recommend: prewarm requests don't bump public toolkit count

The prewarm header was added to bypass per-IP rate limits + optional
force-refresh; it did NOT skip recordGeneration, so every prewarm call
bumped the public /api/toolkit-count counter. That's fine in principle
(the request was really served) but visually wrong: a 100-query
prewarm run would jump the homepage display from '400 toolkits built'
to '500' in a way that reads as inorganic.

Wrap each of the four recordGeneration call sites (L1 / L2 / legacy /
L3) with 'if !prewarm'. A valid prewarm token now warms the caches
without leaving a trace in toolkit_generations. The per-request log
line ('[recommend] L1 quick-match hit' etc.) still fires — ops
visibility preserved.

Scoped intentionally: the prewarm flag comes from a valid token + the
X-Stockyard-Prewarm header, so an unauthenticated scraper can't set
this header and have it quietly exempt them from the counter. Token
lives in STOCKYARD_PREWARM_TOKEN on Railway.

Verification:
  - go build ./internal/site/...: clean
  - go vet ./internal/site/...: clean
  - Existing site tests: pass (no recommend-flow tests exist today;
    the prewarm path is covered by the integration-style /api/recommend
    call we'll make from the prewarm script post-deploy)
  - Secret scan: 0 hits

Author: claude

internal/site/recommend.go

@@ -411,7 +411,9 @@ func (r *Recommender) HandleRecommend(w http.ResponseWriter, req *http.Request)
 				result := personalize(cached, businessName)
 				result.Slug = slug
 				result.Cached = true
-				r.recordGeneration(slug, desc, "L1", req.RemoteAddr, len(result.Tools))
+				if !prewarm {
+					r.recordGeneration(slug, desc, "L1", req.RemoteAddr, len(result.Tools))
+				}
 				log.Printf("[recommend] L1 quick-match hit: %q → %s", normalized, slug)
 				w.Header().Set("Content-Type", "application/json")
 				json.NewEncoder(w).Encode(result)
@@ -428,7 +430,9 @@ func (r *Recommender) HandleRecommend(w http.ResponseWriter, req *http.Request)
 		if cached, ok := r.recCache.Get(normalized); ok {
 			result := personalize(cached, businessName)
 			result.Cached = true
-			r.recordGeneration(result.Slug, desc, "L2", req.RemoteAddr, len(result.Tools))
+			if !prewarm {
+				r.recordGeneration(result.Slug, desc, "L2", req.RemoteAddr, len(result.Tools))
+			}
 			log.Printf("[recommend] L2 cache hit: %q", normalized)
 			w.Header().Set("Content-Type", "application/json")
 			json.NewEncoder(w).Encode(result)
@@ -450,7 +454,9 @@ func (r *Recommender) HandleRecommend(w http.ResponseWriter, req *http.Request)
 			// Migrate to new cache for future hits
 			r.recCache.Set(normalized, oldSlug, &result)
 			result2 := personalize(&result, businessName)
-			r.recordGeneration(oldSlug, desc, "legacy", req.RemoteAddr, len(result2.Tools))
+			if !prewarm {
+				r.recordGeneration(oldSlug, desc, "legacy", req.RemoteAddr, len(result2.Tools))
+			}
 			log.Printf("[recommend] legacy cache hit (migrated): %q → %s", normalized, oldSlug)
 			w.Header().Set("Content-Type", "application/json")
 			json.NewEncoder(w).Encode(result2)
@@ -610,7 +616,9 @@ func (r *Recommender) HandleRecommend(w http.ResponseWriter, req *http.Request)
 		genSlug, desc, string(resultJSON))
 
 	log.Printf("[recommend] L3 %s: %q → %s (%d tools)", modelUsed, normalized, genSlug, len(result.Tools))
-	r.recordGeneration(genSlug, desc, "L3", req.RemoteAddr, len(result.Tools))
+	if !prewarm {
+		r.recordGeneration(genSlug, desc, "L3", req.RemoteAddr, len(result.Tools))
+	}
 
 	finalResult := personalize(result, businessName)
 	w.Header().Set("Content-Type", "application/json")