hexa-lang: Instructions file for Claude Code

hexa-lang

Project

Native .hexa compiler with an embedded theorem atlas + hx package manager. No LLVM anywhere: source→own-IR→native, linked by hexa_ld — byte-identical self-host fixpoint (gen3 ≡ gen4).

⭐ SELF-HOST DONE CRITERION (user standing · the finish line): self-host is "done" only when the toolchain needs ① no hexa_cc.c (the emitted hexa→C transpile fallback at self/native/hexa_cc.c — hexa build/run lowers every flow through the native source→own-IR→native path, never the C-transpile delegate) · ② no runtime.c (the ~5.5k-LOC emitted-C runtime/libc-syscall substrate — ported to native .hexa/.s/own-emit so nothing compiles a generated .c) · ③ no clang (nor any external C compiler) to build OR use hexa (install/build/run never shell out to clang/gcc — the own object-emit backend + hexa_ld do the whole compile+link). All three are emitted (gitignored) artifacts today, not hand-written source — the self/*.c hand-source floor is already ∅ (git ls-files self/**/*.c == ∅, #4356 cuda + #4352 timegm), but file-∅ of hand-written C is NOT the DONE line; the DONE line is eliminating all three ①②③. ⚡ ACTIVE GOAL (user standing · 2026-07): drive all three ①②③ to DONE via a perpetual 3-lane fleet (selfhost-done-3axis-fleet — lane1=①hexa_cc, lane2=②runtime_c, lane3=③clang-flip, fire-on-arrival until each crosses; per-axis status SSOT = ARCHITECTURE.json activeGoal + convergence.records[]). Latest: axis-③'s LINK half is demonstrated — the own ELF linker links own .o + runtime.a → static ET_EXEC (exit 42, ZERO binutils, data-seg/bss covered, opt-in --linker=hexa, #4711–#4718); the COMPILE half (--backend=native --emit=obj) already exists, so the axis-③ frontier is the default-flip. Guardrail: release integrity > self-host progress — every flip opt-in-first → byteeq 3-target GREEN + shipping smoke → default-ON; the sole inviolable = no-LLVM.

Progress toward that line (measured tracks): the emitted-C substrate is a reducible RUNTIME-PORT target whose authoritative nm-UND libc floor is exposed per-CI (advisory dump in nobaseline-gate.yml, #4360); flips drop libc UNDs one family at a time — strtod-tail native default-ON (#4651, T_mis=0 vs glibc) · own-start default-ON drops atexit/environ/__libc_start_main (FLIP-7) · MULTIOBJ print-family native redirect default-ON (HEXA_RT_STDIO_NATIVE, #4856) drops printf/fprintf/perror/snprintf glibc-stdio UND from the ship runtime.a (own raw write(2), 3-target byteeq + ON-path GREEN; opt-out =0). fortify __*_chk dropped via -D_FORTIFY_SOURCE=0 (#4361) · perf-neutral syscall leaves routed to raw-svc (getpid #4358 · setpgid #4364) · compiler-synth memcpy/memset — a reference-matched glibc-style overlapping-load-dispatch native body measures 1.3–2.0× faster than libc for the small-copy (struct) regime + parity at large (summer, clang-18; the naive-byte-loop "perf-wall" is FALSIFIED), so the -fno-builtin-memcpy flip is a measured-viable next round gated on byteeq 3-target + real-workload perf, not a permanent wall. The clang-dependency (③) is the native-emit/self-assemble endgame — own object emit + hexa_ld replacing the clang compile+link, so a consumer needs only the hexa binary. Every formula-bearing fn must cite an atlas law (@cite)/@verify/@grace, else the build refuses to emit (HX8004). Domain tracking is retired.