Prune ghost words from affix expansion and stop frequency loader creating words

- Add PruneOrphanWords post-import step that deletes words with no
  definitions, synonyms, translations, pronunciations, etymology,
  synsets, or frequency data
- Remove INSERT path from CETEMPublico loader — frequency data should
  annotate existing words, not create new ones (was inflating pt-PT
  from 136K real words to 718K)
- Add hasLetter filter to hunspell, affix, and dicionario loaders to
  block punctuation entries like ".", ",", "(" from entering the DB
- Move hasLetter helper to scowl.go alongside other shared filters

Impact: en 655K→129K, pt-PT 1.5M→136K, fr 80K→56K (on reimport)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

internal/loader/scowl.go

@@ -8,6 +8,7 @@ import (
 	"os"
 	"path/filepath"
 	"strings"
+	"unicode"
 )
 
 type SCOWLLoader struct{}
@@ -148,3 +149,12 @@ func containsNonLatin(s string) bool {
 	}
 	return false
 }
+
+func hasLetter(s string) bool {
+	for _, r := range s {
+		if unicode.IsLetter(r) {
+			return true
+		}
+	}
+	return false
+}