The filter stored and compared words by std::hash<std::string> in size_t,
which differs between standard libraries and platforms, so a .dcf made on
one system never matched on another and the block list never worked
there (issue 215).
- ChatFilterCore.h: 64-bit FNV-1a over the entry's bytes, ASCII lower
case, fixed-width uint64_t everywhere stored or compared.
- .dcf version 3: little-endian magic, version, longest entry in words,
uint64 count and sorted uint64 hashes. Version 2 files are refused:
the allowed words cache is rebuilt from its .txt, an old
blocklist.dcf is logged as unreadable.
- The servers build blocklist.dcf from a plain blocklist.txt next to
them (one word or phrase per line) when it is newer.
- Blocked entries can be phrases: runs of consecutive words up to the
longest entry, the whole run marked. Whitelist chat still checks one
word at a time, as the client does.
- Dashboard: the chat filter API reads blocklist.dcf the same way
(status, phrase length), accepts blocked phrases, refuses allowed
ones, and explains phrase matches in its message test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
c1bcda8dd1