[wleeaf]

I checked \X against GraphemeBreakTest.txt from Unicode 17.0.0, which matches PCRE2's unicode_version. I used the 32-bit library and matched an anchored \X from each boundary in turn, with pcre2_match (interpreter and JIT) and with pcre2_dfa_match. 43 of the 766 lines failed, the same ones in all three. Two rows of PRIV(ucp_gbtable) in pcre2_tables.c cause 27 of them:

• The Regional Indicator row has only the RI bit. \X therefore ends the cluster before an Extend, SpacingMark or ZWJ character that follows a regional indicator. UAX Update to Unicode 14.0.0 #29 rules GB9 and GB9a say not to break there. So does pcre2pattern ("Do not end before extending characters or spacing marks or the zero-width joiner (ZWJ) character"). Examples:
  • \x{1F1E6}\x{300} gives two clusters instead of one.
  • In \x{1F1E6}\x{1F1E7}\x{200d}\x{1F1E8}, the ZWJ ends up on its own instead of joining the flag.
• The Prepend row has every bit except Extended_Pictographic. GB9b says not to break after a prepend character, and pcre2pattern says "Do not end after prepend characters". So \x{6dd}\x{a9} (ARABIC END OF AYAH, COPYRIGHT SIGN) gives two clusters instead of one.

The change adds ESZ to the Regional Indicator row and the Extended_Pictographic bit to the Prepend row. PRIV(extuni), the matcher's own \X loop and the JIT read the same table, so all engines change together. The RI pairing code is unchanged: a following Extend becomes the new left property, so a regional indicator after it still breaks.

The other 16 failing lines all involve GB9c, the Indic conjunct rule from Unicode 15.1, which PCRE2 doesn't implement (pcre2pattern doesn't list it). Supporting it would need the Indic_Conjunct_Break property in the UCD tables, so I left it out.

Tests:

• testinput4 and testinput7 (DFA) get the same new subject lines in the ^\X block:
  • a regional indicator followed by extend, spacing mark and ZWJ;
  • prepend followed by Extended_Pictographic, other and CR.
• The new lines replace the subject line "There are no Prepend characters, so we can't test Prepend, CR". Prepend characters such as U+0600 and U+06DD exist now.
• With the old table, the RI and Prepend + Extended_Pictographic lines produce different output. The Prepend + other and Prepend + CR lines are there for coverage and pass either way.
• Perl 5.38.2's \X gives the same clusters for all six new subjects, so test 4 stays Perl-compatible.

Configurations run:

• gcc 13 Release, 8/16/32-bit with JIT: ctest passes all 4 tests, and RunTest 4 7 passes with and without JIT.
• gcc 13 Debug with -fsanitize=address,undefined: RunTest 4 7 passes for all three widths.
• I did not run the clang devcontainer presets or Valgrind.

ChangeLog: entry 4 under 10.50.