ci: fix local scheduling and Windows test baselines (#4941)

* ci: repair local nextest scheduling and Windows fixtures

* test: refresh MSVC robustness references and isolate GNU policy

Integrate the Windows reference records from #3361 and the target-environment
selection and recording policy from #4159. Keep shared records and comparison
tolerances intact, and retain seed 36 as an existing failure tracked by #3583.

Co-authored-by: Ray <raykholo@users.noreply.github.com>
Co-authored-by: chris <christopher.nourse@gmail.com>

* test: make CI subprocess and revolve fixtures deterministic

---------

Co-authored-by: Ray <raykholo@users.noreply.github.com>
Co-authored-by: chris <christopher.nourse@gmail.com>
This commit is contained in:
cronos
2026-10-24 11:33:10 +00:00
committed by GitHub
parent 4612c5e6a9
commit 592f6a6fee
9 changed files with 625 additions and 90 deletions
+29
View File
@@ -68,6 +68,35 @@ with doc tests alongside. Arguments after `--` go to each test binary, for examp
`cargo xtask test -- shell`. To run the test binaries sequentially instead, use
`cargo test --workspace --profile ci`.
## Test-level scheduling for full local CI
Install [cargo-nextest](https://nexte.st/docs/installation/pre-built-binaries/), then set
`SOLIDCRAFT_TEST_RUNNER=nextest` when running `cargo xtask ci`. For example, on PowerShell:
```powershell
$env:SOLIDCRAFT_TEST_RUNNER = 'nextest'
cargo xtask ci
```
This optional backend builds the workspace tests once, then schedules individual tests using
the `local-ci` nextest profile, which inherits `ci`. The action matrix runs first, then the longest geometry
cases start before the other costly tests, so the ordinary queue does not stall behind an
all-slot reservation. It **ignores the default filter**,
so the heavy tests below still run, does not stop after a test failure, and does not retry failures.
Tests marked `#[ignore]` remain ignored, as with the ordinary full local CI.
The default concurrency is at most four tests; `NEXTEST_TEST_THREADS` overrides it. The action
matrix's child workers default to the smaller of that budget and the host CPU count;
`ACTION_MATRIX_JOBS` explicitly overrides this worker count. The local profile reserves CPU
slots for the matrix and keeps its scheduling changes separate from GitHub CI. On Windows,
the `solidcraft-gpu` and app packages, plus UI offscreen/GPU tests, use a serial group. The Wasm check and doc tests still run alongside,
and the web-offline gate still follows a successful native/Wasm/doc run.
The backend keeps Cargo and binary metadata under `target/xtask-nextest/` and reuses it for the
native run, freeing Cargo to run the side gates. Nextest prints each test's result and duration
as it finishes. `SOLIDCRAFT_TEST_RUNNER=libtest` (or leaving it unset) retains the original runner;
`cargo xtask test` and its Cargo/libtest argument forwarding are unchanged. Unknown runner names
fail explicitly.
## CI sharding and heavy tests
On GitHub, one job builds the tests into a [cargo-nextest](https://nexte.st/) archive and four