feat(xpkg): auto-stamp install_dir for wrapper packages

[Sunrisepeak]

Summary

Make wrapper-package idempotency a framework concern instead of a per-package idiom. After the Install hook returns success and install_dir is empty, auto-write a .xim-installed marker so xlings's catalog probe (is_directory && !is_empty) doesn't keep flagging the package as not-installed and re-triggering the hook on every dependent install.

Context

xim-pkgindex's linux-headers.lua and glibc.lua already work around this in their install hooks by writing the marker manually:

function install()
    local install_dir = pkginfo.install_dir()
    if not os.isdir(install_dir) then os.mkdir(install_dir) end
    io.writefile(path.join(install_dir, ".xim-installed"), pkginfo.version())
    return true
end

That's good for known wrapper packages but every new wrapper-shape package hits the bug until the author finds and reproduces the idiom. Move it to the framework once.

Stamp file

Format: key=value INI. Grep / cat / read all friendly. schema = 1 reserves room to add fields later.

schema = 1
name = wrapper
version = 1.0.0
platform = linux

Behavior contract

Trigger	Auto-stamp behavior
Install hook succeeded + install_dir empty	✅ writes .xim-installed
Install hook succeeded + install_dir already non-empty	❌ no-op (don't stomp on hook-laid content)
Install hook failed	❌ no-op (failures shouldn't leave stamps)
Non-Install hooks (Config, Uninstall, …)	❌ no-op (auto-stamp is Install-specific)

Compatibility with existing manual stamps

Existing wrapper packages (linux-headers / glibc) that already write their own stamp continue to work. Their install hook populates install_dir first; the auto-stamp's is_empty gate sees non-empty, so no double-write or format conflict.

Stamp file format differs (manual = single-line version string; auto = key=value INI) but the file is a marker — content isn't load-bearing for the catalog probe. Future tooling that wants to parse it should handle both formats.

Decision NOT made: catalog xvm fallback

Originally proposed adding a "catalog falls back to xvm::has_version when install_dir is empty" as a belt-and-suspenders. Rejected on review:

install_dir 非空 should remain the single source of truth. If a stamp goes missing for any reason (manual cleanup, fs corruption), we want the catalog to truthfully report "not installed" so xlings re-runs the install, regenerating the stamp. Adding xvm DB as a secondary signal would mask such bugs and introduce dual-authority drift.

Self-healing via re-install > silent fallback.

Test plan

ExecutorTest: 26/26 pass (4 new auto-stamp cases):

• AutoStamp_WritesStampWhenInstallDirEmpty — wrapper case, stamp content fields verified
• AutoStamp_SkipsWhenInstallDirNonEmpty — hook populated install_dir → no double-write
• AutoStamp_SkipsWhenInstallHookFailed — failures don't leave a stamp
• AutoStamp_NotWrittenForNonInstallHooks — Config/Uninstall hooks don't trigger

LoaderTest 11/11, ModelTest 5/5, IndexTest 18/18 also still pass.

Cross-repo follow-up

After merge + v0.0.34 tag + mcpplibs-index 0.0.34 entry: a separate xlings PR will bump add_requires("mcpplibs-xpkg 0.0.33") → "0.0.34". xlings has zero code changes — same executor.run_hook(HookType::Install, ctx) call point now picks up the new auto-stamp behavior transparently.