Record site-name provenance instead of inferring it from the name's shape

Recipe extraction no longer falls back to the URL's host: a stored
recipes.site_name is now always a name the page gave (publisher or
og:site_name), and null when it named none. dinners.SourceOf is the one
documented rule for every fetched recipe's source: the page's name is
shown as it stands; otherwise the site is named by the link's host only
when the link itself is shown as free of secrets, else null. The recipe
detail, idea results and the recipe-candidate offer endpoint all apply
it; candidates no longer carry an extraction-time hostname fallback, so
their site_name is nullable (frontend updated to match).

Migration 0020 clears host-shaped site names (dotted host names, IP
addresses) stored before the change, in recipes and in legacy idea job
results, so already-stored rows read under the same rule.

Maestro 966066e3-0e25-4b8b-9bb0-0694c47c9106

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

internal/recipes/extract_test.go

@@ -251,10 +251,10 @@ func TestExtractShapes(t *testing.T) {
 			},
 		},
 		{
-			name: "site name from host without publisher or og:site_name",
+			name: "no site name without publisher or og:site_name, never the host",
 			html: page("", `{"@type":"Recipe","name":"Stew"}`),
 			check: func(t *testing.T, r Recipe) {
-				if r.SiteName != "example.com" {
+				if r.SiteName != "" {
 					t.Errorf("SiteName = %q", r.SiteName)
 				}
 			},