Target config: data.code_page silently accepts non-EBCDIC pages (shift_jis -> katakana 'sign characters'); a PC-COBOL (ASCII / Shift_JIS) estate cannot be declared #3958

Found while preparing the opensourcecobol4j fresh estate (#3806, epic #3803). This is a pre-trial finding; no porting was run.

Defect

data.code_page is documented as "the EBCDIC code page for zoned-decimal sign overpunch". validate() checks it with zoned_sign_characters(), which says it rejects anything that is "not a known EBCDIC code page". In fact it only checks codecs.lookup(cp), so any Python codec passes silently, and the "sign characters" are whatever that codec maps 0xC0-0xC9 / 0xD0-0xD9 to:

>>> zoned_sign_characters("shift_jis")
('ﾀﾁﾂﾃﾄﾅﾆﾇﾈﾉ', 'ﾐﾑﾒﾓﾔﾕﾖﾗﾘﾙ')      # half-width katakana, not sign characters
>>> zoned_sign_characters("cp1252")
('ÀÁÂÃÄÅÆÇÈÉ', 'ÐÑÒÓÔÕÖ×ØÙ')

CobolRecords gets those characters baked in (records_source), and EbcdicDecoderUtil / sortKey use the page as if it were EBCDIC. So a config naming the ASCII-family page an estate really uses produces wrong zoned-decimal codecs, and nothing warns about it. The estate here is opensourcecobol4j, GnuCOBOL or Micro Focus data: Shift_JIS or UTF-8 text with ASCII zoned decimals.

It also means the Japanese PC estate (#3806) cannot be declared honestly. There is no way to say "the data is Shift_JIS, zoned decimals are ASCII (0x30-0x39, with the dialect's negative convention)". The only accepted values are EBCDIC pages, or a non-EBCDIC page that is silently misread.

Minimal reproduction (origin/main 3f8bdb6)

import tempfile
from pathlib import Path
from gitgalaxy.tools.cobol_to_java.java_target import load_target
p = Path(tempfile.mkdtemp()) / "t.yml"
p.write_text("data:\n  code_page: shift_jis\n")
print(load_target(p).data.code_page)      # -> shift_jis   (accepted; expected: ConfigError, or an ASCII-family mode)
p.write_text("data:\n  code_page: nosuch\n")
load_target(p)                            # -> ConfigError: not a known EBCDIC code page  (only unknown codecs fail)

Expected

Either:

• reject non-EBCDIC pages with the error the code already promises; or
• (better, for PC-COBOL estates) support an ASCII-family data encoding: the text codec (shift_jis / cp932 / utf-8 / cp1252), ASCII zoned-decimal signs (the dialect's own convention), and ASCII byte-order collation.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TmgE1CwLgaqxqKEgQn5Dw5