kencc: `auto <type> <name>;' parses -- root the C23 declarator at LNAME

[staalmannen]

The first pcc run over EDG's generated C stopped on the fourth file:

array_new_aligned.c:25 syntax error, last name: __T41392176

autoadlist' began with xdecor', and xdecor -> xdecor2 -> tag -> ltag -> LNAME | LTYPE', so after LAUTO the parser could SHIFT a typedef name as the variable being declared. That collides with cname: LAUTO', the storage-class reduction, and yacc resolves shift/reduce in favour of SHIFT -- so auto size_t x;' took the C23 deduction path, read size_t' as the NAME, and died on the next token.

EDG's C writes `auto ;' for every local: 116,079 of them across 103 of its 133 files, so one conflict blocked the whole port.

-std= would not have fixed this. The conflict is resolved when yacc builds the table, so no runtime flag has a say; and there is no dialect disagreement to arbitrate, because C23 kept auto' as a storage-class specifier and its type inference applies only where auto' is the sole type specifier. Both standards want the same answer. The grammar can decide alone, since size_t' is LTYPE and x' is LNAME: a new autoxdecor' roots the first declarator at LNAME, so LTYPE leaves the shift set and that lookahead reduces to the storage class. *' still shifts -- auto *p = &x;' is deduction, while auto *p;' as a storage class would need implicit int.

Measured with bison, since kencc's sources need Plan 9's <u.h> and <libc.h> and cannot be compiled on the host:

                    before   after

shift/reduce 26 22
reduce/reduce 6 6 (pre-existing, LTYPEOF)
state 417 (adecl) 4 2 LTYPE and '(' shifts gone
state 517 (forexpr) 4 2 same, renumbered from 513

Claude-Session: https://claude.ai/code/session_01WGAwvvTwDg2yknFkmZ3qzs