edg: -C on the archive too, and the finalization pair EDG leaves open

[staalmannen]

cpfe linked and then crashed on every run with fault read addr=0x2 pc=0x2'. lib_src/main.c' defines void _main(void)' -- EDG's static-init helper, the one a target's own crt is expected to call -- and _main' is 6l's default entry point (6l/obj.c:305' INITENTRY = "_main"; arch/amd64/main9.s:3' TEXT _main(SB)). 6l makes that name an SXREF at startup and satisfies it from the first archive on the link line that defines it, and `cmd/edg' names libedg.a ahead of libap. So the program's entry became EDG's helper: it ran _Z12__call_ctorsv, which returns immediately (vars.c:49 has __head null and munch_ctors.c defines _ctors[1] = {0}), and then executed a plain RET.

main9.s reads the kernel's argument block at 0(SP), so 0(SP) at entry holds argc -- which is what that RET popped and jumped to. cpfe --help' is argc 2. Falsifiable with no rebuild: cpfe' alone must have given pc=0x1 and `cpfe a b c' pc=0x4.

-C is not implicated. There was no duplicate for dupokall to silence: main9.$O is pulled only to satisfy the entry symbol, its other global _tos being referenced from profile.c alone, so with libedg.a answering first that member never reached the link.

Renamed rather than dropped from OFILES: main.c is compiled with -D_main=__edg_main. Nothing anywhere references _main, so removing the member would have linked equally well and would have thrown away the only entry point into EDG's static-initialisation machinery.

The static-init gap is recorded, not fixed: nothing now calls _Z12__call_ctorsv, so a C++ program cpfe translates would not run its static constructors. cpfe does not need it and does not emit a call to _main either -- the name occurs in the 133-file corpus exactly twice, both in lib_src/main.c, with no string literal of it in the emitter.

One nm says _main is alone: all 51 lib_src files through gcc, then nm --defined-only -g, gives exactly four external definitions that are neither _Z-mangled nor __-prefixed -- _ctors, _dtors, _main, _new_handler -- and the other three collide with nothing. cmd/edg/mkfile asserted that every definition under lib_src is _Z-mangled or __-prefixed; that sentence is the one that would have found this, and it is corrected there.

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