sdk: centralize cross-compile function availability

[brandonpayton]

Why

The software development kit (SDK) tells package build tools what Kandelo supports. Autoconf is the common configure-time tool that tests those capabilities, and musl is Kandelo's C standard library.

Make the Kandelo SDK's Autoconf target description authoritative enough for dependency-first Homebrew builds, instead of requiring each formula to repeat target availability overrides.

Kandelo's Wasm link path permits undefined imports. A conventional AC_CHECK_FUNC link probe can therefore report yes for a function absent from musl and the SDK archives; legacy probes can also report false negatives for exported functions by calling them with invalid signatures. GNU tar, Coreutils, Dash, Bash, Diffutils, Gettext, Bison, and the dependency roots exposed both directions during the tap migration.

Dynamic loading has another false-positive path: musl's weak dlfcn stubs link but only report "Dynamic loading not supported." An uncached AC_SEARCH_LIBS(dlopen) therefore returns none required, omits -ldl, and prevents the SDK compiler wrapper from selecting Kandelo's functional Wasm dlopen glue.

What changed

• record three verified-present musl functions: bsd_signal, isalpha, and stpcpy
• record 39 verified-absent extension functions used by dependency and leaf package waves
• include the final formula-surface audit results: Windows GetSystemTimeAsFileTime, FreeBSD cap_rights_limit, and glibc feenableexcept/obstack_printf
• record verified-working strerror(0) and strcasecmp behavior so gnulib consumers neither select unnecessary replacements nor abort while cross-configuring
• route AC_SEARCH_LIBS checks for dlopen, dlsym, dlclose, and dlerror through -ldl, which the SDK translates to libc/glue/dlopen.c
• keep unadorned ac_cv_func_dlopen=no, report the explicit SDK library as ac_cv_lib_dl_dlopen=yes, and remove contradictory duplicate defaults
• preserve caller overrides with ${var=default}, with automated coverage for defaults and overrides
• document that reusable sysroot/toolchain availability belongs in sdk/config.site; package recipes own only package-specific runtime, dependency-library, or semantic probes

No syscall behavior, host behavior, Wasm layout, ABI surface, generated artifact, Node path, or browser path changes.

Validation

Current stacked-head validation through scripts/dev-shell.sh on exact head 128a3f714fe5583cd25210f4e899b12258b3eb7f:

• SDK Vitest suite: 7 files, 73 tests passed, including sdk: preserve static linker argument order #913 linker ordering, config-site sourcing, caller-override coverage, -ldl selection, and compile integration
• scripts/check-abi-version.sh: ABI snapshot, C header, generated TypeScript bindings, and ABI version remained in sync
• git diff --check c6849bde6..HEAD: passed

The target-fact changes are unchanged by the #913 restack. Consumer and symbol-audit evidence recorded on the prior reviewed head includes:

• every added absent function was compared with all wasm32 sysroot archives; none exports the named symbol
• the final tap/open-formula audit classified library-dependent probes such as OpenSSL/ncurses and package-semantic probes such as PHP syslog as formula-owned rather than moving them into the SDK
• GNU Bash 5.2.37 configure/link integration using only package-specific facts produced the asserted target configuration
• GNU Wget 1.25.0 built with Gnulib's real group_member replacement instead of its obsolete always-false package stub
• GNU Make 4.4.1 selected the SDK -ldl glue and invoked a side-module callback
• GNU Coreutils 9.6, GNU tar 1.35, and Dash 0.5.12 built without formula-local reusable availability overrides and passed focused runtime coverage
• pristine GNU Diffutils 3.12 resolves both string checks as cached yes and configures without package-local function facts
• the current tap Cflow formula also built without this PR by truthfully selecting Gnulib's strerror replacement, proving these facts optimize and centralize target identity rather than masking a required replacement

Kernel, libc-conformance, POSIX, and browser suites were not rerun because this supplies build-time Autoconf cache facts only. Consumer executables were exercised through Kandelo; no bottles or package archives are published by this PR.

Merge order

This PR is stacked directly on fully-green #913. Merge #913 first, then retarget this PR to main; its two purpose commits remain the complete review diff.

Review state

Current head: 128a3f714fe5583cd25210f4e899b12258b3eb7f, based on exact #913 head c6849bde6cfbe51837227789840ecf2b05791070, on ABI 39. This is a main-repository PR for Brandon's review and must not be auto-merged.

Formula source acceptance can proceed in parallel, but final bottles must be rebuilt after this PR lands so they consume SDK-owned facts rather than duplicated formula overrides.

[github-actions]

Phase B-1 matrix build status — pr-874-staging

ABI v39. 71 built, 0 failed, 71 total.

Package	Arch	Status	Sha
icu	wasm32	built	7adac0d6
libcurl	wasm32	built	09364632
libcxx	wasm32	built	eb8b0f23
libcxx	wasm64	built	98408792
libiconv	wasm32	built	33ebf32a
libpng	wasm32	built	f6b060eb
libxml2	wasm32	built	ed004619
libzip	wasm32	built	3a18fa37
openssl	wasm32	built	2534f50d
openssl	wasm64	built	b1f606da
sqlite	wasm32	built	38d1e709
sqlite	wasm64	built	74ba7df5
zlib	wasm32	built	ed1f0382
zlib	wasm64	built	fe65fab5
bc	wasm32	built	74fe9422
bzip2	wasm32	built	fbf3db6e
coreutils	wasm32	built	4ad93277
curl	wasm32	built	8d9826d0
dash	wasm32	built	6aa80963
diffutils	wasm32	built	157e6de9
dinit	wasm32	built	32700f6b
fbdoom	wasm32	built	adfb0c3d
file	wasm32	built	dd352e11
findutils	wasm32	built	7c850a20
gawk	wasm32	built	97df82fe
git	wasm32	built	6c1ba9dd
grep	wasm32	built	16fdbb01
gzip	wasm32	built	21778374
hello	wasm32	built	7f3a5ebd
kandelo-sdk	wasm32	built	2e05e690
kernel	wasm32	built	5ad0bc5c
less	wasm32	built	c1d48473
lsof	wasm32	built	141d3a95
m4	wasm32	built	8fb73304
make	wasm32	built	8255da12
mariadb	wasm32	built	d7742568
mariadb	wasm64	built	c844b652
modeset	wasm32	built	bc6ee39f
msmtpd	wasm32	built	6eecbf15
nano	wasm32	built	9a4f539b
ncurses	wasm32	built	0ad43612
netcat	wasm32	built	04b88dfd
nginx	wasm32	built	8b0eaf1d
php	wasm32	built	65dc7d82
posix-utils-lite	wasm32	built	fe0864d3
ruby	wasm32	built	7b786f57
sed	wasm32	built	290ef527
spidermonkey	wasm32	built	3934d30d
tar	wasm32	built	c18feaf9
tcl	wasm32	built	50406646
unzip	wasm32	built	007b2902
userspace	wasm32	built	a0479d3e
vim	wasm32	built	116c55b3
wget	wasm32	built	4aa9cc71
xz	wasm32	built	540b9422
zip	wasm32	built	59ba947e
zstd	wasm32	built	5b01d1d5
bash	wasm32	built	3d825116
mariadb-test	wasm32	built	5f58870f
mariadb-vfs	wasm32	built	8b85364a
mariadb-vfs	wasm64	built	67f214cb
nethack	wasm32	built	c9baab43
node	wasm32	built	64a3eb03
spidermonkey-node	wasm32	built	4c31f9e3
vim-browser-bundle	wasm32	built	655462a3
nethack-browser-bundle	wasm32	built	1200fdd1
rootfs	wasm32	built	f73d1843
shell	wasm32	built	37b5964d
lamp	wasm32	built	729d141a
node-vfs	wasm32	built	5cfb8cd8
wordpress	wasm32	built	1252eb49

_{Auto-generated; replaced on each push. Raw data in the publish-status workflow artifact.}

[brandonpayton]

Wget 1.25.0 provides concrete downstream evidence for the new group_member target fact in this PR.

Without the shared ac_cv_func_group_member=no fact, Autoconf links a probe successfully because the Kandelo linker permits unresolved Wasm imports, even though musl provides neither the declaration nor the symbol. Wget then suppresses its bundled gnulib replacement and fails to compile at the undeclared group_member call.

With this PR stacked unchanged in the ABI 18 integration:

• configure reports checking for group_member... (cached) no
• generated config has HAVE_GROUP_MEMBER 0
• the build compiles libgnu_a-group-member.o
• the unmodified upstream Wget source completes its Homebrew source build
• the final fork-instrumented artifact passes the literal Node Homebrew test and exact Chromium HTTPS retrieval

This is a target capability fact, not a Wget-specific cache override; the Wget formula retains only its package-specific strerror(0) runtime probe result.

[brandonpayton]

Libmagic supplied a focused regression probe for the existing fmtcheck target fact.

The file 5.45 cross configure uses AC_REPLACE_FUNCS(fmtcheck). Without the canonical ac_cv_func_fmtcheck=no, Kandelo's unresolved-import link behavior can make that probe appear successful even though musl provides neither a declaration nor a symbol; compilation then fails in softmagic.c instead of selecting upstream's replacement implementation.

Commit 7e0c99881 adds focused coverage that:

• reads ac_cv_func_fmtcheck=no from sdk/config.site;
• proves a caller-provided result still overrides the default.

Validation through the repository dev shell:

• npx vitest run sdk/test/configure.test.ts: 7/7 passed
• npm --prefix sdk test: 57/57 passed
• git diff --check: passed

No ABI or generated artifact changes.

[brandonpayton]

Exact-head CI exposed one integration failure after ac_cv_func_group_member=no became authoritative: the legacy registry Wget recipe still injected a file-local always-false group_member stub after Gnulib's generated unistd.h declaration. Wget itself failed with a linkage-declaration conflict, and the same failure propagated into shell, lamp, node-vfs, and wordpress.

Fixed at 4a21076a0: remove the obsolete source injection, let Gnulib provide the configured replacement, and bump Wget's build revision from 2 to 3 because archive bytes change.

Fresh local evidence through scripts/dev-shell.sh:

• worktree-local musl SDK build completed;
• Wget 1.25.0 normal registry build completed;
• generated lib/Makefile records HAVE_GROUP_MEMBER = 0;
• libgnu_a-group-member.o was built and the final module contains 0026c91e t group_member;
• full SDK suite: 7 files, 59 tests;
• build-script syntax and git diff --check passed.

The five failed staging jobs were all manifestations of this one Wget build failure and should rebuild from revision 3 on this head.

[brandonpayton]

Absorbed by rebase-merged #967. Main is now 6efb411 with the sealed tree b133bfa018f7e2bc84c4845edf2f945b2f844954, and ABI 41 activation completed successfully in run 29550069085. The absorption audit confirmed this PR\x27s behavior and tests are present in that merged tree, so this source PR is closed and its branch is deleted.