Fix a ReadConsoleOutputCharacter regression

src/host/ft_host/CJK_DbcsTests.cpp

@@ -164,6 +164,8 @@
     BEGIN_TEST_METHOD(TestInvalidTrailer)
         TEST_METHOD_PROPERTY(L"IsolationLevel", L"Method")
     END_TEST_METHOD()
+
+    TEST_METHOD(TestNarrowSurrogate);
 };
 
 bool DbcsTests::DbcsTestSetup()
@@ -2183,3 +2185,18 @@
 
     DbcsWriteRead::Verify(expected, output);
 }
+
+// The various console APIs that read back from the buffer are generally incompatible with UTF16 and surrogate pairs.
+// ReadConsoleOutputCharacterW in particular has a nLength parameter which is a column count but also the buffer size.
+// This makes it impossible to reliably return arbitrarily long graphemes per-cell in the output buffer.
+// The test ensures that we replace them with U+FFFD which makes the behavior more consistent for the caller.
+void DbcsTests::TestNarrowSurrogate()
+{
+    const auto out = GetStdHandle(STD_OUTPUT_HANDLE);
+    wchar_t buf[3];
+    DWORD read;
+
+    VERIFY_WIN32_BOOL_SUCCEEDED(WriteConsoleOutputCharacterW(out, L"a\U00010000b", 4, {}, &read));
+    VERIFY_WIN32_BOOL_SUCCEEDED(ReadConsoleOutputCharacterW(out, &buf[0], ARRAYSIZE(buf), {}, &read));
+    VERIFY_ARE_EQUAL(std::wstring_view(L"a\U0000FFFDb"), std::wstring_view(&buf[0], read));
+}

src/host/output.cpp

@@ -222,7 +222,12 @@ std::wstring ReadOutputStringW(const SCREEN_INFORMATION& screenInfo,
             // Otherwise, add anything that isn't a trailing cell. (Trailings are duplicate copies of the leading.)
             if (it->DbcsAttr() != DbcsAttribute::Trailing)
             {
-                retVal += it->Chars();
+                auto chars = it->Chars();
+                if (chars.size() > 1)
+                {
+                    chars = { &UNICODE_REPLACEMENT, 1 };
+                }
+                retVal += chars;
             }
         }