Fix katakana lookup: normalize halfwidth forms to fullwidth

__test__/main/core/entries.test.ts

@@ -472,4 +472,17 @@ test("Test Japanese words", () => {
   expect(createLookupWordsJa("走った")).toEqual(expect.arrayContaining(["走る"]));
   expect(createLookupWordsJa("おいた")).toEqual(expect.arrayContaining(["おく", "おいる"]));
   expect(createLookupWordsJa("19az")).toEqual(expect.arrayContaining(["１９ａｚ"]));
+  // Halfwidth katakana must also reach fullwidth headwords (manga/UI text),
+  // keeping the usual prefix decomposition on both spellings.
+  expect(createLookupWordsJa("ﾃﾚﾋﾞ")).toEqual(expect.arrayContaining(["ﾃﾚﾋﾞ", "テレビ", "テレ", "テ", "ﾃﾚﾋ", "ﾃﾚ", "ﾃ"]));
+  expect(createLookupWordsJa("ｶﾞｽ")).toEqual(expect.arrayContaining(["ｶﾞｽ", "ガス", "ｶﾞ", "ｶ", "ガ"]));
+  // Conversion is scoped to halfwidth katakana: other characters must not get
+  // compatibility-folded (NFKC over the whole string would turn ① into "1").
+  const mixed = createLookupWordsJa("①ﾃﾚﾋﾞ");
+  expect(mixed).toEqual(expect.arrayContaining(["①ﾃﾚﾋﾞ", "①テレビ", "①テレ", "①ﾃﾚ"]));
+  expect(mixed).not.toContain("1テレビ");
+  // Fullwidth katakana runs keep cutting into all prefixes (show-all-candidates design)
+  expect(createLookupWordsJa("ソフトウェアエンジニア")).toEqual(
+    expect.arrayContaining(["ソフトウェアエンジニア", "ソフトウェア", "ソフト", "ソ"]),
+  );
 });

src/main/core/entry/ja.ts

@@ -10,6 +10,17 @@ import rule from "../rule";
 const RE_ALPHABETS_NUMBERS = /[A-Za-z0-9]/g;
 const FULLWIDTH_OFFSET = 0xfee0;
 
+// Halfwidth katakana (U+FF61-FF9F, common in manga/UI text) is converted to
+// its fullwidth form so ﾃﾚﾋﾞ looks up the same headwords as テレビ.
+// Only maximal halfwidth-katakana runs are normalized: NFKC over the whole
+// string would fold unrelated characters (e.g. ① -> 1). Normalizing each run
+// lets NFKC compose voiced marks with the preceding letter (ｶ + ﾞ -> ガ)
+// without touching anything outside the run.
+const RE_HALFWIDTH_KATAKANA_RUN = /[\uFF61-\uFF9F]+/g;
+
+const convertHalfwidthKatakana = (s: string): string =>
+  s.replace(RE_HALFWIDTH_KATAKANA_RUN, (run) => run.normalize("NFKC"));
+
 const createLookupWordsJa = (sourceStr: string): string[] => {
   const str = sourceStr
     .substring(0, 40)
@@ -20,13 +31,25 @@ const createLookupWordsJa = (sourceStr: string): string[] => {
 
   result.push(sourceStr); // Add the original word
 
-  for (let i = str.length; i >= 1; i--) {
-    const part = str.substring(0, i);
-    result.push(part);
+  // For halfwidth input, keep both chains: every candidate retains its usual
+  // prefix decomposition (show-all-plausible-candidates design), e.g.
+  // ﾃﾚﾋ -> ﾃﾚﾋ/ﾃﾚ/ﾃ plus テレビ/テレ/テ. Without halfwidth katakana the
+  // replace is a no-op, so no second chain is added.
+  const chains = [str];
+  const converted = convertHalfwidthKatakana(str);
+  if (converted !== str) {
+    chains.push(converted);
+  }
+
+  for (const chain of chains) {
+    for (let i = chain.length; i >= 1; i--) {
+      const part = chain.substring(0, i);
+      result.push(part);
 
-    if (i >= 2) {
-      const deinedWords = rule.doJa(part);
-      result.merge(deinedWords ?? []);
+      if (i >= 2) {
+        const deinedWords = rule.doJa(part);
+        result.merge(deinedWords ?? []);
+      }
     }
   }
   return result.toArray();