The core/tests/ harness

Most differential tests render on the software adapter (prefer_software) so they hold on any GPU; the exceptions say so below. Run the whole suite:

cargo nextest run -p rlx-core     # what CI runs (per-test process isolation)
cargo test -p rlx-core            # single binaries only — see the caveat below

Most of these files share one test binary (ADR-0204). core/tests/suite/ and standalone/tests/suite/ each hold the files that need no binary of their own, as modules of a main.rs, so a test’s name carries its file: easing::…, and cargo nextest run -p rlx-core --test suite easing:: runs one file’s tests. A file stays a top-level tests/*.rs when a binary() selector in .config/nextest.toml names it (the GPU suites below that -P fast excludes, and the clock-reading tests scheduled alone), when it reads the clock, or when it reads a process-level quantity such as its own memory. The rule for which is which lives in core/tests/suite/main.rs, and hygiene.rs fails on a clock exemption inside either suite/.

Use nextest for the whole suite. preset’s zero-allocation assertion is not the reason: its allocator hook counts per thread, so a concurrently-running test’s allocations never reach the count, under either runner and inside the shared suite binary alike.

The reason is that stock cargo test runs a binary’s tests as threads in one process, which the GPU tests do not survive. Several of them build and drop a Renderer (and so a wgpu device) concurrently, and the driver aborts the process with STATUS_ACCESS_VIOLATION — transition’s tests every run on WARP, --lib intermittently at teardown, --lib render::post since Plan 0035. This is a runner artifact, not a failing test: the same binaries pass in full under cargo nextest run, which gives each test its own process. If a cargo test invocation aborts with 0xc0000005, re-run it under nextest (or -- --test-threads=1) before concluding anything about coverage.

Tests that need a real GPU (background_composite, and the in-crate dual-live dissolve check in render/mod.rs) skip themselves when only a software rasterizer is present, per ADR-0016. WARP mis-renders both: the fullscreen-scene + background pipeline set, and — once a dissolve allocates its blend targets mid-run — what the feedback trails stage resolves to.

In this section

• The preset sweeps fan out in batches
• What the five preset gates can and cannot see
• Golden baselines
• The tonemap and pixel-level assertions
• The display write dithers, and every baseline moved once
• A lit backdrop is a distinct test configuration, not a variant
• Asserting that something moved — the feedback fixtures
• The in-frame geometry fraction, and the four things it cannot see