[draft] Combined: bgfx MSAA+mipmaps fix + Canvas FB-pool survivability (CI smoke)

[bkaradzic-microsoft]

Note

Draft PR for CI smoke-test only. Combines the two weekend feature branches into one tree to exercise all platform builds at once. Do not review or land from here — splits below.

This combined branch contains two independent fixes:

1. MSAA + mipmaps fails on D3D11/D3D12/Vulkan (bgfx) #1714 — bgfx MSAA + mipmaps fix (bn-fix-msaa-mips on BabylonJS/bgfx, picked up via bgfx.cmake bump). Sets MipLevels=1 on the MSAA target texture in all four backends (D3D11, D3D12, Vulkan, Metal); the resolve target still gets m_numMips. Without this, MSAA-enabled render targets that also request mipmaps fail validation (E_INVALIDARG on D3D11/12, VUID on Vulkan, MTL validation on Metal). Three-level pin chain:

   • BabylonJS/bgfx branch bn-fix-msaa-mips @ b617ec8f4
   • BabylonJS/bgfx.cmake branch bn-bump-bgfx-msaa-mips @ 2bd6f6214 (advances bgfx submodule)
   • This branch advances CMakeLists.txt GIT_TAG for bgfx.cmake to 2bd6f6214

2. Canvas framebuffer-pool survivability fix (cascade root cause). Many playground tests were marked excludeFromAutomaticTesting with reasons like "Suspected to cause Win32 D3D11 cascade crash". Empirically: the cascade is V8 GC pacing outpacing bgfx's default 128-slot framebuffer pool; whichever test ran when the pool finally exhausted got blamed, varying from run to run, so dozens of innocent tests accumulated the exclusion.

   This branch:

   • Bumps default BGFX_CONFIG_MAX_FRAME_BUFFERS from 128 → 512 in Dependencies/CMakeLists.txt and propagates CMake -D overrides to bgfx's compile defines (the existing CI -D BGFX_CONFIG_MAX_FRAME_BUFFERS=256 was previously a silent no-op).
   • Replaces three bare assert(handle.idx != bgfx::kInvalidHandle) sites with std::runtime_error throws + texture cleanup, so pool exhaustion no longer abort()s the process. The Canvas Context::Flush JS binding now catches and rethrows as Napi::Error so just the offending test fails instead of killing the whole sweep.
   • Re-enables 112 playground tests in config.json that were marked excluded as cascade victims. All 112 verified to pass when run sequentially as one batch on Win32 D3D11 Release with the FB pool bump.

CI expectations

This is the first time these two changes have been exercised on the non-Win32-D3D11 backends. Things to watch for:

• bgfx MSAA + mipmaps fix: validation layers on D3D12 / Vulkan / Metal should all be clean now where they previously errored. If any backend still errors, the fix needs further work in the respective renderer_*.cpp.
• Canvas FB-pool survivability: tests should pass without --include-excluded. The 112 re-enabled tests' new validated lines should appear in the sweep output. A test that genuinely fails on a non-Win32-D3D11 backend (real rendering bug, not cascade) will show up as a new pixel-diff or JS error in that backend's CI log.