feat(compat): the eight general-purpose libraries a modern Vulkan app needs

[Sunrisepeak]

Audited Stehsaer/vulkan-rt's dependencies against this index: of its thirteen, four were already here (vulkan-headers, vulkan-hpp, vulkan-memory-allocator, nlohmann_json). This adds the rest except SDL3, which is a package-sized job of its own — see below.

Nothing here is specific to that project. glm, doctest, argparse and libcoro are the general answers to math, testing, argument parsing and coroutines; the cpptrace/libassert pair is the general answer to "an assertion that tells you what the values were".

package	version	shape
compat.glm	1.0.2	header-only, include root is the tarball root
compat.doctest	2.4.12	header-only; main comes from a macro, not a feature
compat.argparse	3.2	header-only
compat.mio	2023.3.3	header-only, commit-pinned (upstream has never cut a tag)
compat.gzip-hpp	0.1.0	header-only over compat.zlib, + two #include_next shims
compat.cpptrace	1.0.4	whole-source glob + defines
compat.libassert	2.2.1	8 TUs, depends on cpptrace
compat.libcoro	0.16.0	core half only — the tarball cannot build the rest

CN mirrors, all eight

gitcode.com/mcpp-res/<slug>, upstream re-hosted byte-identically and verified as such: every mirror was downloaded back and its sha256 compared against the declared one — 8/8 identical. That equality is what lets a single sha256 serve both arms of the url table.

Four of the eight are not "expose the headers and move on"

compat.gzip-hpp is frozen at a point where it no longer compiles. utils.hpp names uint8_t six times and decompress.hpp once, and neither includes <cstdint>:

include/gzip/utils.hpp:14:32: error: 'uint8_t' does not name a type

It is a consumer-side failure — the package has no C++ TU of its own, so cxxflags = {"-include", "cstdint"} (the repair compat.redis-plus-plus uses) cannot reach the TU that breaks. The one lever a descriptor has is a same-named shim first on the include path reaching upstream's with #include_next, which is exactly what compat.catch2 does for its <new> defect. Two shims, because decompress.hpp and utils.hpp are independent entry points, and include_dirs must list mcpp_generated first.

compat.cpptrace is shape E without the config header. Six symbol backends, six unwinders, three demanglers — and every backend file opens with its own #ifdef, so the source list is one glob and the whole configuration lives in defines, the way curl's unselected TLS backends compile to nothing. The backends chosen are the dependency-free ones on every platform: libgcc's _Unwind_Backtrace + dladdr + __cxa_demangle on unix, DbgHelp on windows. ⚠️ The cost is documented rather than hidden: dladdr does not read DWARF, so frames carry function names and no file:line. A libdwarf feature can add that later without changing what a consumer writes.

compat.libassert takes cpptrace as a real dependency where upstream FetchContents a private copy, so a consumer linking both gets one. magic_enum turns out not to be needed at all — LIBASSERT_USE_MAGIC_ENUM is switched on only under LIBASSERT_BUILD_TESTING, and the library sources never mention it.

compat.libcoro ships the core only, and the tarball decides that. Upstream's networking half resolves DNS through c-ares, which lives in vendor/c-ares — a git submodule, so an empty directory in any GitHub archive (measured: 0 files). The source list is upstream's LIBCORO_SOURCE_FILES before the networking block appends to it, transcribed rather than globbed, because scheduler.cpp, poll.cpp, the epoll/kqueue notifiers and all of net/ do not self-guard. coro/export.hpp is generate_export_header output the tarball does not carry — without it the first TU dies with a fatal error — so it is snapshotted, and CORO_STATIC_DEFINE goes in defines because it decorates declarations consumers include.

The tests assert behaviour, not compilation

• glm — cross product, column-major storage, and that GLM_FORCE_DEPTH_ZERO_TO_ONE really maps the near plane to 0 (the assertion that tells Vulkan's clip space from OpenGL's). Both switches are set by the member, not the package: they change the meaning of types crossing a library boundary.
• gzip-hpp — gzip magic and that a repetitive payload actually shrinks, not just that it round-trips.
• cpptrace — a four-deep chain must yield ≥4 frames with distinct return addresses; a single-frame stub passes a count check but not that one.
• libassert — drives the failing path and reads the report back: expression text, message, both decomposed operand values, and a non-empty stacktrace. That last one is the proof cpptrace arrives transitively, since the project declares only libassert.
• libcoro — 200 coroutines through a thread pool under a coro::mutex.
• doctest / argparse / mio — subcases run independently; a missing required positional must throw; a write through the mapping must reach the file.

Also in this PR

A separate first commit fixes something I got wrong in #369: vk::to_string IS part of the vulkan module. vulkan.hpp includes vulkan_to_string.hpp itself (guarded only by VULKAN_HPP_NO_TO_STRING), so reading vulkan.cppm's include list as exhaustive was the error. The claim is now an assertion in the member rather than a comment.

Not included: SDL3

Deliberately, and the state of the work is worth recording. The Linux half is already solved — SDL_build_config.h generated with this index's toolchain and the index's own X11 headers (yielding Video drivers: dummy offscreen x11(dynamic), the same shape compat.sdl2 uses), and the selected source set maps to exactly 73 whole directories + 1 partial (src/core/linux, skipping the dbus/IME files — the same set compat.sdl2 skips).

What is missing is the macOS and Windows source sets, which cannot be verified anywhere but CI, and compat.sdl2 — the closest analogue — is a 923-line descriptor. Bundling that here would make eight finished packages hostage to SDL3's iteration cycles. Happy to push it onto this same branch before merge if preferred.