Re-encode carefully after any unencodable code point

Apple's iconv converts several code points at once and, when one of them
is unencodable, restores the shift state of the whole batch, while the
output written for the preceding code points keeps it. Re-running the
conversion one code point at a time keeps the descriptor in sync with
the output, so the shift state can be reset before the replacement.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: serhiy-storchaka

Objects/unicodeobject.c

@@ -8526,25 +8526,15 @@ _PyUnicode_EncodeIconv(const char *encoding, PyObject *unicode,
             }
             /* A positive result counts nonreversible conversions: iconv()
                substituted an unencodable character instead of failing with
-               EILSEQ (musl and *BSD citrus do this).  Treat it as unencodable
-               and re-run one code point at a time to locate it. */
+               EILSEQ (musl and *BSD citrus do this). */
             if (ret > 0) {
-                if (!careful) {
-                    careful = 1;
-                    probe_cd = iconv_open_or_set_error(encoding, source,
-                                                       encoding);
-                    if (probe_cd == (iconv_t)-1) {
-                        goto done;
-                    }
-                    iconv(cd, NULL, NULL, NULL, NULL);
-                    out = PyBytesWriter_GetData(writer);
-                    outend = out + PyBytesWriter_GetSize(writer);
-                    up = ustart;
-                    continue;
+                if (careful) {
+                    /* The probe reported the code point as encodable, but it
+                       was substituted in the current shift state; drop it and
+                       report it. */
+                    out = out_before;
+                    up -= unit;
                 }
-                /* This code point was substituted; drop it and report it. */
-                out = out_before;
-                up -= unit;
             }
             else if (careful && up < uend) {
                 continue;
@@ -8566,6 +8556,23 @@ _PyUnicode_EncodeIconv(const char *encoding, PyObject *unicode,
             goto done;
         }
 
+        if (!careful) {
+            /* iconv() can reset the shift state of cd on an unencodable code
+               point, while the output written for the preceding code points
+               stays in the shifted state.  Re-run one code point at a time:
+               then nothing is written for the failed one. */
+            careful = 1;
+            probe_cd = iconv_open_or_set_error(encoding, source, encoding);
+            if (probe_cd == (iconv_t)-1) {
+                goto done;
+            }
+            iconv(cd, NULL, NULL, NULL, NULL);
+            out = PyBytesWriter_GetData(writer);
+            outend = out + PyBytesWriter_GetSize(writer);
+            up = ustart;
+            continue;
+        }
+
         /* An unencodable code point at *up; one input unit is one code point. */
         Py_ssize_t pos = (up - ustart) / unit;
         Py_ssize_t newpos;