Fix macOS build and run issues + input/simulation deadlock

[lugt]

Continuation of #1777 (Fix macos build and run issues by @Skosulor), rebased onto current master with the review feedback addressed and an additional deadlock fix that the parallel threading exposes on macOS.

What this does

macOS build/run fixes (from #1777)

1. Run the presenter on the main thread. QGuiApplication must be created and operated on the main thread. On macOS 26 (Tahoe) the platform plugin asserts in -[NSMenu _setMenuName:] ("API misuse: setting the main menu on a non-main thread") otherwise. The simulation is moved to a worker thread instead.
2. Remove the #ifndef __APPLE__ guard that stopped Qt events from being processed on macOS (gui_application_impl.cpp).
3. Fix shaders that crashed glLinkProgram on macOS. The macOS OpenGL driver crashes (SIGSEGV in glpLLVMCGSwitchStatement during glLinkProgram) when compiling a switch statement in world2d.frag.glsl — switched to if/else. Also upgraded the GUI shaders from GLSL 1.2 (#version 120, attribute/varying) to 3.3 core (in/out, layout(location=)), and renamed the texture uniform to tex to avoid the GLSL reserved keyword.

Deadlock fix (new, exposed by parallel sim/presenter)

Once the simulation runs in a worker thread and the presenter on the main thread, mouse interaction deadlocks within a few seconds:

• Main thread (presenter): Controller::process holds Controller::mutex, then calls bind.transform() → EventLoop::create_event (takes EventLoop::mutex). Lock order: Controller → EventLoop.
• Worker thread (simulation): EventLoop::reach_time holds EventLoop::mutex, executes DragSelectHandler::invoke → select_cb → Controller::set_selected (takes Controller::mutex). Lock order: EventLoop → Controller.

Classic AB-BA deadlock (confirmed via lldb attach + thread backtrace all, both threads in __psynch_mutexwait).

Fix: narrow Controller::process's lock scope to only protect the outqueue operations; bind.transform() now runs without holding Controller::mutex. Both threads then take the locks in the order EventLoop → Controller, eliminating the cycle. The bindings' callbacks (get_controlled, get_selected, reset_drag_select, …) already take Controller::mutex themselves and are safe to call unlocked because it is a recursive_mutex.

This is the same class of issue @Skosulor reported in #1777 (mutex lock failed: Invalid argument), just triggered through the drag-select path.

Review feedback (@heinezen) addressed

• Engine: thread functions (time loop, simulation) extracted into run_time_loop() / run_simulation(); all thread spawning moved into Engine::loop(); comments explain why the main thread runs the presenter in FULL mode (Qt requirement) and the simulation in headless mode. Thread-count log is now printed after spawning (so the number is correct).
• window_settings moved from Engine into Presenter (stored as a member, passed via the Presenter constructor); Presenter::run() and init_graphics() no longer take it as a parameter.
• gui.cpp: removed the static constexpr for the texture uniform name; the name is passed inline (now \"tex\").
• maptexture.frag.glsl: uniform renamed texture_ → tex.

Two suggestions intentionally not applied — would like your input

1. identity.vert.glsl: kept gl_Position = vertex_position;. In GLSL 330 core, gl_Position does not automatically contain the current vertex position (it's a write-only output with undefined default); removing it would collapse the GUI quad to the origin (0,0,0,1). This may be a confusion with compatibility-profile gl_Vertex. Could you double-check @heinezen?
2. world2d.frag.glsl / demo_2_obj / demo_4_obj: kept if/else instead of switch, with a comment explaining the macOS driver crash (glpLLVMCGSwitchStatement SIGSEGV during glLinkProgram). This is a driver bug, not a shader-authoring issue — confirmed by the crash stack. A switch reliably crashes on macOS even though it's valid GLSL.

Verification

Built and run on macOS 26.3 (Apple Silicon), Homebrew Qt 6.11.1 + LLVM clang 22.1.8, Python 3.14. Tested:

• Game window opens, terrain renders, all render stages (Skybox/Terrain/World/HUD/Screen) initialize.
• Mouse interaction (drag-select, click) no longer deadlocks; window closes cleanly with all threads exiting (Draw loop exited → Game simulation stopped → Time loop stopped → presenter exited).
• renderer_stresstest 0 runs end-to-end (shaders link, entities spawn, FPS stable), confirming the shader fixes.

Supersedes #1777. Happy to split the deadlock fix into a separate PR if preferred.

🤖 Generated with Claude Code

[lugt]

@heinezen I've just continued the work left by #1777, and done some fix so that it could actually build on macOS 26 Tahoe, could you help to review and merge this, and I am happy to further change the code if needed.