Just in case it might be of interest, if a motion on this proposal were raised in a UTC meeting, I suspect Microsoft would vote against encoding.
Peter From: Unicode <[email protected]> On Behalf Of [email protected] via Unicode Sent: May 6, 2026 11:09 AM To: Philippe Verdy via Unicode <[email protected]>; Philippe Verdy <[email protected]> Subject: Re: Odp: Pd: Missing legacy Arabic encoding The ReadConsoleOutputW function, by definition, captures the tiles into the lpBuffer, which is a random access array of CHAR_INFO structure, whose horizontal and vertical size is specified in dwBufferSize. Since lpBuffer is random access, this implies one CHAR_INFO structure per character tile. The API therefore fundamentally imposes a strict memory layout that cannot be violated. The fact that some Unix-like terminals such as Windows Terminal may support features outside the scope of the CHAR_INFO structure for compatibility with ANSI escape codes or WSL programs does not invalidate the compatibility considerations for legacy DOS/Win16/Win32 programs that require all character tiles to fit in the CHAR_INFO structure for random access, because 4 byte CHAR_INFO structure of Win32 is intended to be a fully backwards compatible extension of the 2 byte VGA text mode tile structure of DOS/Win16. Dnia 06 maja 2026 19:58 Philippe Verdy via Unicode<mailto:[email protected]> < [email protected]<mailto:[email protected]> > napisał(a): Windows can use other ways to map 16-bit codes in its *legacy* Console buffer (using old CHAR_INFO structure), it can perfectly internally use compatibility characters, or PUAs of the BMP, and still present an API that exposes connforming sequences. You're talking about an old implementation that was built even long before the Arabic script was extended (and newer scripts using contextual joining behaviors, that have never been part of the BMP, shcih as Adlam, and other scripts like Mongolian that also may need such sequences with ZWJ/ZWNJ controls, or with other formatting characters like those specific to Mongolian like FVS1...FVS4 and MVS, or those common to many Bhramic scripts, that the *legacy* Console did not support. The *legacy* console was not built to support more than one plane (including many CJK cgaracters). The newer console can! Le mer. 6 mai 2026 à 18:37, [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>> a écrit : Have you read the L2/26-077 proposal? Using ZWJ or ZWNJ would not work for the compatibility purposes at all as already explained in the proposal. This is because ZWJ or ZWNJ would take the space of one character tile in the CHAR_INFO structure. Suppose that you're trying to map 0xD0 from FP164 to a sequence of U+FE7C U+200D U+064B (ﹼً). The legacy application fills the 80×25 screen with all 0xD0 tiles. You subsequently try to capture the tiles with a Win32 program by using ReadConsoleOutputA into an 80×25 buffer of 2000 tiles. This succeeds and captures 0xD0 into all the tiles. You then try to capture the tiles using ReadConsoleOutputW into an 80×25 buffer. Each sequence U+FE7C U+200D U+064B would take a sequence of three CHAR_INFO structures to store, meaning 6000 such structures for the whole screen. But the 80×25 buffer has only room for 2000 instances of the structure (one per character tile). Since CHAR_INFO stores 16-bit character code, by that same logic the compatibility characters would have to be in BMP for it to work. In Windows 95 Vietnamese and Windows 95 Thai, there are instances where one character tile takes multiple CHAR_INFO structures, causing visual width to be smaller than logical width, and when that happens, the remaining space at the end of the line is left blank, allowing for CP1258/CP874 combining characters in those systems to map 1:1 to their Unicode equivalents. Windows 3.1/95/98/ME Arabic don't work that way and don't use combining characters or ZWJ sequences, so visual width is always equivalent to logical width, each character tile maps 1:1 to a CHAR_INFO structure and all characters may fill the entire line, which would be impossible if some of those characters were mapped to composition sequences or non-BMP characters. Since there is currently no sufficient evidence of user community that would need to use those mappings, there are no plans for those characters to be added to Unicode, and therefore the only solution for the ReadConsoleOutputW to work properly in this case is to use agreed upon private use mappings for those compatibility characters. Dnia 06 maja 2026 18:06 Philippe Verdy via Unicode<mailto:[email protected]> < [email protected]<mailto:[email protected]> > napisał(a): You actually don't need any new compatibility characters for Arabic contextual forms, or for other contextual forms in other joining scripts (like Adlam, or even Mongolian whichbis a LTR script). You just have to prepend or append a ZWJ or ZWNJ formatting control to the unified letter if you want to override its default contextual presentation form. Le mar. 5 mai 2026, 00:48, Asmus Freytag via Unicode <[email protected]<mailto:[email protected]>> a écrit : The issue at hand is the distinction between a theoretical gap and real-life problem. You have demonstrated that there are specifications that, if chained in the right way, can lead to ambiguities or gaps in interchange. What we don't have is an actual use case with real-life consequences for a set of existing users, not hypothetical ones. When it comes to encoding decisions based on existing documents, there is a strong presumption that once sufficiently many documents exist that contain a character, that this character will be needed in digitizing these documents, whether immediately, or eventually (e.g. in the case of future scholarly studies). Also, the texts themselves exist, barring accidents, in permanence. Therefore, it is justified to consider irrevocably allocating a character that will map to this source in perpetuity, even though each encoded character carries a small cost for implementers. However, when it comes to legacy characters, there's an additional cost that is imposed, and that is based on the fact that characters that are encoded solely for compatibility will usually violate one or more of the other encoding principles, something that incrementally complicates the standard. Even for people who never intend to use that character. Therefore, the SEW is on solid ground when it demands not only a hypothetical scenario, but evidence of actual impact on actual users. Not only whether some application could invoke an API, but whether such applications exist and are used today to access documents encoded using the legacy characters in a way that is compromised irreparably by not having an encoding for them. A./ On 5/4/2026 10:24 AM, [email protected]<mailto:[email protected]> via Unicode wrote: In UTC 187 Minutes, "Asmus Freytag noted that the fact that lists of things existed in the past does not make these things plain text. Ned Holbrook pointed out that the purported issue occurs in a closed system, not in public interchange.". However, the arguments in the proposal do not merely hinge on the encodings being lists of characters, but specifically points out methods to interchange text, including an example of copying terminal output and pasting to Notepad, where the copying invokes the mapping of the current terminal codepage to UCS-2 (as is CHAR_INFO compatible) and the pasting writes it into plain text. Win32 is also not a closed system, as Win32 can capture the tiles of the output of Windows 3.1 Arabic DOS/Win16 programs and Windows 95/98/ME Arabic DOS/Win16/Win32 programs, but Win32 can also interact with public text interchange systems by reading and writing to files and network. I'm not saying that Unicode absolutely must include those characters, but those kinds of misleading claims are causing users to misunderstand what the proposal is about, and I don't want Unicode to be relying on uninformed decisions to evaluate proposals. Dnia 18 kwietnia 2026 13:36 [email protected] via Unicode<mailto:[email protected]> < [email protected] ><mailto:[email protected]> napisał(a): The SEW subsequently explained that the actual reason is due to insufficient evidence of user community that would need to use the resulting mapping. Despite Win32 being a highly popular platform with plenty of backwards compatibility and native UCS-2 terminal support, the specific use cases of installing codepages into Windows NT and using terminal tiles from Windows 3.1/95/98/ME are not sufficiently documented, making it difficult for any user communities to form around it. So it seems like the idea of standardizing legacy Arabic terminal BMP mappings is a dead end for now. Dnia 17 kwietnia 2026 22:59 [email protected] via Unicode<mailto:[email protected]> < [email protected] ><mailto:[email protected]> napisał(a): The Recommendations in L2/26-100 claim that Microsoft's documentation of legacy Arabic encodings is available at https://learn.microsoft.com/en-us/typography/legacy/legacy_arabic_fonts. However, that article only demonstrates two encodings of TrueType fonts, which are used in Windows 3.1 but are completely different from the eight terminal encodings. Unlike the TrueType encodings which represent internal shaping mappings and are not used for text interchange, the terminal encodings have been demonstrated to be directly used in text interchange through int 10h and ReadConsoleOutputA/WriteConsoleOutputA as already demonstrated in L2/26-077. The Recommendations also claim that the proposal does not demonstrate any need for interchange or encoding, but the proposal actually demonstrated such a need due to the logical extension of the Win32 terminal API to the functions ReadConsoleOutputW/WriteConsoleOutputW, which are in Windows NT and may be used on the output of previously ran programs (including those that used the legacy Arabic terminal encodings), which given the CHAR_INFO structure, therefore implies a need for all the tiles to map to BMP for interchange. I'm not objecting to the SEW's conclusion of "Users are expected to use PUA.", which can indeed be used to provide a mapping even if not standardized, but the reasoning given was flawed. Dnia 09 stycznia 2026 17:25 [email protected]<mailto:[email protected]> < [email protected] ><mailto:[email protected]> napisał(a): The following Win32 C code will output 256 characters in system console codepage into the character grid, capture those character tiles in UCS-2 if possible, and then output the current console codepage number. #include <windows.h> #include <stdio.h> int main(){ HANDLE hConsole=GetStdHandle(STD_OUTPUT_HANDLE); CHAR_INFO screen[256]; COORD size={16,16,}; COORD pos={0,0,}; SMALL_RECT rect={0,0,15,15,}; for(int i=0;i<256;i++){ screen[i].Attributes=0xF0; screen[i].Char.AsciiChar=i; } WriteConsoleOutputA(hConsole,screen,size,pos,&rect); CHAR_INFO screenu[256]; if(ReadConsoleOutputW(hConsole,screenu,size,pos,&rect)){ for(int i=0;i<256;i++) printf("%04X ",screenu[i].Char.UnicodeChar); } else{ printf("error %08X\n",GetLastError()); } printf("codepage %u",GetConsoleOutputCP()); } In most cases, whenever a legacy Win32 codepage is used, the application can run on Windows NT to capture the UCS-2 mapping of those character cells to the BMP (although for CJK codepages a more complex setup would be necessary due to thousands of fullwidth characters with 2-byte sequences). However, in Arabic versions of Windows 9x (95/98/ME) the resulting character set has many presentation forms that are not in Unicode. This is the result when running on Windows ME: https://i.imgur.com/QFm3SkI.png in 10×20 font, https://i.imgur.com/KUbLQ0A.png in 10×18 font (same result also appears in Windows 95/98). 5×12, 7×12, 8×12, 10×18, 10×20, and 12×16 bitmap fonts have been attested with that character set (VGAOEM.FON, 8514OEM.FON, DOSAPP.FON). The 10×20 font has slightly different mapping than the other sizes: 0x93 is ö instead of ô, and 0x97 is missing (causing the following characters on the same line to be drawn at the wrong position). It also claims to be using codepage 720, but many characters differ from their CP720 mappings, including the bundled CP_720.NLS mappings (for example, ـ (U+0640 ARABIC TATWEEL) is 0x95 in CP720, but in the console 0x95 is ش instead, and the tatweel is at 0xFF). On Windows 9x, ReadConsoleOutputW is not supported so the UCS-2 mappings of the console character tiles cannot be captured (error 0x00000078 ERROR_CALL_NOT_IMPLEMENTED). When that program runs on Arabic versions of Windows NT, the visual output is of the CP437 character set if one of the bundled bitmap fonts is used (https://i.imgur.com/RxjtxMH.png), or the CP720 set if Lucida Console is used, with the Arabic letters either having glitchy font substitution (NT 4.0, NT 5.0/2000) or the .notdef glyph (NT 5.1/XP and up). In fact, it seems that the only Arabic bitmap fonts that occur in Windows NT are CP1256 fonts, which are not used in terminals. So this appears to be one of those permanent Windows compatibility regressions that occured when Windows 9x ended, where the terminals can no longer render legacy Arabic text. Even if the user managed to use registry hacks to set the font to Courier New or Simplified Arabic Fixed, it would still use the CP720 mapping which is not compatible with the Windows 9x set. It appears that in the Windows 9x Arabic terminal character set, 244 characters ( ﺀﺁﺂﺃﺄﺅﺇﺈﺊﺋﺍﺎﺏﺑﺓ►◄↕ﺕ¶§ﺗﺙ↑↓→←ﺛﹰ▲▼ !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~ﺝﺟﺡéâﺣàﺥçêëèïîﺧﺩﺫﺭﺯôﺳûùﺷﺻ£ﺿﻁﻅﻉﻊﻋﻌﻍﻎﻏﻐﻑﻓﻕﻗﻙﻛ«»ﹱ▒ﹲ│┤ﹴﹶﹷﹸ٠١٢٣ﹹﹺ┐└┴┬├─┼ﹻﹾ٤٥٦٧٨٩،ﹿﱞﱟﱠﳲﱡﳳﱢ┘┌؛؟¤ﻝﻟﻡﻣﻥﻧµﻩﻫﻬﻭﻯﻰﻱﻲﻳﳴﹼﹽﺱﺵﺹﺽﹳ°·■ـ) are already in Unicode, but 12 characters are not in Unicode: • 6 of them are pieces of lam-alef ligatures (0xDD, 0xDE, 0xF9, 0xFB, 0xFC, 0xFD) • 2 of them are shadda with fathatan ligatures without or with tatweel (0xD0, 0xD1) — in some legacy Microsoft fonts, shadda with fathatan is mapped to private use U+E818 • 4 of them are disunifications of seen/sheen/sad/dad occuring either with or without tail — ﹳ (U+FE73 ARABIC TAIL FRAGMENT) was originally encoded in Unicode 3.2 for CP864 compatibility; in that codepage, the forms of seen/sheen/sad/dad attach to the tail fragment — forms with included tail: 0x92, 0x95, 0x98, 0x8A — forms without tail (attaching to tail fragment like in CP864): 0xF3, 0xF4, 0xF5, 0xF6 If someone tried to make a Win32 console implementation and tried to implement both Windows 9x Arabic terminal character set compatibility and wide string API (ReadConsoleOutputW) compatibility simultaneously, then they would run into the issue that there is currently no standardized mapping to handle that scenario. What should Windows 9x Arabic console compatible implementations do in that case?
