feat(xpkg): subos.env declarations, and one source for the Lua stdlib — 0.0.48

[Sunrisepeak]

What

subos.env{var, op, value, binding} — a package declares an environment variable its subos must export to every process entering it.

PATH and RPATH cover linking, not discovery. A GL driver is found through LIBGL_DRIVERS_PATH, an EGL vendor through __EGL_VENDOR_LIBRARY_DIRS, a font config through XDG_DATA_DIRS. None of those can be linked in, and the process that has to see them is the user's own binary — which xlings never wraps, so xvm.add's per-shim envs cannot reach it.

Carried to the consumer as an op = "subos_env" entry with its own var/value/mode fields. The recipe's set/prepend choice travels as mode because op is already the category the consumer dispatches on. Slice 1 implements set and prepend; append is refused and returns false rather than quietly behaving like set.

The probe rule this establishes

Recipes must write:

if type(subos.env) == "function" then ... end

not if subos.env then. import() answers an unknown module with a permissive proxy whose every key is a truthy callable table — so the truthiness test is true on clients that predate this module, the recipe takes the new branch, and the call evaporates. Install succeeds, nothing is configured, nothing complains.

if xvm.files then in the V2 spec is safe only because xvm is a module those clients already ship, so the missing field really is nil. A missing module never is. SubosEnv_ProbeMustTestTypeBecauseUnknownModulesAreTruthy pins both halves.

Build: mcpp only

xmake and CMake entry points are gone (the CMake one still declared project(mcpplibs-templates) and a templates_test target — it never built this library). mcpp build / mcpp test is the only path, which CI already used.

The Lua stdlib now has one copy under version control. build.mcpp generates the embedded .cppm into MCPP_OUT_DIR instead of it being committed next to its own source.

Those two had already drifted: xvm.files shipped in 0.0.47 as an edit to the generated file alone, and src/lua-stdlib/xim/libxpkg/xvm.lua never had it. An xmake build would have regenerated from that stale source and silently deleted the feature — nothing reads both copies, so nothing could notice. This restores the content to the .lua and removes the shape that allowed it.

Also refreshes the CI pins (xlings 0.4.69 → 2026.8.4.2, mcpp 0.0.109 → 2026.8.4.1). Those rot silently and fail looking like network errors.

Verified

mcpp test — 100 tests pass (the 4 macOS-only elfpatch cases stay filtered on a Linux runner, as before).

Consumed by openxlings/xlings subos slice 1: .agents/docs/2026-08-05-subos-minimum-design.md.