdsValid, ParseAllowlistCIDR (bare
  IP widening + traversal/typo rejection), FraudFeedbackValid
- Integration: thresholds round-trip + invalid ordering rejection,
  allowlist CRUD + duplicate 409 + invalid CIDR 400 + IP auto-widen,
  feedback record + upsert + cross-tenant 404 + invalid verdict 400,
  viewer can read / editor can write / outsider gets 404
- Full integration suite green (315.8s, all 36 top-level tests pass)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

internal/api/rsvps.go

@@ -33,6 +33,7 @@ type rsvpHandler struct {
 	events     *storage.EventRepo
 	rsvps      *storage.RSVPRepo
 	accessLogs *storage.AccessLogRepo
+	allowlist  *storage.AllowlistRepo
 	scorer     fraudScorer
 	pub        rsvpPublisher
 }
@@ -322,6 +323,26 @@ func (h *rsvpHandler) scoreAccess(
 		h.logger.Error("create access log", "err", err)
 	}
 
+	// Block G: allowlist short-circuit. If the request IP matches a CIDR
+	// the host has explicitly trusted (office Wi-Fi, family network), we
+	// skip the fraud engine entirely — score 0, low band. Best-effort:
+	// any error reading the allowlist falls through to normal scoring so a
+	// dropped DB connection doesn't lock guests out of an event.
+	if h.allowlist != nil {
+		if matched, label, err := h.allowlist.Matches(r.Context(), event.ID, ip); err == nil && matched {
+			reason := "allowlisted"
+			if label != "" {
+				reason = "allowlisted: " + label
+			}
+			return fraud.Decision{
+				Score:   0,
+				Risk:    "low",
+				Reasons: []string{reason},
+				Used:    true,
+			}, fingerprint, ip, true
+		}
+	}
+
 	decision := h.scorer.Score(r.Context(), fraud.ScoreInput{
 		EventID:     event.ID,
 		GuestID:     guest.ID,
@@ -332,6 +353,14 @@ func (h *rsvpHandler) scoreAccess(
 		UserAgent:   r.UserAgent(),
 		Referrer:    r.Referer(),
 	})
+
+	// Block G: re-band the score using this event's thresholds. The
+	// engine's `Risk` field becomes advisory; the API is the source of
+	// truth for "what counts as block here". This lets a strict-event
+	// host set Block=70 while a casual-event host sets it to 95 without
+	// touching the engine.
+	decision.Risk = event.Thresholds().Band(decision.Score)
+
 	if fraud.IsBlock(decision) {
 		writeJSON(w, http.StatusForbidden, submitRSVPResponse{
 			Decision: decision,