C: struct-tag redefinition misfiles a field into the prior record → self-referential record → compiler SIGSEGV

• Type: bug (cparser struct layout / tag redefinition). Track A (shared symtab.inc + record field pool).
• Found: 2026-07-08 game-library ladder, ENet unity probe (feature-game-library-candidate-suite).

Symptom

The compiler SIGSEGVs (stack overflow) compiling the ENet unity build. Crash site RecordHasManagedFields (symtab.inc) recursing forever on a self-referential record: struct in_addr ends up with a field ip_dst whose UFldRec_ points back at in_addr itself.

Mechanism (diagnosed)

ENet includes <netinet/tcp.h> / <netdb.h> / <poll.h>, which crtl does NOT provide (see bug-c-crtl-missing-net-headers-enet), so the HOST headers are pulled — and they REDEFINE struct in_addr (already defined by crtl's netinet/in.h) and define struct ip_opts { struct in_addr ip_dst; ... }. On the tag redefinition + adjacent struct layout, ip_dst (a struct ip_opts member of type in_addr) is appended into in_addr's own field range and recorded with UFldRec_ = in_addr, i.e. a record containing itself by value — impossible in C, and it makes RecordHasManagedFields (and any UFldRec_ walk) recurse without bound.

Two fixes

1. Robustness (do regardless): RecordHasManagedFields — and any recursive UFldRec_ walk — must not infinite-loop on a cyclic/self record (depth cap or visited-set; a compiler must never crash on malformed input).
2. Root cause: struct-tag REDEFINITION must not append fields into the existing record's contiguous range (or must reject the redefinition), and a by-value field of the record's own type must be rejected, not recorded.

Repro

tools/install_lib_candidates.sh enet then compile test/gamelib/enet_probe.c (unity include of the ENet core .c) with -Ilibrary_candidates/enet/include -Ilibrary_candidates/enet. SIGSEGV.

Gate

ENet unity compiles (or errors cleanly); a minimal tag-redefinition repro; no compiler crash; self-host byte-identical.

Log

• 2026-07-08 — resolved, commit f9fe98f8.