Cross-platform SDL3 migration: Phase −1 → Sprint 7 (Win32 removal, rendering rewrite, .NET AOT, upstream merge)

[yesid-bocanegra]

Summary

End-to-end cross-platform port of the MU Online client. Starts from a
Windows-only, DirectX / WinINet / DirectSound / Win32-GDI stack on a
monolithic MSVC build. Ends at a native macOS arm64 + Linux x64 + Windows
MSVC client on SDL3 + SDL_GPU + SDL_ttf + miniaudio + libcurl + glslang +
.NET 10 Native AOT, plus an spdlog-based unified logging facade.

Stats: 250 commits, 1,188 files changed (+222,960 / −68,063).
Base: main. Head: cross-platform-sdl-migration-merged.

The work is organized as PCC / BMad sprints (Sprint 5 → Sprint 7) plus a
mid-flight merge of 406 upstream commits. Each section below maps directly
to the stories in the commit log.

Phase −1 — Modular reorganization (foundation)

Moves 692 source files from a flat directory into 20 module directories and
splits the monolithic CMake build into 7 independent static libraries +
MUGame + Main. Every later phase depends on these boundaries.

• Main/, Core/, Protocol/, Network/, Data/, World/, Gameplay/
• UI/Framework/, UI/Windows/, UI/Events/, UI/Legacy/
• Audio/, RenderFX/, Platform/

CMake targets: MUCommon (INTERFACE) → MUCore, MUProtocol, MUData,
MURenderFX, MUAudio, MUThirdParty, MUPlatform → MUGame → Main.
PCH shared via REUSE_FROM MUCore.

Build system, dependencies & tooling

• CMake presets: windows-x64, macos-arm64-debug, linux-x64.
• C++20; warnings-as-errors (/WX, -Werror).
• Quality gates via ./ctl check: clang-format, cppcheck, and a clang-tidy
  portability gate (bugprone-sizeof-expression,
  bugprone-suspicious-memset-usage,
  bugprone-misplaced-pointer-arithmetic-in-alloc) that catches the
  sizeof(pointer) vs sizeof(element) bug class invisible on 32-bit.
• Dependencies via FetchContent/CPM: SDL3, SDL_ttf 3.2.2, spdlog 1.15.3,
  GLM 1.0.1, miniaudio, Catch2 3.7.1, OpenSSL, libcurl, glslang + spirv-cross,
  libjpeg-turbo.
• Shader pipeline: HLSL → SPIR-V (glslang) → MSL (spirv-cross), with
  committed reference blobs in src/shaders/compiled/. Replaces SDL_shadercross.
• .NET 10 Native AOT library (MUnique.Client.Library.{dll,dylib,so}) built
  automatically as a Main dependency; platform-correct lib extension via
  MU_DOTNET_LIB_EXT; LIBRARY_PATH injects Homebrew + OpenSSL dirs so the
  AOT linker finds brotli, icu4c, libssl/libcrypto.
• CI: Windows native MSVC (MinGW cross-compile job removed — it shipped a
  crippled SDL3-disabled build), macOS arm64, Linux x64. Artifact upload
  for macOS and Linux on push.

Sprint 5 — Audio (DirectSound → miniaudio)

• 5-2-1 / 5-2-2 — MiniAudioBackend replaces DirectSound for BGM and
  SFX. DSPlaySound.cpp free functions delegate to g_platformAudio;
  Set3DSoundPosition upgraded from stub to per-frame position tracking;
  path normalization (\ → /) in LoadSound.
• 5-3-1 — 17 Catch2 tests for WAV (16-bit mono/stereo), MP3, OGG decode
  via ma_decoder; headless-CI-safe.
• 5-4-1 — Independent BGM + SFX volume (IPlatformAudio, 0.0–1.0
  floats), persisted in GameConfig with migration from legacy
  VolumeLevel; safe on uninitialized backend.

Sprint 6 — Gameplay regression coverage (EPIC-6)

Catch2 RED-phase component test suites validating the gameplay invariants
the SDL3 port must preserve. Five file-scan / struct-layout suites:

• 6-1-1 auth / character select, 6-1-2 world navigation (grid
  geometry, MOVEINFODATA gates, TW_* flags), 6-2-1 combat (34
  TEST_CASEs across AC-1..AC-6), 6-2-2 inventory / trading / shop (slot
  grid + equipment + storage enum), 6-3-1 social systems,
  6-3-2 quest / pet / PvP / duel constants, 6-4-1 UI windows.

Sprint 7 — Cross-platform completion (EPIC-7)

Platform & build (7-3 through 7-5)

• 7-3-0 UTF-8 continuation-byte validation + overlong-sequence rejection
  (RFC 3629) on the MultiByteToWideChar boundary; remove stale <codecvt>
  includes.
• 7-5-1 mu_swprintf POSIX wide-string format helper; iterative macOS
  arm64 zero-warning build sweep.

Win32 removal (7-6)

• 7-6-1 macOS native build — 9 GDI stubs in PlatformCompat.h, Homebrew
  LLVM toolchain.
• 7-6-2 Win32 <windows.h> / <tchar.h> purge from string-heavy TUs.
• 7-6-3 Data-layer Win32 removal + OpenSSL crypto (OPENSSL_cleanse on
  key material, proper error reporting).
• 7-6-4 Cross-platform CPU-usage API — 1.0 = 100% of all cores
  (formula divides by m_numProcessors); signed guard against clock
  anomaly wraparound; documented not-thread-safe constraint.
• 7-6-5 Win32 console APIs (SetConsoleTextAttribute,
  FillConsoleOutputCharacter, GetConsoleScreenBufferInfo) replaced with
  ANSI escape sequences and a 16-entry RGBI → SGR table. Windows 10+ enables
  ENABLE_VIRTUAL_TERMINAL_PROCESSING automatically.
• 7-6-6 ShopListManager HTTP: WinINet → libcurl
  (BuildURL + ConfigureCurl + curl_easy_*); URLDownloadToFile →
  inline libcurl download; Win32 types (TCHAR, INTERNET_PORT, DWORD,
  BOOL) → wchar_t/unsigned short/uint32_t/bool.
• 7-6-7 ErrorReport diagnostics cross-platform; integer-overflow fix
  in RAM detection (static_cast<int64_t> instead of int).
• 7-8-1..4 Audio interface Win32-type purge; gameplay header
  cross-platform (SKILL_REPLACEMENTS inline linkage, transitive include
  fixes); test compilation fixes (header self-containment, WZResult
  self-assignment guard, PadRGBToRGBA bounds).
• 7-8-4 .NET 10 Native AOT build.

Rendering migration (7-9 — the largest sub-epic)

The OpenGL fixed-function → SDL_GPU migration, delivered as 12 stories plus
~40 runtime-fix commits after the first screen went black.

• 7-9-1 macOS game loop & render path migration.
• 7-9-2 IMuRenderer interface extensions: BeginScene/EndScene,
  Begin2DPass/End2DPass, ClearScreen, RenderLines, IsFrameActive.
• 7-9-3 Entry-point unification: MuMain() replaces WinMain on SDL3.
• 7-9-4 DirectSound removal validation (file-scan tests for zero
  Win32-wave and zero IDirectSound* symbols).
• 7-9-5 Kill-all-cross-platform-stubs: file-scan tests for the absence
  of wglCreateContext, etc. — dead-code audit turns GREEN when stubs are
  deleted.
• 7-9-6 Raw GL → MuRenderer migration across ~100 files; 628
  glColor* calls removed.
• 7-9-7 GLM 1.0.1 adoption (GLM_FORCE_DEPTH_ZERO_TO_ONE for
  Metal/Vulkan clip space, right-handed coords preserved); replaces
  ~150 lines of hand-rolled mat4::. Apple-Silicon depth format fix
  (D24_UNORM → D32_FLOAT). Depth-read-only pipeline variants for
  transparent/additive passes. Matrix-stack overflow diagnostics.
• 7-9-8 SDL_ttf 3.2.2 GPU text engine; CUIRenderTextSDLTtf
  implements IUIRenderText; glyph-atlas warmup; 4 HFONT → TTF_Font*
  mapping; Y-up ortho with SDL_ttf's pre-negated vertex Y.
• 7-9-9 Text-input boxes on SDL3: focus hygiene, shared-buffer fix
  (per-instance InputBoxTexture), 8×16 bitmap font scaling against
  g_fScreenRate_y, GiveFocus alternation from overlapping hit regions,
  SendLogin wchar_t → MU_C16 marshalling to .NET (PtrToStringUni).
• 7-9-10 CUITextInputBox → SDL_ttf direct rendering; deletes the GDI
  bitmap-font pipeline, per-instance GPU textures, UploadText, and ~367
  lines of section-splitting math; blinking caret actually visible now.
• 7-9-11 foundation — CUITextInputBox::Reset() + Configure(InputBoxConfig)
  declarative setup; extern singletons centralized in UIControls.h;
  SetIsPassword(bool) runtime toggle.

Post-migration runtime fixes (not story-tagged, large volume):

• MVP & pipeline state. Full matrix stack implementation replacing
  no-op SetMatrixMode/LoadIdentity/Rotate/Translate/PushMatrix
  stubs; mu_gluPerspective as real perspective matrix. Vertex attribute
  layout fix (normals/UVs/colors were mapped to wrong shader locations,
  producing solid yellow across the entire 3D scene). ScreenSize uniform
  push fix (every vertex was clipped to infinity → black screen despite
  559 draw calls/frame). HiDPI viewports use swapchain physical dims, not
  logical window size.
• Blend, depth, cull. 45-entry pipeline matrix across blend mode × depth
  mode × 2D/3D layout; back-face culling on opaque 3D (CCW front face);
  LESS depth for 3D (blocks same-depth glow self-overlay), LESS_OR_EQUAL
  for 2D (normal UI stacking); LEQUAL on depth-read-only for chrome/
  bright/glow overlays (recovers classic SetDepthFunc(GL_LEQUAL) intent).
  Depth write only for BlendMode::Alpha. Color-mask and stencil-test
  draw-skip paths to suppress shadow-volume artifacts until stencil lands.
  RenderColor default alpha changed to semi-transparent black so tooltip,
  message-box, inventory highlight, and progress-bar backgrounds render
  correctly (~114 call sites).
• Near-plane & sprite clipping. World-space billboard construction via
  camera right/up vectors extracted from CameraMatrix; SV_ClipDistance
  for proper near-plane triangle clipping (after a w<0.1 vertex discard
  attempt collapsed legitimate geometry). Deferred-draw command system:
  vertices for a full frame collected and copied BEFORE the render pass,
  eliminating the 1-frame vertex delay that caused streaking when particle
  counts varied between frames. Vertex buffer 4 MB → 16 MB.
• Texture lifecycle & state sync. ResolveTextureId() helper to honor
  BindTexture global-state semantics from OpenGL-era callers while
  SDL_GPU requires explicit per-draw texture IDs (~25 call sites across
  ZzzBMD, ZzzLodTerrain, CSWaterTerrain, effects). Disabled-texture
  path returns white texture so texture × vertexColor = vertexColor
  preserves glDisable(GL_TEXTURE_2D) behavior. BeginOpengl now syncs
  renderer state alongside CPU-side tracking variables. Dangling sampler /
  texture guards in deferred command replay (Metal objc_retain crashes
  on freed Obj-C objects across scene transitions). CachTexture
  optimization removed (it bypassed direct BindTexture callers,
  desyncing state).
• Viewports & scissor. SetViewport / SetScissor overrides for
  UI-embedded 3D previews (CharMake photo viewer, NPC shops). Sticky
  scissor re-applied before every draw command (Metal drops it on pipeline
  bind). glViewport2 Y-flip removed (SDL_GPU uses top-left origin).
  Begin2DPass resets viewport to full screen (previously a no-op, UI
  rendered inside shrunken 3D viewport — left black bars).
• Dynamic texture upload for GDI/bitmap-font paths. QueueTextureUpdate
  processed in EndFrame copy pass — replaces the glTexSubImage2D
  removed during migration. Fixes text input boxes and any remaining UI
  text on the GDI bitmap-font path.

Input (SDL3) fallout

• Button hover-before-click (clicks accepted when UP→OVER transition
  happens in the same frame as the click).
• g_hWnd sentinel (HWND)1 so GetFocus() comparison in
  CNewUIManager::UpdateKeyEvent matches and hotkeys work
  (I=inventory, C=character, F1-F5 skill bar).
• Mouse push/edge flag lifecycle: accumulate across PollEvents,
  clear in RenderScene after scene logic consumes them. Preserves
  click detection when BUTTON_DOWN+UP land in the same event batch
  (common at startup).
• GetAsyncKeyState log-spam fix (~9,000 writes/sec removed).
• Text-input focus hygiene: IsAnyInputBoxFocused() suppresses global
  hotkeys while typing; hotkeys gated on m_bSDLHasFocus AND m_iState != UISTATE_HIDE, not just a non-null tracker; destructor
  calls MuStopTextInput when the destructing box was focused.
• PollEvents text buffer preserved across calls (previously cleared
  every frame-throttling tick, losing typed characters).
• Close-button click leak: CNewUISystem::Hide consumes in-flight
  left-click state so closing a menu doesn't issue a click-to-move
  at the close-button screen position.
• Enter key chat open on non-editor builds (#ifdef _EDITOR gate was
  swallowing VK_RETURN outside the editor).

Network / .NET Native AOT

• Thread-safe packet queue (mutex-protected, drained per-frame):
  PostMessage is a no-op on SDL3, so HandleIncomingPacket was
  silently dropping every packet → server timed out and disconnected.
• .NET side: ConcurrentDictionary for connections map;
  ConnectionWrapper dispose-once + send-after-dispose guards.
• SIOF (static init order fiasco) in PacketBindings_*.h:
  ResolvePacketBindings() re-resolves all 191 ClientToServer inline
  function pointers after the library handle loads. Previously only
  ConnectServer + ChatServer were resolved → segfaults at
  SendPing post-login and SendRequestCharacterList on character
  select. Also stripped UTF-8 BOMs from auto-generated bindings.
• munique_client_library_handle SIOF: moved from inline in
  Connection.h to Connection.cpp after g_dotnetLibPath.
• SIGSEGV handler conflict: .NET Native AOT uses SIGSEGV for managed
  null-ref checks and GC write barriers. Our POSIX crash handler
  intercepted these, printed a crash report, prevented recovery →
  every .NET send terminated the process. On MU_ENABLE_SDL3, skip
  SIGSEGV installation and let .NET own it (SIGABRT/SIGBUS still
  installed).
• PtrToStringAuto → PtrToStringUni fix in the XSLT template:
  Auto reads UTF-8 on macOS but the C++ side passes UTF-16 via
  MU_C16. Only the first character of each string survived. Affects
  35+ string parameters across ClientToServer, ChatServer,
  ConnectServer.

Logging (7-10-1)

spdlog 1.15.3 integrated as MUCore PUBLIC dep. MuLogger.h facade with
11 named loggers (core, network, render, data, gameplay, ui,
audio, platform, dotnet, gameshop, scenes). Rotating file sink
(512 KB × 3) + colored stderr sink (warn+). Crash handler file descriptor
preserved via O_WRONLY|O_APPEND.

Migrated:

• 277 g_ErrorReport.Write call sites → per-module loggers
• 339 LOG_CALL macro sites → SPDLOG_TRACE + direct calls
• 140 g_ConsoleDebug->Write sites (MCD_* → spdlog levels)
• ~29 fprintf(stderr) diagnostics → SPDLOG_DEBUG

Deleted ErrorReport.{h,cpp} and muConsoleDebug.{h,cpp}; extracted
MuSystemInfo. Runtime console commands $loglevel / $loggers
preserved via MuConsoleCommands. Generic fmt::formatter<T> for enums
handles fmt 11.x breaking change.

Data-layer crash fixes

• PetProcess::LoadData — sizeof(float*) used where sizeof(float)
  was meant. Invisible on 32-bit Windows (sizeof(ptr) == sizeof(float) == 4),
  caused a 2× heap overflow and SIGBUS on arm64. Related lint added to
  the quality gate.
• Text_Eng.bmd / Text_Eng_decrypted.bmd — macOS case-insensitive
  filesystem collapsed them into one file; restored from git history.
  GlobalText::Load uses mu_narrow_path() for POSIX separator
  normalization.
• mu_narrow_path() centralized in PlatformCompat.h — unified helper
  for backslash → forward-slash conversion on POSIX, replacing scattered
  local shims (NarrowPath in GlobalBitmap, wcstombs buffers in
  ShopList / Path, std::filesystem::path in CSQuest).
• OpenFilterFile double-fclose removed (UB).
• mu_mbclen no longer returns 0 on null byte — caused infinite loops on
  wide strings containing characters with null low bytes (e.g. Korean
  U+AC00).
• Skill-book tooltip SIGSEGV: GetSkillByBook computed
  SkillAttribute[-1] for ITEM_SCROLL_OF_POISON (offset mismatch vs
  ITEM_SCROLL_OF_METEORITE). Explicit remap + bounds check.
• Guild-make window: RenderGuildColor refactored to render direct colored
  quads instead of mutating shared BITMAP_GUILD.Buffer 81×/frame (which
  produced last-writer-wins under the deferred pipeline and triggered a
  Metal GPU abort after ~10 seconds). CUITextInputBox::GetText switched
  to wmemcpy with explicit length cap to eliminate wcsncpy null-pad
  overflow into 400-byte stack buffers.
• Friend-add SIGSEGV: HandleMessage(TXTRETURN) now validates the
  iParam2 heap-range before dereferencing; ReturnText zero-inits
  empty input, passes correct buffer size, and frees on early-return
  (long-standing 1024 B/empty-Enter leak).
• mu_wchar_to_char16 capped at 4096 iterations to prevent walking off
  unterminated buffers into unmapped memory.

Upstream merge (f5d1d73e — 406 commits, 29 conflicts resolved)

Major upstream PRs integrated:

• bugfix/fix mu helper not healing party members #304, bugfix/self-position skills not working in mu helper #305 — Mu Helper: heal-party index increment fix, IsMonster()
  filter on self-defense (was coercing pTarget->Object.Kind = KIND_MONSTER),
  TargetX/TargetX typo fix, self-position AoE skills (Nova, Hellfire,
  Inferno) dispatch path, pickup dedup token.
• fix(items): correct divine stick of archangel display #306 — Divine Stick of Archangel: IsDivineArchangelWeaponItem /
  IsDivineArchangelWeaponModel helpers unify the wing-type / item-name
  color / tooltip color / model registration / texture-load checks.
• fix(npc): wandering merchant Zyro proper model + allow interactions #310 — Wandering Merchant Zyro added to IsForcedNpcMonsterType
  allowlist.
• add right click drop item support and refactor CNewUIMyInventory::HandleInventoryActions #311 (architecturally significant) — CNewUIInventoryActionController
  • IInventoryActionContext interface; ~1,500 lines of helpers moved off
    CNewUIMyInventory. Bounds-check hardening on iSourceIndex against
    [MAX_EQUIPMENT_INDEX, MAX_MY_INVENTORY_EX_INDEX).
• fix: remove FPS-dependent spawn offsets from skill effects #312 — Skill-position fix via DetermineMonsterObjectKind extraction
  • IsForcedNpcMonsterType.
• Fix login scene cleanup ordering #313 — Login scene race: ReleaseLogoSceneData() moved before
  character-request send.
• Fix JumpTime push movement  #314 + Fix walk jitter #324 — Walk jitter (+ FPS_ANIMATION_FACTOR →
  * FPS_ANIMATION_FACTOR) and post-move angle snap.
• fix(exp): experience bar overflows #316 — Exp bar overflow (saturating arithmetic, __int64 widening).
• fix(inventory): prevent phantom/ghost items in inventory #317 — Inventory resync on pickup (phantom items).
• fix(tooltips): significantly improve rendering performance #321 — Ground-item tooltip perf (GroundItemLabelDescriptor + budget).
  Port to SDL3 in 1eeac9aa: _snwprintf_s → mu_swprintf_s, GDI/GL cache
  gated behind MU_ENABLE_SDL3, falls back to direct g_pRenderText
  rendering (perf regression acknowledged, labels still render every frame).

Other upstream fixes: defenseSuccessRate calc; lucky-set / restricted-item
constants to mu_define.h; VectorScale double-scaling removal in
ZzzEffectJoint.

Post-merge fixups:

• 1cceb074 — _struct.h → mu_struct.h include rename in the
  upstream-added IInventoryActionContext.h.
• 1eeac9aa — SDL3 ground-item-label cache port (see fix(tooltips): significantly improve rendering performance #321 above).
• 1de9b2a5 — Restore 9 per-tier tooltip text colors (blue = excellent,
  red = ancient, yellow = set, etc.) — casualty of the glColor* sweep in
  Story 7-9-6.
• e81ac588 — RenderBar tints (pet life, siege castle timer, hunting
  progress) + inventory durability warnings (yellow/orange/red slot tints)
  — same root cause as 1de9b2a5.
• cb96217c — PR review fixes: UTF-8 locale facet in IniFile.h;
  CONFIGURE_DEPENDS on shader-blob glob.
• 12109089 — Repo hygiene: untrack build_validation/ (8 CMake
  artifacts), 2 .DS_Store files, Main.sln.DotSettings.user, and 3
  .paw/metrics/*.jsonl workflow-telemetry files; extend .gitignore
  to cover build_validation/ and .paw/metrics/.
• ec483c50 — Add docs/upstream-merge-features.md smoke-test guide
  (referenced below).

Review feedback addressed (cb96217c)

From the gemini-code-assist review:

• IniFile.h: std::locale("") replaced with fixed
  std::codecvt_utf8<wchar_t> facet via a scoped helper. config.ini now
  byte-identical across Windows/Linux/macOS.
• CMakeLists.txt:131 shader-blob glob: CONFIGURE_DEPENDS added.

Intentionally deferred (documented in PR thread):

• Rule-of-Three = delete on legacy CList / CDimension — pre-existing
  tech debt; proper fix is std::unique_ptr / std::list, out of scope.
• include_directories → target_include_directories at
  src/CMakeLists.txt:277 — cosmetic lint, no functional impact.

Smoke-test guide

docs/upstream-merge-features.md — 15-minute play-through exercising
each upstream PR merged during this branch's life.

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request executes a major architectural refactoring (Phase -1) of the MU Online client source code. The primary objective is to transition from a legacy, monolithic, Windows-only build to a modular, cross-platform architecture. By reorganizing the source into logical modules and hardening the build system with CMake, the project is now prepared for the upcoming SDL3 migration. The changes also introduce modern C++20 standards, unified logging via spdlog, and robust cross-platform build tooling, ensuring consistent development environments across Windows, macOS, and Linux.

Highlights

• Modular Source Reorganization: Moved 692 source files from a flat directory structure into 20 logical module directories, significantly improving project maintainability and build scalability.
• CMake Build System Refactoring: Split the monolithic build into 7 independent static libraries and 1 coupled MUGame target, facilitating modular compilation and future SDL3 migration.
• Cross-Platform Tooling: Introduced CMake presets for Windows (MSVC), macOS (arm64), and Linux (x64), along with new quality gates (clang-format, cppcheck, clang-tidy) and git hooks.
• Renderer & Audio Modernization: Adopted GLM for matrix math and SDL_ttf for GPU-accelerated font rendering, while removing legacy DirectSound in favor of a cross-platform audio abstraction.
• .NET Native AOT Integration: Implemented cross-platform .NET 10 Native AOT library integration, enabling server communication on non-Windows platforms.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Ignored Files

• Ignored by pattern: .github/workflows/** (4)
  • .github/workflows/ci.yml
  • .github/workflows/mingw-build-dev.yml
  • .github/workflows/mingw-build-pr.yml
  • .github/workflows/mingw-build.yml

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request implements a comprehensive reorganization and cross-platform port of the MU Online client, introducing a modular CMake build system, HLSL shader compilation, and a unified logging facade. It replaces legacy DirectSound code with a miniaudio backend and refactors the codebase into distinct static libraries to enforce dependency boundaries. Feedback highlights the need to explicitly disable copy operations for classes managing raw pointers to prevent memory corruption, the importance of using fixed encodings for configuration files, and the recommendation to avoid CMake globbing and global include directories in favor of explicit file listing and target-based configuration.

[gemini-code-assist]

The CList class manages raw pointers (m_pHead, m_pTail) but lacks a copy constructor and assignment operator. This violates the Rule of Three and will lead to double-free errors or memory corruption if a CList instance is copied. Since this is legacy code being reorganized, you should at least explicitly delete the copy operations to prevent accidental usage.

// cppcheck-suppress [noCopyConstructor, noOperatorEq]
template <class T> CList<T>::CList() : m_pHead(new CNode<T>()), m_pTail(new CNode<T>())
{
    m_pHead->SetNext(m_pTail);
    m_pTail->SetPrev(m_pHead);
}

// Explicitly disable copying to satisfy Rule of Three and prevent memory corruption
CList(const CList&) = delete;
CList& operator=(const CList&) = delete;

[gemini-code-assist]

Similar to CList, the CDimension class manages a raw pointer m_pData but does not implement copy control. Copying an instance will result in two objects sharing the same pointer, leading to a double-free in the destructor. Please delete the copy constructor and assignment operator.

// cppcheck-suppress [noCopyConstructor, noOperatorEq]
m_pData = new T[m_nSize];
}

CDimension(const CDimension&) = delete;
CDimension& operator=(const CDimension&) = delete;

[gemini-code-assist]

Using std::locale("") relies on the environment's default locale, which can vary across systems and lead to inconsistent encoding behavior when reading or writing config files. For a cross-platform game client, it is safer to explicitly use a UTF-8 locale or a fixed encoding to ensure config.ini remains portable between Windows, Linux, and macOS.

[gemini-code-assist]

Using the global include_directories command is generally discouraged in modern CMake. It affects all targets defined in this directory and subdirectories, which can lead to header pollution and unexpected conflicts. It is better to use target_include_directories on specific targets or the MUCommon interface library.

target_include_directories(MUCommon INTERFACE "${sdl3_SOURCE_DIR}/include")

[gemini-code-assist]

Using file(GLOB_RECURSE) to collect source files is discouraged by CMake best practices. CMake cannot automatically detect when new files are added to the filesystem unless a re-configure is manually triggered. Listing source files explicitly is preferred for reliable build dependency tracking.

[gemini-code-assist]

The use of file(GLOB) here to find pre-compiled blobs has the same issue as source file globbing: CMake won't know if the contents of the compiled/ directory change unless the project is re-configured. Since these are build artifacts, consider using a more deterministic approach or explicitly listing the expected blobs.

[Mosch0512]

i will test it and look at the code in detail this weekend :D

[Mosch0512]

Review

Branch reviewed: cross-platform-sdl-migration-merged
Build target: MinGW-w64 i686 cross-compile from WSL/Linux (in-tree cmake/toolchains/mingw-w64-i686.cmake); also reproduced on MSVC + vcpkg via CLion.

Verdict:

The branch is too large to land as a single PR, and on the configurations its CI doesn't cover (MinGW i686 cross-compile, MSVC + vcpkg from CLion / Visual Studio) it currently fails to configure or build. Both problems are fixable, but the size problem is the gating one - until the branch is sliced up, the technical issues below are also effectively unreviewable.

---

This branch needs to be split before it can merge

cross-platform-sdl-migration-merged currently sits at ~1,188 files changed, +68k / -223k lines vs main. That's effectively unreviewable as a single PR, blocks every other contributor's branch from rebasing cleanly (we hit ~46 conflicts in just the first commit during a test rebase), and means any mistake won't surface until the whole thing lands. No amount of fixing the technical issues below changes the fact that the change-set itself is the biggest blocker.

The good news: the work is already factored into clear, mostly-independent stories (the Story X.X.X tags in commit messages and code comments). They map naturally onto a sequence of small PRs that can each go to main and be released incrementally.

Phase	PR scope	Approx. story refs	Land independently?
1	Repo restructure only - move files into Core/, World/, Gameplay/, UI/Framework/, UI/Legacy/, RenderFX/, Platform/, Network/, Data/, Scenes/. Pure git mvs + matching #include path updates. No behavior change.	7.0.x	yes - touches paths only
2	CMake target split - break the monolithic Main into MUCommon, MUCore, MUProtocol, MUData, MURenderFX, MUAudio, MUPlatform, MUThirdParty, MUGame, Main. Library boundaries surface coupling problems early.	4.x.x	yes - only touches CMakeLists.txt and minimal includes
3	PlatformCompat shim + cross-platform types - mu_base_types.h, mu_define.h, mu_types.h, PlatformTypes.h, PlatformCompat.h, mu_get_app_dir, mu_narrow_path, mu_swprintf, mu_wcstok. Lets non-Windows toolchains compile something.	7.6.x	yes - can land before SDL3 even appears
4	Logging migration - introduce mu::log (spdlog wrapper), then convert CErrorReport and CmuConsoleDebug call sites in batches (Network/, Scenes/, Gameplay/, GameShop/). Retire the old globals once empty.	7.10.x	yes - can run in parallel with later phases as long as both APIs coexist
5	CMake build hygiene - toolchain files (linux-x64, macos-arm64, mingw-w64-i686), warning flag gating per compiler (issue 6), preset cleanup (issue 9), clang-format/clang-tidy/cppcheck configs, dev Makefile.	7.6.x	yes - pure tooling
6	MuRenderer abstraction - introduce the interface and ship the existing OpenGL backend behind it as MuRendererGL. Game code keeps working unchanged. This is the keystone phase: nothing else can land until this exists.	4.2.x, 7.9.x	yes - backwards-compatible by design
7	Audio migration to miniaudio - MUAudio library + miniaudio-backed wzAudio shim.	7.6.6 + audio stories	yes - independent of rendering
8	CURL/OpenSSL cross-platform shop/credentials - replace Win32 wininet/wincrypt with libcurl + OpenSSL. Make CURL optional with a header-only fallback (issue 8) so MinGW / un-vcpkg'd MSVC contributors aren't blocked.	3.4.x, 7.6.x	yes - independent
9	SDL3 platform layer - SDLWindow, SDLEventLoop, SDLKeyboardState, input shims, text input. Window/input only; rendering still routes through MuRendererGL.	2.2.x, 7.x.x	depends on phase 3 (PlatformCompat) and phase 6 (renderer abstraction)
10	SDL3 GPU backend (MuRendererSDLGpu) - implements the renderer interface introduced in phase 6 against SDL_gpu. Add the HLSL→SPIR-V→MSL shader pipeline. Until this lands, both backends ship and a CMake flag picks between them.	4.x, 7.9.x	depends on phase 6
11	Remove the legacy OpenGL backend - delete MuRendererGL, drop GLEW, retire the gl* immediate-mode calls (issue 4). Only after phase 10 is proven on all three platforms.	7.9.3	depends on phase 10
12	Tests + CI matrix - tests/ tree, GitHub Actions matrix for Linux x64 / macOS arm64 / MinGW i686 / MSVC.	various	yes - bolt-on

Why this is worth the extra work upfront:

1. Each phase is reviewable. Phase 1 is a git mv audit; phase 6 is a single new interface; phase 10 is the hard rendering work in isolation. Reviewers can actually catch bugs.
2. Each phase is releasable. main keeps shipping after every merge. Players don't wait six months for the next playable build.
3. Other contributors' branches survive. Right now any branch (camera, UI, gameplay) cherry-picked off main becomes nearly unrebaseable against this branch. With incremental landing, a contributor only needs to react to one change at a time.
4. Bisecting becomes possible again. When a regression appears, you bisect against a sequence of small commits, not a single 290k-line megacommit.
5. The technical issues below shrink. Most of them are detected automatically once phase 5 (CMake hygiene + cross toolchains) and phase 12 (CI matrix) are in. They never accumulate to "150+ errors at once."

The core insight: the abstraction phases (3, 4, 6) and the implementation phases (9, 10, 11) are decoupled by design. There is no technical reason they all have to land together - only the inertia of the existing branch keeps them coupled.

---

TL;DR - what's broken under the hood

Even setting the size problem aside, the branch builds on the platforms its CI exercises (Linux/macOS clang, Windows MinGW64) but several gaps prevent any other configuration from producing a binary:

1. Logging migration is incomplete - g_ConsoleDebug, g_ErrorReport, MCD_* constants are still referenced in many files after Story 7.10.1 replaced them with mu::log::Get(...).
2. SDL3 helpers are declared inside the wrong #ifdef branch in Platform/PlatformCompat.h, so they're invisible to MinGW (_WIN32-defined) builds.
3. POSIX-only types (u_char, u_int64) and headers (<sys/utsname.h>) are used in unconditionally compiled code paths.
4. Several files still call OpenGL immediate-mode functions (glColor3f, glPushMatrix, glBegin/glEnd, …) after the SDL3 GPU backend made them unavailable (Story 7.9.3).
5. Win32 macro collision - MuPlatform::CreateWindow(...) clashes with <windows.h>'s CreateWindow → CreateWindowW macro on MinGW.
6. CMake warning relaxations are Clang-only - they trigger errors when GCC validates -Wno-error=<unknown-flag> under -Werror.
7. Stale relative #include paths that pre-date the directory restructure.
8. CURL is unconditionally REQUIRED - blocks configure on every Windows configuration that doesn't have a system curl: MinGW cross-compile, MSVC + vcpkg (vcpkg integration is silently bypassed by issue 9), and stock Visual Studio without manual prep.
9. CMakePresets.json hard-codes toolchainFile on the base Windows presets - downstream presets inheriting from windows-x64 can't supply their own toolchain file (e.g. vcpkg.cmake) without surgery, because the inherited toolchainFile field always wins over a CMAKE_TOOLCHAIN_FILE cache variable.

Each item is reproducible, narrowly scoped, and can be fixed independently - and most of them naturally fall inside one of the phases in the table above.

---

CI coverage today

Worth flagging up front: the existing .github/workflows/ci.yml covers four jobs - Quality Gates (Ubuntu, no build), Linux Native (Ubuntu + system GCC), macOS Native (Brew clang), and Windows Native (windows-latest + MSYS2 MinGW-w64 x86_64, not MSVC, despite the CMakePresets.json windows-base description saying "MSVC"). The toolchain files (toolchain-x64.cmake, toolchain-x86.cmake) don't actually force MSVC - they trust whatever compiler is on PATH, and the CI installs MinGW.

Implications:

• MSVC has zero CI coverage. Anyone building from Visual Studio is the first to hit MSVC-specific issues. Recommend either adding an MSVC job, or removing the misleading "MSVC" wording from the preset description.
• The MinGW issues this review documents are reproduced on MinGW-w64 i686. Some - u_char / u_int64, <sys/utsname.h> guarding, CreateWindow macro collision, mu_wcstok 2-vs-3-arg shim - should also surface on the CI's MinGW64 path. If the existing Windows job is currently green, it's worth re-checking the latest run logs to see whether they're truly clean or compiling with warnings that should be errors. Issues 1 (logging migration) and 4 (lingering gl* calls) are language-level - they will fail any compiler.
• CLion + MSVC + vcpkg cannot configure the project today. Tested on Windows: opening the repo in CLion, picking the bundled windows-x64 (or any inheriting) preset under the Visual Studio toolchain, and pointing CMake at C:/vcpkg/scripts/buildsystems/vcpkg.cmake - configure still fails on find_package(CURL REQUIRED) because the preset's hardcoded toolchainFile overrides the user-supplied vcpkg one. Same outcome for the auto-generated cmake-build-x86debug-mueditor profile CLion creates from windows-x86-mueditor. Concretely: there is no out-of-the-box configuration in which a Windows developer using MSVC + vcpkg can produce a build.

---

Reproduction

# From repo root:
cmake -S . -B build-mingw \
  -G Ninja \
  -DCMAKE_TOOLCHAIN_FILE=cmake/toolchains/mingw-w64-i686.cmake \
  -DCMAKE_BUILD_TYPE=Release \
  -DMU_ENABLE_SHADER_COMPILATION=OFF \
  -DMU_ENABLE_DOTNET=OFF
cmake --build build-mingw -j 8

Configure currently fails on find_package(CURL REQUIRED) (item 8). With CURL stubbed locally, configure succeeds and the build runs through MUCore/MUPlatform/MURenderFX/MUData (after the other fixes) before halting in MUGame on the logging-migration and lingering-GL items.

---

Issues

1. Incomplete logging migration (HIGH - blocks compilation)

Story 7.10.1 replaced CErrorReport and CmuConsoleDebug with mu::log::Get(...), but the call sites in many files still reach for the old API. Symbols not declared anywhere on the branch:

• g_ConsoleDebug
• g_ErrorReport
• MCD_NORMAL, MCD_RECEIVE, MCD_ERROR
• header Utilities/Log/muConsoleDebug.h

Affected (non-exhaustive - grep g_ConsoleDebug\|g_ErrorReport\|MCD_ to enumerate):

• Network/WSclient.cpp - at least lines 350, 354, 358, 467, 480, 482, 490, 534
• Scenes/MainScene.cpp - extern CErrorReport g_ErrorReport; declared but never defined
• Scenes/CharacterScene.cpp - same extern, plus #include "Utilities/Log/muConsoleDebug.h"
• Scenes/SceneManager.cpp - g_ErrorReport.Write(...) (~line 603-604)
• GameShop/NewUIInGameShop.cpp - g_ConsoleDebug.Write(...) (~lines 615, 705)

Fix: finish the migration to mu::log::Get("module")->info(...) / warn(...) / error(...). The pattern is already correctly applied in Gameplay/Items/PersonalShopTitleImp.cpp - same shape works for the rest.

2. SDL3 helpers in wrong preprocessor branch (HIGH)

Platform/PlatformCompat.h is structured as #ifdef _WIN32 ... #else ... #endif (lines 3 → 50 → 2285). The SDL3 helpers are declared inside the #else branch (#ifdef MU_ENABLE_SDL3 ... #endif nested at ~line 924-960), so they're only visible to non-Windows builds.

MinGW defines _WIN32, so the following declarations are unreachable from MinGW translation units that include PlatformCompat.h:

• extern char g_szSDLTextInput[32];
• extern bool g_bSDLTextInputReady;
• void MuStartTextInput(); / void MuStopTextInput();
• inline wchar_t MuSdlUtf8NextChar(const char*&);

Resulting failures:

• Platform/sdl3/SDLKeyboardState.cpp - SDL_Window, SDL_StartTextInput, SDL_StopTextInput not declared (file includes PlatformCompat.h expecting it to bring in <SDL3/SDL.h>).
• Platform/sdl3/SDLEventLoop.cpp - g_sdl3KeyboardState not declared.
• ThirdParty/UIControls.cpp - MuStopTextInput, g_bSDLTextInputReady, g_szSDLTextInput, MuSdlUtf8NextChar not declared.

Fix: lift the SDL3 helper block out of the platform #ifdef and gate it only on MU_ENABLE_SDL3. Same root cause applies to mu_narrow_path (used by Data/GlobalText.h's template - GCC's two-phase name lookup needs the declaration visible regardless of platform).

3. POSIX-only types in unconditional code (MEDIUM)

• RenderFX/ZzzBMD.cpp:1022 - static_cast<u_char>(this->StreamMesh). u_char is from <sys/types.h> (POSIX), not on MinGW. Fix: unsigned char.
• Network/WSclient.cpp:540-550 - u_int64 received; ... u_int64 actual;. Same problem. Fix: uint64_t.
• Core/MuSystemInfo.cpp:13 - #include <sys/utsname.h> is unconditional, and uname() is called in the unconditional path of MuGetSystemInfo(). Fix: wrap both in #ifndef _WIN32, fall back to a Windows-shaped string from RtlGetVersion / compile-time identifier.

4. Lingering OpenGL immediate-mode calls (MEDIUM)

Story 7.9.3 removed the OpenGL backend and GLEW headers, but several files still emit immediate-mode draw calls. Those symbols don't exist anywhere in the build:

• RenderFX/ZzzEffectFireLeave.cpp:547,572,573,598 - glColor3f, glPushMatrix, glTranslatef, glPopMatrix.
• Gameplay/Characters/ZzzObject.cpp - glColor3f/glColor4f/glColor3fv at ~25 sites between lines 786-7840.
• GameShop/NewUIInGameShop.cpp:85,312-337 - glColor4f, glMatrixMode, glPushMatrix, glLoadIdentity, glClear, glPopMatrix.

Fix: route through the renderer abstraction. mu::GetRenderer() already exposes SetMatrixMode, PushMatrix, LoadIdentity, MultMatrix, Translate, Rotate, ClearDepthBuffer, SetViewport. A small extension may be needed for color (e.g. SetVertexColor(r,g,b,a)) and for legacy glBegin(GL_QUADS)-style geometry, since there's no debug-primitive helper on the abstraction yet.

5. Win32 macro collision in MuPlatform.h (MEDIUM)

Platform/MuPlatform.h:16 declares static bool MuPlatform::CreateWindow(...). On MinGW with <windows.h> included earlier, CreateWindow is #defined to CreateWindowW / CreateWindowA, so the class declaration becomes garbage:

error: macro "CreateWindowW" requires 11 arguments, but only 4 given

Fix: either #undef CreateWindow (and CreateWindowA/W for paranoia) at the top of MuPlatform.h, or rename the method to something like CreateGameWindow / CreateMainWindow.

6. CMake warning relaxations assume Clang (MEDIUM)

MUCommon's INTERFACE compile options propagate -Werror -Wall -Wextra plus a list of -Wno-error=<name> exceptions. Many entries in the MU_LEGACY_WARNING_RELAXATIONS list are Clang-only flags:

-Wno-error=deprecated-anon-enum-enum-conversion
-Wno-error=deprecated-enum-compare
-Wno-error=null-conversion
-Wno-error=null-arithmetic
-Wno-error=logical-op-parentheses
-Wno-error=reorder-ctor
-Wno-error=pointer-bool-conversion
-Wno-error=tautological-constant-out-of-range-compare
-Wno-error=tautological-overlap-compare
-Wno-error=delete-non-abstract-non-virtual-dtor
-Wno-error=pragma-once-outside-header
-Wno-error=bitwise-op-parentheses
-Wno-error=sometimes-uninitialized
-Wno-error=extern-initializer
-Wno-error=tautological-pointer-compare
-Wno-error=c++11-narrowing
-Wno-error=writable-strings
-Wno-error=constant-conversion

Under -Werror, GCC treats -Wno-error=<unknown-flag> as a hard error.

Fix: wrap each Clang-only entry in $<$<CXX_COMPILER_ID:Clang,AppleClang>:...>. Also add the GCC-flavoured set that the legacy code needs:

$<$<CXX_COMPILER_ID:GNU>:-Wno-error=address>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=parentheses>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=implicit-fallthrough>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=narrowing>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=class-memaccess>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=stringop-truncation>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=format>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=format-security>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=maybe-uninitialized>
$<$<CXX_COMPILER_ID:GNU>:-Wno-error=array-bounds>

Same relaxation list needs applying to MURenderFX / MUData / MUCore (currently it's only on MUGame and Main) - the legacy Zzz* files live in those targets too.

7. Stale relative include paths (LOW)

A handful of files still use ../-prefixed includes referencing pre-restructure locations:

• Scenes/SceneManager.cpp, MainScene.cpp, CharacterScene.cpp, LoginScene.cpp, SceneCommon.cpp
• Network/WSclient.cpp - Guild/GuildCache.h, MUHelper/MuHelper.h
• UI/Legacy/ZzzInterface.cpp
• UI/Windows/HUD/NewUIMiniMap.cpp
• Gameplay/Items/PersonalShopTitleImp.cpp - Guild/UIGuildInfo.h
• RenderFX/ZzzOpenglUtil.cpp - ShellExecute used without including <shellapi.h> or declaring it for Win32.

Fix: drop the ../ and old subdirectory components - the include path already covers the new locations.

Related: Data/GlobalText.h should #include "Platform/PlatformCompat.h" directly rather than relying on transitive includes, otherwise the template Load(...) at line 162 fails GCC's two-phase name lookup for mu_narrow_path.

8. CURL is unconditionally REQUIRED (HIGH - blocks configure on every untested Windows path)

Originally scoped to MinGW, but reproduces on MSVC + vcpkg too. Configure halts at:

CMake Error at .../FindPackageHandleStandardArgs.cmake:230 (message):
  Could NOT find CURL (missing: CURL_LIBRARY CURL_INCLUDE_DIR)
Call Stack:
  src/CMakeLists.txt:364 (find_package)

Where it reproduces:

• MinGW i686 cross-compile from WSL - no apt-provided libcurl.a for the i686 mingw target, and _deps/mingw-i686/lib/ only ships libjpeg.a / libturbojpeg.a.
• MSVC + vcpkg in CLion - even with vcpkg pre-installing curl:x64-windows, the toolchain conflict in issue 9 prevents vcpkg's CMake integration from loading, so find_package(CURL) falls back to system search and finds nothing.
• Stock Visual Studio without manual curl prep - same root cause, no system curl.

OpenSSL has the same shape but already gracefully degrades - configure logs OpenSSL NOT found: mu_encrypt_blob/mu_decrypt_blob will use identity no-op fallback. CURL should follow that pattern.

Options:

1. Add prebuilt binaries to _deps/<triplet>/ for each Windows target.
2. Use FetchContent to build curl from source for the cross target.
3. Add a feature flag (MU_ENABLE_SHOP_DOWNLOADS) so the shop/banner downloader can be excluded from the build, falling back to a no-op stub when off.
4. Have CMake auto-stub CURL::libcurl (and provide a header-only stub) when CURL isn't found, with a clear STATUS message that download features will be inert at runtime - mirrors the existing OpenSSL fallback.

Option 4 is the lowest-friction path to "any Windows contributor can clone and configure" - it matches the OpenSSL precedent already in the tree.

9. CMakePresets.json hardcodes toolchainFile on the base Windows presets (HIGH - blocks vcpkg / sysroot overrides)

windows-x64 and windows-x86 set:

"toolchainFile": "${sourceDir}/toolchain-x64.cmake"

toolchainFile in CMake presets is a top-level field and takes precedence over a CMAKE_TOOLCHAIN_FILE cache variable in the same or any inheriting preset. So a downstream preset like:

{
  "name": "msvc-x64-vcpkg",
  "inherits": "windows-x64",
  "cacheVariables": {
    "CMAKE_TOOLCHAIN_FILE": "C:/vcpkg/scripts/buildsystems/vcpkg.cmake"
  }
}

…silently keeps using toolchain-x64.cmake and ignores the vcpkg integration entirely. That's how issue 8 manifests on MSVC: vcpkg installs curl, but its cmake-side hooks never load, so find_package(CURL) can't see the install.

The two existing toolchain files (toolchain-x86.cmake, toolchain-x64.cmake) don't actually do anything - they just set CMAKE_GENERATOR_PLATFORM (which is a no-op under Ninja Multi-Config, the configured generator). Removing them or making them opt-in would unblock vcpkg / sysroot users without changing CI behaviour.

Fix (any of):

1. Move toolchainFile off the base presets and into an explicit windows-x64-default-toolchain leaf preset. Other leaves (windows-x64-vcpkg, windows-x64-mingw) override freely.
2. Delete the toolchain files entirely - they only set CMAKE_GENERATOR_PLATFORM, which the Ninja Multi-Config generator ignores.
3. Have the toolchain file include() $ENV{CMAKE_TOOLCHAIN_FILE_OVERRIDE} first so an env var can chain in vcpkg / a sysroot file. (Ugly but non-breaking.)

Option 1 or 2 is the right answer - option 3 just papers over the problem.

---

What works

To set expectations: after the configure-time CURL workaround, the following targets currently build clean on MinGW with the changes from this review applied:

• MUCommon, MUCore, MUProtocol, MUData, MURenderFX, MUAudio, MUPlatform, MUThirdParty, MUImgui
• All FetchContent dependencies (SDL3, SDL3_ttf, glm, spdlog)

The remaining build errors are concentrated in MUGame, almost entirely from issues 1 (logging) and 4 (lingering GL calls).

[Mosch0512]

Closing this PR.

As detailed in the review above, the change-set is too large to land or review as a single unit (~1,188 files, -68k / +223k LOC vs main), and the technical issues are effectively
unreviewable until the branch is sliced up. The work itself is valuable and already factored into clear, mostly-independent stories - it just needs to land incrementally.

Suggested path forward: re-open the work as the sequence of smaller PRs outlined in the Review (phases 1–12),
each targeting main and releasable on its own. The abstraction phases (repo restructure, CMake target split, PlatformCompat, logging, MuRenderer interface) are decoupled from
the implementation phases (SDL3 platform layer, SDL_GPU backend, legacy GL removal) by design — there's no technical reason they have to land together.

Benefits of the incremental approach:

• Each PR is reviewable and bisectable.
• main keeps shipping playable builds between merges.
• Other contributors' branches stay rebaseable.
• The cross-platform/CI gaps (CURL, vcpkg toolchain, MinGW logging migration, lingering gl* calls, etc.) get caught one phase at a time instead of all at once.

Happy to review the smaller follow-up PRs as they come in. Thanks for the huge amount of work here - this isn't a rejection of the work, just of landing it as one piece.