Windows: encode and decode text as UTF-8 instead of trusting the locale

setlocale(LC_ALL, "") maps to the ANSI code page on Windows -- 936 (GBK)
on a Chinese system -- so wcrtomb() produces GBK where the protocol wants
UTF-8, and mbrtowc() reads UTF-8 as GBK. Typed Chinese therefore reached
the model as mojibake, and once the buffer held correct UTF-8 the input
line rendered as mojibake instead. Windows' C runtime has no ".UTF-8"
locale to switch to, so the conversion has to be explicit.

os_compat.[ch]
  * add os_utf8_encode() and os_utf8_decode(), locale-independent, with
    range and overlong-form checks
src/tui.c
  * input_put_wc(), the command-palette filter and the ask_user buffer use
    os_utf8_encode() instead of wcrtomb()
src/textsel.c
  * cst_utf8_decode() uses os_utf8_decode() instead of mbrtowc()

src/textsel.c

@@ -3,23 +3,29 @@
 #endif
 
 #include "cst_textsel.h"
+#include "os_compat.h"
 
 #include <stdlib.h>
 #include <string.h>
 #include <wctype.h>
 
 int cst_utf8_decode(const char *s, size_t avail, wchar_t *wc, size_t *eaten) {
+  unsigned long cp = 0;
+  int n;
   if (avail == 0)
     return -1;
-  mbstate_t st;
-  memset(&st, 0, sizeof(st));
-  size_t n = mbrtowc(wc, s, avail, &st);
-  if (n == (size_t)-1 || n == (size_t)-2 || n == 0) {
-    *wc = (unsigned char)s[0];
+  /* os_utf8_decode(), not mbrtowc(): the latter follows setlocale(), which on
+   * Windows means the ANSI code page, so UTF-8 text would be decoded as GBK
+   * and rendered as mojibake. */
+  n = os_utf8_decode(s, avail, &cp);
+  if (n <= 0) {
+    /* Not a valid sequence: fall back to one raw byte, as before. */
+    *wc = (wchar_t)(unsigned char)s[0];
     *eaten = 1;
     return 0;
   }
-  *eaten = n;
+  *wc = (wchar_t)cp;
+  *eaten = (size_t)n;
   return 0;
 }