On Vulkan, the Windows default, a presentation stall longer than a second crashed the client. wgpu-hal 27.0.4's `acquire_texture` maps a `VK_TIMEOUT` from its post-acquire fence wait to `DeviceError::Unexpected`. wgpu-core treats that as a lost device, and Bevy then panics on "Couldn't get swap chain texture". Upstream tracks this as gfx-rs/wgpu#9029, still open in wgpu 30.
wgpu-hal 27.0.4 is now vendored under `third_party/wgpu-hal` through `[patch.crates-io]`, as `kira` is. Only `src/vulkan/instance.rs` differs. A timeout of the previous submission's wait, of the acquire, or of the acquired image's fence now waits again within a 30 s budget, then fails as upstream does. The Vulkan spec leaves the semaphore and fence unaffected by a timed-out call, so the retry is legal. Each retry logs, with the raw result. `BENILLA.md` says how to check the patch and when to drop it.
Tests: the lock resolves to the patch. Built for Linux and Windows.
Fixes#2373🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On a Windows checkout every text file but Rust took CRLF, since `.gitattributes` pinned LF for `*.rs` alone. A test that matches a shader's source on `\n` (`a_fade_twin_blends_by_the_stage_zero_alpha_op`) failed there, and any `include_str!` of a shader, Lua, XML or TOML file read different bytes than on macOS and Linux.
`* text=auto eol=lf` now pins LF for every text file, keeping the two byte-for-byte exceptions (the vendored benilla-re tables and kira). Every file in the index is LF already, so a renormalize changes nothing stored; only what a Windows checkout writes changes.
Fixes#1998🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Three tests failed on a Windows checkout:
- one split a source path on `/`;
- one split its own `include_str!` on `"\n}\n"`, which a CRLF checkout lacks;
- one captured a traceback's addon folder with a greedy `.*`.
Each now normalises the separator or line end, or bounds the capture. A sweep found the same path forms in four more source-walking tests (`fire_scan`, `world_api_wall`, `reference_strings` and `front_end`), now normalised too.
`.gitattributes` pins Rust sources to LF and leaves `third_party/kira`'s CRLF files as they are.
Built and tested on macOS. The Windows cases are reasoned from the path and line-ending forms, not run.
Fixes#798🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`reference/` held benilla's own copies of the reference's tables, some derived by its own scripts from older inputs: the fire-site table kept one event id where a site fires two, and credited glue-screen sites to in-game events.
Now `reference/` is benilla-re's tables at the commit `reference/PIN` names, byte for byte: bindings, events, glue events, messages, event arguments, verb events, the glue surface, the Lua globals, chat types and groups, and CVars. `scripts/sync-reference.py` copies them, and its `--check` runs in the gates and in CI. Four generators and nine old files go, and every reader moves to the new columns.
New tests hold the chat type, chat group and CVar tables to the copies; the CVar rows already filed are listed as known.
Fixes#245Fixes#960
Part of #244🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Sam Sung