[bensaadmucret]

What and why

Two misreads of an SDL_GPU main's files, both found bringing up MARVEL Cosmic Invasion (010059D020C26000 v0, build D6A37DEAFF3F5AEA, FNA/SDL_GPU):

• graphics: a main that imports only vkGetInstanceProcAddr — SDL's Vulkan loader, which an SDL game links for WSI and may never use to draw (Balatro renders with OpenGL, Absolum with NVN) — was read as OpenGL when it also imports EGL and names glShaderSource, which is SDL's unused GL backend. vkGetDeviceProcAddr does not tell the renderer either: the WSI path names it to load the swapchain procs. A vkCmd* name can only come from a table a renderer resolves command-recording procs through — MARVEL's main names 30 of them. vk_names counts them; the loader import plus any of those is graphics=vulkan, with its own survey reason. A main whose own nvnBootstrapLoader import wins as before.
• gl_src: MARVEL's main holds four bare #version headers (MojoShader prefixes the GLSL it regenerates at run time) which counted as embedded GLSL sources, so the build translated four empty shaders and failed in metallib. A #version string now counts only with a body behind it (>64 bytes to its NUL) — still looser than glsl_sources' void main rule, on purpose.

Adds the finding to docs/reference/packing.md (A Vulkan game read as OpenGL through SDL's loader).

Checked

• Tests: python3 -m unittest dm.tools.tests.test_profile_guess — 37 tests, including the new SdlVulkan class: loader alone with GL names → gl (Balatro's shape), loader only and no GL name → nvn (Absolum's shape), own nvnBootstrapLoader beating the SDL signal, vk_capture set for an SDL_GPU game without SPIR-V files, and a bare #version header not counting as embedded GLSL. test_docs passes (6 tests).
• Games run, with title ID and version: MARVEL Cosmic Invasion 010059D020C26000 v0: the provisional profile now comes out graphics=vulkan + vk_capture, and with the sibling PRs the game boots and draws (intro video, 60 fps). Before, the profile said gl and the build captured nothing.
• Devices and OS: Apple M6, macOS 27.0
• regress.py impact names profile / every build for this change. The gl_src tightening can only lower that counter (bare headers stop counting); the vk_names path triggers only when a main imports the loader alone and names vkCmd* procs — a shape the supported SDL games do not have (Balatro: OpenGL, names no vkCmd*; Absolum: NVN, own nvnBootstrapLoader). regress.py check --games was not run locally; flagging it honestly for CI.
• One of four PRs from the same bring-up (the others: EOS platform handle, vk_shaders entry point, vkCmdBlitImage + SDL_GPU device procs).

Docs

• The finding is in its game note or its docs/reference/<subsystem>.md entry, and python3 dm/tools/docs_symptoms.py was run (the symptom index)
• A chain tool changed: dmsh_chain.py --header was run and dm/gfx/nvn/dmsh_chain.h is part of the change
• No game files, keys or data extracted from a game in the change or the attachments