docs(gpu): #1050 run the GPU tests locally on a software adapter

docs/gpu-pipeline.md says how to run lightcraft-gpu's tests on a machine
without a GPU: LIGHTCRAFT_GPU_ALLOW_SOFTWARE=1 admits Mesa's lavapipe on
Linux and WARP on Windows (Metal on macOS needs nothing), and
LIGHTCRAFT_SKIP_LOG shows whether anything still skipped for want of an
adapter. The comments on the variable and in equivalence.rs no longer
say CI runs them: the platform CI legs of #1148 are not landing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

crates/gpu/src/ctx.rs

@@ -758,9 +758,8 @@ fn refuse_adapter(device_type: wgpu::DeviceType, allow_software: bool) -> bool {
 }
 
 /// `LIGHTCRAFT_GPU_ALLOW_SOFTWARE=1`: accept a software adapter. For tests on machines without a
-/// GPU (CI runners: Mesa lavapipe on Linux, WARP on Windows), so the CPU ↔ GPU equivalence tests
-/// run there instead of skipping; the app never sets it (software rendering is slower than the CPU
-/// pipeline).
+/// GPU (Mesa lavapipe on Linux, WARP on Windows), so the CPU ↔ GPU equivalence tests run there
+/// instead of skipping; the app never sets it (software rendering is slower than the CPU pipeline).
 fn software_allowed() -> bool {
     std::env::var("LIGHTCRAFT_GPU_ALLOW_SOFTWARE").is_ok_and(|v| matches!(v.trim().to_ascii_lowercase().as_str(), "1" | "true" | "yes" | "on"))
 }

crates/gpu/tests/it/equivalence.rs

@@ -2,8 +2,8 @@
 //! `lightcraft_gpu::render`, compared in 8-bit sRGB.
 //!
 //! Bounds (documented in `docs/gpu-pipeline.md`): mean |Δ| < 0.5 LSB and max |Δ| ≤ 3 LSB per
-//! channel for every case. Skips (passes with a note) when no GPU adapter exists; CI runs them on
-//! software adapters with `LIGHTCRAFT_GPU_ALLOW_SOFTWARE=1` (lavapipe, WARP) and on Metal.
+//! channel for every case. Skips (passes with a note) when no GPU adapter exists;
+//! `LIGHTCRAFT_GPU_ALLOW_SOFTWARE=1` runs them on a software adapter (lavapipe, WARP) instead.
 
 use std::sync::Arc;
 
@@ -899,7 +899,7 @@ static BIG_SOURCE_TESTS: std::sync::Mutex<()> = std::sync::Mutex::new(());
 
 /// The big photo of these tests: 10.24 MP, a 117 MiB source. That is over what a view keeps with its
 /// stages (96 MiB, so it is shared) and within a 128 MiB storage binding (the WebGPU default, and
-/// the limit of Mesa's lavapipe in CI), so the source goes to the device on every adapter. A source
+/// the limit of Mesa's lavapipe), so the source goes to the device on every adapter. A source
 /// over the device's limit is resampled on the CPU instead and never uploaded, which would leave
 /// these tests nothing to check.
 const BIG: usize = 3200;

docs/gpu-pipeline.md

@@ -50,8 +50,8 @@ Overrides (read once at launch):
 - Both `LIGHTCRAFT_GPU=0` and `LIGHTCRAFT_GPU_BACKEND=off` also keep AI Denoise, the face models
   and the AI mask model (SAM 3 on Metal) on the CPU.
 - `LIGHTCRAFT_GPU_ALLOW_SOFTWARE=1`: accept a software adapter (llvmpipe / lavapipe, WARP), which is
-  otherwise refused because the CPU pipeline is faster. For tests on machines without a GPU (CI);
-  the app never sets it.
+  otherwise refused because the CPU pipeline is faster. For running the GPU tests on a machine
+  without a GPU (below); the app never sets it.
 
 **DX12 shader compiler (issue #471).** Every instance (window and compute) compiles DX12 shaders with
 FXC (`d3dcompiler_47.dll`, part of Windows; `lightcraft_gpu::backend::backend_options`). wgpu's
@@ -275,15 +275,30 @@ bit-identical, and one (the point colour range overlay) has max 2 LSB on 0.0002%
 mean 0.0003 LSB. That is the same picture as on Metal, inside the bounds
 (`cargo test -p lightcraft-gpu --release --test it equivalence:: -- --nocapture`).
 
-**In CI** (`.github/workflows/ci.yml`, issue #1050) they run on every platform: on Metal on the macOS
-runners, and on software adapters on Linux (Mesa lavapipe) and Windows (WARP, DX12 with FXC), which
-`LIGHTCRAFT_GPU_ALLOW_SOFTWARE=1` admits (tests only: the app never sets it). The step fails if they
-skip. Measured on lavapipe (CI and locally) and WARP: all 134 compared renders within **max 1 LSB,
-mean ≤ 0.0003 LSB**, as on the M4 Pro; on the macOS runner's (paravirtual) Metal device one case reaches
-2 LSB (point colour range overlay, 0.0002% of pixels over 1), the rest max 1. The suite takes about
-50 s on lavapipe, 30 s on the macOS runner and 2–7.5 min on WARP. lavapipe's
-storage bindings stop at 128 MiB (the WebGPU default), so the two big-source tests use a 117 MiB
-source: above what a view keeps with its stages (96 MiB) and small enough to be uploaded there.
+**Running them without a GPU (issue #1050).** `Gpu::new` turns down software adapters, so on a
+machine without a hardware GPU (a VM, a container) every GPU test skips.
+`LIGHTCRAFT_GPU_ALLOW_SOFTWARE=1` lets them through for one run (tests only: the app never sets it):
+
+```sh
+rm -f /tmp/gpu-skips.jsonl   # the log is appended to; an absolute path, as tests run in crates/gpu
+LIGHTCRAFT_GPU_ALLOW_SOFTWARE=1 LIGHTCRAFT_SKIP_LOG=/tmp/gpu-skips.jsonl cargo test -p lightcraft-gpu --tests
+grep '"reason":"no GPU adapter' /tmp/gpu-skips.jsonl   # no output: nothing skipped for want of an adapter
+```
+
+- **Linux:** Mesa's lavapipe (Vulkan; `mesa-vulkan-drivers` on Debian and Ubuntu). Its storage
+  bindings stop at 128 MiB (the WebGPU default), so the two big-source tests use a 117 MiB source:
+  above what a view keeps with its stages (96 MiB) and small enough to be uploaded there.
+- **Windows:** WARP, the DX12 software rasterizer that ships with Windows (shaders through FXC, as
+  above).
+- **macOS:** Metal is a hardware adapter, so the tests run without the variable.
+
+A machine with a hardware GPU picks it over the software adapter, variable or not. Each test that
+skips prints `skipped (gpu): …` and appends a line to the `LIGHTCRAFT_SKIP_LOG` file; with a software
+adapter the only one left should be the device-loss test (`no GPU device, or one that can't be lost on
+purpose here`). Measured 2026-10-11 on lavapipe and WARP, and again 2026-10-30 on lavapipe (Mesa's
+llvmpipe on LLVM 20, 143 compared renders), every compared render stayed within **max 1 LSB, mean ≤
+0.0003 LSB**, as on the M4 Pro. The equivalence binary takes about 20 s on lavapipe and several
+minutes on WARP.
 
 Remaining differences come from f32 vs f64 coordinate math, fast-math transcendental functions on
 Metal and summation order in the box filters — all far below one 8-bit step.