build-macos: published b1/b2 corrupt Genshin's asset reads (toolchain); configure against an installed engine

[pharmacykitty]

What

• build-macos/README.md: records that the published b1 and b2 engines corrupt the data Genshin reads during scene loads, the evidence that it is the build and not the source, and how to configure against an installed engine.
• build-macos/configure-from-engine.sh: configures a tree with Homebrew's headers and an installed engine's own x86_64 dylibs, sonames pinned to what the engine's unix halves dlopen.

Symptom

During a teleport or the login scene: HK4EUpload: error asr_003: BlobSignature not match [...blocks/00/<blk>@<offset>] by the dozen (spanning ~600 KB of one block), or Unity's The file 'CAB-…' is corrupted! / Position out of bounds / Mismatched serialization, then an endless loading screen or c0000005 in GenshinImpact.exe (once Trying to allocate 18446744073676197384B, a negative length from the garbage).

Ruled out, each by test

Game files (every failing .blk matches the server md5), Wine's file I/O (14,700 overlapped + unbuffered + mapped reads per block, byte-compared, clean on b1 and on Yaagl's wine-11.0), AVX advertising, the fps unlocker, the presenter and upscaler, the GPU identity, D3DMetal vs DXMT (both corrupt on b2), the launcher bundle, memory (RSS flat 4–6 GB for hours), msync lost wakes (sweep = 0).

The matrix

engine	path	result
published b1	Sevoflurane, D3DMetal 4.0b2	corrupt, 2–3 loads
published b1	wine steam.exe GenshinImpact.exe script, DXMT	clean 5 h 16, 96 loads
published b2	Sevoflurane, DXMT	corrupt, load 1
published b2	script, DXMT	corrupt ×2, load 1–2
Yaagl wine-11.0 + DXMT 0.80	Yaagl script	clean 1 h 45, 20 loads
same b2 commit, built locally	script, DXMT	clean 2 h 10, 21 loads

Local build: M4 Max, Xcode 27.0 27A5194q (Apple clang 21.0.0, ld-1328.2), configure-from-engine.sh recipe (x86_64 only, SDL on, no gstreamer), installed over an APFS clone of the published b2 keeping its arm64 wineserver, loader, DXMT/D3DMetal payloads, winegstreamer.so and winemetal.so.

Published x86_64-unix/*.so: LC_BUILD_VERSION sdk 27.0, tool ld 27037.1. Local: sdk 27.0, ld 1328.2. Both chained fixups; neither contains AVX code. So: toolchain, a flag that differs from the README recipe, or the i386+x86_64 build. A rebuild of b2 with Xcode 27.0's linker, or x86_64-only, checked through a few Genshin teleports, would tell which.

Also seen

• One hang on the local build (not corruption): presents stop, all game threads parked, one GCD worker spinning in Rosetta runtime frames with audio HAL frames nearby. Samples kept.
• Per-game renderer dxmt through WINEDLLPATH_PREPEND died at startup on the 1.0/1.1-beta app with b1 (c0000005 in kernel32 ~12 s); b2's loader change presumably covers it, untested here.

🤖 Generated with Claude Code