build-and-verify

name	build-and-verify
description	Build C-Kermit (make linux → wermit), regenerate ckcpro.c after editing ckcpro.w, and verify changes are behavior-preserving — SOURCE_DATE_EPOCH byte-identical rebuild gate, per-object .o comparison, macOS/BSD

Building and verifying C-Kermit

Build

There is no autoconf/configure; the hand-maintained makefile (~1200 lines) contains per-platform targets (Linux, macOS, modern BSD only) that invoke make recursively, passing options through KFLAGS/LIBS environment variables. Do not use make -e.

The standard build on this machine (Linux):

make linux

Other useful invocations:

make clean          # removes objects, the wart binary, and generated ckcpro.c
make check          # reports SUCCESS if an executable wermit exists (rm wermit before building so this is meaningful)

The resulting binary is always named wermit, regardless of target. There is no test suite; verification is building and running the binary. ckubuildlog is a Kermit script used to produce build-log table entries after a successful build, not part of the build itself.

A build leaves wermit, wart, and the generated ckcpro.c in the tree; all three are gitignored (along with *.o) — do not commit them. make clean removes wart and ckcpro.c but not wermit. (resume-claude.sh in the tree is a local Claude Code session helper, also not part of the build.)

Most platform targets (including linux) build via the xermit link target, which uses the select()-based CONNECT module ckucns.c. The older wermit link target uses the fork()-based ckucon.c instead.

Protocol changes: ckcpro.w and wart

ckcpro.c (the protocol state machine) is generated, not hand-edited. It is produced from ckcpro.w (a lex-like grammar) by the wart preprocessor (built from ckwart.c). This conversion is NOT run automatically by the platform targets — after editing ckcpro.w you must run, before the normal build:

make wart
make ckcpro.c

Verification methodology

Every "behavior-preserving" change in this tree is verified at the strongest applicable level. Pick the level that matches the change:

1. Byte-identical whole-binary rebuild — for pure source-spelling changes (macro expansion, typedef renames, cosmetic reformatting). Build original and modified trees with a fixed timestamp and compare:

   make clean && SOURCE_DATE_EPOCH=1750000000 make linux   # in each tree
   cmp wermit.orig wermit
   

   The build must be fully clean (make clean first). A non-clean rebuild relinks stale .o whose __DATE__ perturbs the GNU build-id, giving a spurious early-byte diff; stale .os also make a whole-binary diff misleading in the other direction.

2. Per-object .o comparison — when one file legitimately differs (e.g. an intentional user-visible change, or clang-format InsertBraces perturbing register allocation). Do the fixed-epoch clean build of both trees, compare all *.o pairwise, and explain every non-identical object (e.g. disassembly at -O0 showing only label renumbering).

3. Token-stream proof — for mechanical text substitutions, show replacer(git HEAD) == new file exactly, or that a comment/string-aware token stream is identical before/after. This covers #ifdef branches that never compile on Linux.

4. macOS/BSD path checks — make linux alone does NOT cover the macOS/BSD #ifdef branches. Check them with:

   gcc -E -DBSD44 [-DMACOSX10 | -D__OpenBSD__ | -D__NetBSD__ | -DFREEBSD] file.h   # preprocessor-level claims
   cc -fsyntax-only -DBSD44 -DMACOSX ... file.c                                     # syntax-level
   

   Remember: MACOSX10 → MACOSX → BSD44 → SVR4 → SVR3/ATTSV are all live there, and POSIX/CK_POSIX_SIG are NOT (they are Linux-target defines).

5. Runtime smoke tests — after any signal- or timing-related change, run the battery from the signal-handling skill: clean startup/exit, PAUSE 2 completes in ~2.00s (SIGALRM longjmp timer), SIGINT mid-PAUSE aborts back to the command loop.

The make linux build is warning-clean; keep it that way — a new warning is a regression.