ci(compile-smoke): darwin leg + msys cross compile with teeth

[ronaldtse]

Root cause

#122 — no darwin compile smoke. The release-src compile gate (tools/smoke_matrix → _release-line.yml) planned legs only over the shipped scenarios (linux-gnu / linux-musl / msys). Tfs::ReleaseDiff attributed a _darwin patch to the empty set — "no shipped scenario" — so the patch set's darwin-specific hunks (config_status_mainlibs_darwin.patch, configure_extstatic_bundle_loader_darwin.patch, the darwin paths of every base patch) compiled in no gated leg, ever. A darwin-only compile break would ship in the src release and surface downstream in the runtime factory.

#109 — the msys smoke never compiled _WIN32 code, and a 3.4 patch drifted from the c_api ABI. tools/compile_smoke --platform msys ran configure --host=x86_64-w64-mingw32 without pinning CC. Ruby's configure then used the host cc: the tree configured for the host, _WIN32 stayed undefined, and the windows half of every patch — exactly the code the msys scenario exists to compile — was never built, while the leg stayed green. Under a real cross gcc (x86_64-w64-mingw32-gcc 16.1.0) the hole surfaces immediately:

tools/compile_smoke: FAIL 3.4.9 (msys): compile.o did not compile
== make compile.o ==
compiling compile.c
In file included from compile.c:14768:
prism_compile.c: In function 'tfs_string_file_init':
prism_compile.c:44:29: error: passing argument 2 of 'tebako_fs_fstat' from incompatible pointer type [-Wincompatible-pointer-types]
   44 |     if (tebako_fs_fstat(fd, &sb) == -1) {
      |                             ^~~
      |                             struct stat *
ci/include/tebako/fs/c_api.h:80:49: note: expected 'struct tebako_stat *' but argument is of type 'struct stat *'

The drift is in the 3.4 prism_compile_memfs patches, not the ci/include stub. The stub is byte-identical with the c_api authority (the product repo's include/tebako/fs/c_api.h, the header the runtime factory's link unit vendors) on the whole declared surface — the pinned struct tebako_stat block and both declarations:

$ diff <(sed -n '/#if defined(_WIN32)/,/^#endif/p' <authority>) <(same <stub>)
STAT BLOCK IDENTICAL
$ grep -h 'tebako_fs_stat\|tebako_fs_fstat' <authority> <stub> | grep ^int | sort -u
int tebako_fs_fstat(int fd, struct tebako_stat* st);
int tebako_fs_stat(const char* path, struct tebako_stat* st);

The 3.4 patches declared the CRT struct stat sb (32-bit st_size on mingw) and passed it to tebako_fs_fstat, which fills the pinned 56-byte tebako_stat — invisible while the msys smoke compiled no _WIN32 code, an error under any recent cross gcc (and a stack overflow waiting to happen on the msys line: the callee writes 56 bytes into a smaller struct). The 4.0 line already carries the correct shape (struct tebako_stat sb;); 3.4 lagged.

The fix

• patches/3.4/prism_compile_memfs{,_6}.patch — the call sites now use struct tebako_stat sb;, mirroring the 4.0 line's patch verbatim (comment included). Hunk offsets adjusted; tools/lint verifies every 3.4 version still applies cleanly.
• tools/compile_smoke — CC is pinned to x86_64-w64-mingw32-gcc for the msys scenario, and a pinned cross compiler that is not on PATH is a named error with an install hint (never a silent host-tree smoke).
• Smoke universe — Tfs::Versions::SMOKE_ONLY_SCENARIOS = %w[darwin]: no darwin tarball ships (the darwin tree is the base selection plus the _darwin features), but the smoke universe is shipped scenarios + darwin. _darwin patches now attribute to [["darwin", nil]] (was []); base patches attribute darwin too. The build universe is unchanged — the darwin attribution matches no build row, so a darwin-only change still re-rolls no asset (spec-pinned).
• _release-line.yml — smoke legs with platform: darwin run on macos-latest (arm64); the apt toolchain step is skipped there.
• lint-patches.yml — the same compile gate now also runs at change time: tools/smoke_matrix diffed previous-release-tag..HEAD, so exactly the lines whose patch sets changed get legs (vacuously green when none did). Apply-clean is not compile-clean; a broken patch now fails on the PR, not at the release.

Leg definitions added

tools/smoke_matrix v0.0.0-next on this branch (diffs v0.2.38..HEAD — this branch touches patches/3.4 base patches, so the 3.4 tip gets every scenario):

{
  "include": [
    { "line": "3.4", "version": "3.4.10", "platform": "linux-gnu" },
    { "line": "3.4", "version": "3.4.10", "platform": "linux-musl" },
    { "line": "3.4", "version": "3.4.10", "platform": "msys" },
    { "line": "3.4", "version": "3.4.10", "platform": "darwin" }
  ]
}

The darwin leg runs as runs-on: ${{ matrix.platform == 'darwin' && 'macos-latest' || 'ubuntu-latest' }} in both _release-line.yml (the release gate) and lint-patches.yml (the change-time gate).

Evidence

Teeth proof (local, macOS arm64, x86_64-w64-mingw32-gcc 16.1.0): the msys leg now compiles _WIN32 code for real. A deliberate #error planted in the #elif defined(_WIN32) branch of patches/3.4/process_c_tebako_spawn_msys.patch (applied cleanly — tools/lint 3.4.9 OK) FAILS the leg:

tools/compile_smoke: FAIL 3.4.9 (msys): process.o did not compile
== make process.o ==
compiling process.c
process.c:2814:2: error: #error deliberate compile-smoke teeth check (reverted before merge)
make: *** [process.o] Error 1
exit=1

The break reverted, the same leg greens again (exit 0):

OK 3.4.9 (msys, pass2): 11 patched translation units compile (ext/digest/digest.o, win32/winmain.o, dir.o, dln.o, file.o, io.o, main.o, process.o, compile.o, ruby.o, win32/file.o)

(An earlier attempt planted the break in the base process_c_tebako_spawn.patch and the leg stayed green — correctly: the base patch is shadowed by its _msys sibling in the msys selection and never applies there. The break belongs to the msys variant, as above.)

Fixed-state compiles (local):

=== msys real cross compile, 3.4.9 (prism_compile_memfs_6 variant) — the #109 repro version ===
OK 3.4.9 (msys, pass2): 11 patched translation units compile (ext/digest/digest.o, win32/winmain.o, dir.o, dln.o, file.o, io.o, main.o, process.o, compile.o, ruby.o, win32/file.o)
=== msys real cross compile, 3.4.5 (prism_compile_memfs whole-line variant) ===
OK 3.4.5 (msys, pass2): 11 patched translation units compile (ext/digest/digest.o, win32/winmain.o, dir.o, dln.o, file.o, io.o, main.o, process.o, compile.o, ruby.o, win32/file.o)
=== darwin native, 3.4.10 (line tip; the new leg's runner family) ===
OK 3.4.10 (darwin, pass2): 8 patched translation units compile (dir.o, dln.o, file.o, io.o, main.o, process.o, compile.o, ruby.o)

compile.o is the prism path that errored in the repro above — green now.

Before state (this branch's parent): the repro transcript at the top of this body (real cross gcc, unfixed patches — the #109 error, exit 1). With no CC pinned at all, configure on a mingw-equipped host fails even earlier (cross objdump against host objects) — the old leg could only ever green by compiling a host tree.

Regression gates: bundle exec rspec — 121 examples, 0 failures (specs updated for the new darwin attribution and smoke universe, including a spec that the build plan still re-rolls nothing on a darwin-only change). tools/validate_manifests, tools/validate_msys_pairs green. tools/lint green for 3.4.5 / 3.4.9 / 3.4.10 (both prism patch variants apply cleanly). actionlint clean on both changed workflows. CI legs on this PR below.

Closes #122
Closes #109