Files
MetaCrate/docs/release-ci-matrix.md
Chili Palmer c0be04968f
Some checks failed
CI / required (push) Failing after 7m31s
Consolidate required CI gate (#115)
2026-08-12 18:50:04 +00:00

121 lines
6.1 KiB
Markdown

# Consolidated CI and release matrix
MetaCrate has one routine code-ready signal: the `required` job in
`.gitea/workflows/ci.yml`. It runs for every push and pull request on
`ubuntu-latest`, has a hard 15-minute timeout, and cancels an older run for the
same branch or pull request. `.gitea/workflows/release.yml` is the only other
workflow. It is manual, weekly, and tag-only; it never runs for an ordinary
source push or pull request.
The reviewed coverage map is `ci/ci-coverage.json`. It maps every responsibility
from the former 13 workflows to the required gate, the non-routine release
gate, or both, with a rationale. `metacrate-ci-matrix ci-audit` rejects missing
routes, additional workflow fan-out, non-Ubuntu jobs, Skia source-build
fallback, and changes to the 15-minute limit.
## Required gate
The Rust `required-gate` command owns the command graph and records the elapsed
time and result of every stage. YAML performs checkout, one package install,
one coherent persistent-cache inspection, pinned toolchain/tool setup, and exact native-cache
preparation before invoking it. Fast-failure order is:
1. formatting, deterministic generators, API mappings, parity catalogs, and
milestone ownership checks;
2. regenerated API, documentation, dependency, and provenance evidence;
3. exact Skia ARM64 GNU binary installation proof;
4. all-feature workspace library/integration tests, minimal and JPEG 2000
feature isolation, compatibility tests, and concurrency regressions;
5. offline/fake program smoke and downstream API/SemVer consumers;
6. Clippy, warning-free documentation, dependency policy, unused dependency,
provenance, and committed performance-evidence audits.
The driver stops starting work after its 12-minute internal target. This leaves
three minutes for shutdown and artifact upload beneath the workflow's hard
limit. `CARGO_BUILD_JOBS=2` is the bounded ARM64 setting; it permits useful
parallelism without the repeated high-memory compiles caused by the former
13-job fan-out.
Run the structural audit locally on Linux, Windows, or macOS with:
```sh
cargo run --locked -p metacrate-ci-matrix -- ci-audit
```
The full required graph expects the CI native-cache environment. On an ARM64
GNU Linux host with OpenJPEG installed by `tools/install_openjpeg_2_5_4.sh` and
the pinned Skia archive available, reproduce it with:
```sh
OPENJPEG_PREFIX=/path/to/openjpeg \
METACRATE_EXPECTED_HOST=aarch64-unknown-linux-gnu \
METACRATE_SKIA_ARCHIVE=/path/to/skia-binaries-a25a0fdb7d90429aa2d1-aarch64-unknown-linux-gnu-jpegd-jpege-pdf-svg-textlayout-vulkan-webpd-webpe.tar.gz \
cargo run --locked -p metacrate-ci-matrix -- \
required-gate --evidence artifacts/ci/local-required.json
```
The driver is shell-free and constructs child processes with argument arrays,
so command and path mappings are portable. Gitea jobs intentionally remain
Ubuntu-only. Windows GNU and macOS portable surfaces are cross-checked from
Ubuntu by the non-routine release graph; no native runtime claim is inferred.
## Cache identity and invalidation
The dedicated VPS runner's persisted tool-cache mount stores Cargo registry/git
downloads, the two pinned audit binaries, the routine target directory, and
native archives/installations. The target directory is content-keyed by Rust
1.97.1, `aarch64-unknown-linux-gnu`, `Cargo.lock`, the CI coverage and dependency
policies, the Skia feature manifest, and the OpenJPEG installer. Changing any of
these selects a new target generation. This avoids depending on an external
GitHub cache service that is unreachable from the isolated Gitea runner.
The manual `cold_cache` dispatch input clears only those exact keyed Cargo,
tool, and native paths before setup; it exists to produce the required cold-run
measurement. Ordinary pushes and dispatches reuse the last validated generation.
OpenJPEG caches include a three-line identity containing version 2.5.4, commit
`6c4a29b00211eb0430fa0e5e890f1ce5c80f409f`, and runner architecture. The
installer checks this identity and `pkg-config --exact-version` before reuse.
The Skia cache is the exact 0.99.0 ARM64 GNU feature archive with SHA-256
`dd127f458a5e67a79f3936a8aa19f822fe90a1d6a11b50b5f84df2b0519d909c`.
The workflow never sets `FORCE_SKIA_BINARIES_DOWNLOAD`: rust-skia uses the
normal crates.io `.cargo_vcs_info.json` hash. The proof reads rust-skia's
persisted build-script output even on a cache hit and requires
`DOWNLOAD AND INSTALL SUCCEEDED`; the Rust driver rejects missing success output
or any `STARTING A FULL BUILD`. The adapter enables
rust-skia's `no-compile` feature on every target, making an unavailable or
incorrect binary a fast failure rather than a source build.
To invalidate local artifacts, remove only the relevant target directory or use:
```sh
cargo clean --target-dir target/required
cargo clean --target-dir target/release-gate
```
## Non-routine release gate
The Rust `release-gate` command owns clean MSRV and portable cross-target
checks, benchmark compilation, the 16-cycle resource soak, all 16 source
packages, release binaries, native artifact inspection, and the aggregate
release-candidate audit. Reference C# performance regeneration and live-grid,
physical-audio, proprietary-service, signing, and platform-native installer
checks retain their documented credential/device/manual boundaries. Their
committed evidence and fast regression checks remain in routine CI.
## Troubleshooting and runtime evidence
- A host mismatch means the required runner label is not the documented ARM64
GNU VPS. Do not substitute a musl or x86 archive.
- An OpenJPEG identity failure requires deleting only its native cache entry and
rerunning the installer.
- A Skia hash/download failure indicates an upstream release-asset or redirect
reachability problem. Restore the exact archive in the configured cache; do
not enable source compilation.
- A 12-minute driver failure identifies the last completed stage in the uploaded
`required-gate.json`. Optimize or remove duplicate work only after updating
the coverage manifest; do not raise the 15-minute timeout.
The validation record and five qualifying ARM64 run links are maintained in
[`ci-runtime-evidence.md`](ci-runtime-evidence.md).