I did a comparison between lemon (from sqlite) byacc (from https://invisible-island.net/byacc/byacc.html) and bison with a custom command line option to ignore all precedence settings then tested with sqlite3/cfront3 grammar (see here https://github.com/mingodad/lalr-parser-test).
Probably still there is something I'm missing or misunderstood (pull requests are welcome) but again differences between how lemon/byacc/bison parse LARL(1) grammars seems strange:
For sqlite3 grammar:
lemon : zero conflicts
byacc/bison: 53 reduce/reduce conflicts
Again for sqlite3 grammar ignoring all precedences:
lemon/byacc: 827 parsing conflicts, 788 shift/reduce, 39 reduce/reduce
bison: 827 parsing conflicts, 734 shift/reduce, 93 reduce/reduce
Now for cfront3 grammar:
lemon: 19 parsing conflicts, 19 shift/reduce, 0 reduce/reduce
byacc: 24 parsing conflicts, 21 shift/reduce, 3 reduce/reduce
bison: 24 parsing conflicts, 20 shift/reduce, 4 reduce/reduce
Again for cfront3 grammar ignoring all precedences:
lemon/byacc: 1077 parsing conflicts, 1074 shift/reduce, 3 reduce/reduce
bison: 1077 parsing conflicts, 1073 shift/reduce, 4 reduce/reduce
I'll appreciate any help on figuring out what's happening here.