← board

Rung 4: 257 applets

Rung 3 (feature-c-corpus-busybox-141-applets-linked) met its bar at 141 applets / 265 TUs. This is the next one: every applet the configured tree can build, which is 257 and 400 translation units.

257 is the tree's own answer, not a target chosen by us. It is what enabled_applets() reports with everything switched on; the harness asserts the count against include/applet_tables.h rather than trusting a NUM_APPLETS proxy, after that proxy claimed the tree had extras it did not have.

Where it stands

The gcc oracle links all 400 objects and agrees with upstream's own build over 621 cases, so the bar is reachable and the comparison is real.

Closed by attempting it, 2026-09-02, all measured rather than triaged:

cause closed by
getty.c: no tcsetpgrp crtl termios pair, 1799aad1a's predecessor
switch_root.c: no statfs crtl src/sys/statfs.c, plus the whole __NR3264_ family missing from <sys/syscall.h> for aarch64/riscv32
telnetd.c: no FD_ZERO crtl <sys/select.h> + select(2) over three kernel spellings, then 06b6b47e4 — libbb.h reaches fd_set only through <sys/types.h>, and giving <sys/types.h> glibc's #include <sys/select.h> exposed a splice-point bug the crtl auto-pull has had all along
fallocate.c: no posix_fallocate crtl, with the two-error-conventions trap
httpd.c: a 331-byte static const char suffixTable[] bc8fa306b — the block-scope string-array cap
dhcpc.c: "inline asm" NOT busybox's asm: the host's <asm/swab.h>, reached through <linux/filter.h>. Shadowed. Also needed cmsghdr/CMSG_* and AF_PACKET/SOL_PACKET
bc.c: "undeclared identifier passed as argument 1" 7d1b6cc12 — __VA_ARGS__ was substituted raw, so an argument landing inside a self-named macro's own call was never expanded

The count is 257, not the 258 first written here. Both readings agree — NUM_APPLETS in the configured include/applet_tables.h, and the applet list the harness prints and cross-checks against that table in both directions. The 258 was an off-by-one carried into this ticket, the inline-asm ticket and two comments in the harness before anything counted it twice.

Three of those seven were not what the diagnostic said. dhcpc's asm was a host header's; bc.c's "undeclared identifier" was a preprocessor gap four thousand lines away; and six of the eight FAIL lines in the first run all carried bc.c's error, because the harness reported grep error: | tail -1 over a build log shared by every translation unit. That last one is fixed too — per-file logs — and it is the reason this ticket can list causes at all.

Confirmed 2026-09-02 on a clean tree

tools/busybox_diff.sh --targets x86_64 --separate --applets <all 257> at 192c91a4b42f: the gcc oracle links 400 objects and agrees with busybox's own build over 619 cases; pxx compiles 399 of 400. The one refusal is networking/tls_sp_c32.c and the one undefined symbol in the link is curve_P256_compute_pubkey_and_premaster, which is that object and nothing else. --separate is required: coreutils/test.c and shell/ash.c claim ordinary identifiers through their globals macros and collide in both include orders, so no unity can hold the whole corpus — gcc's ceiling too, not pxx's.

Met, 2026-09-02

tools/busybox_diff.sh --targets x86_64 --separate --applets <all 257> at compiler 20953dc0444e: GREEN. 400 objects, one real link, and the linked binary agrees with the gcc oracle over all 619 cases — an oracle that itself agrees with busybox's own separately-linked build.

The last two blockers, both named by the attempt rather than by triage:

What this claim does NOT say

"pxx compiles and links busybox, byte-identical to gcc" is the claim. "crtl is enough to build busybox" is NOT, and it is the sentence a reader supplies for themselves. This build emits 1053 resolved from the host system (/usr/include), not pxx's own headers warnings over 110 distinct headers. Measured the day this closed, by attempting the SAME 400 translation units for i386, where there is no host to fall back on: 64 of them stop on a crtl header gap, 34 distinct headers, led by regex.h and netinet/udp.h at 7 translation units each. On x86-64 not one of those is visible, because the fallback answers every time and answers correctly.

That is the second architecture's ticket (feature-c-corpus-busybox-i386-the-second-architecture), not a defect in this one — but it belongs here, because this is the page anyone quoting the GREEN will read.

Scope

x86-64 only, like rung 3. aarch64 needs an --emit-obj object writer (feature-a-object-output-for-arm32-and-aarch64) and nothing smaller — 141-way unity is impossible for gcc too, measured.

Log