357. A test that paints asks for the library, and a download asks again

Date: 2026-09-17

Status

Accepted. Repairs Snapshot run 12 (commit f716adec). Applies ADR-0338’s rule to one more test and adds retries to :assets’ downloads.

Context

Snapshot run 11 was green. Run 12 was red on four jobs, for two unrelated reasons, both read off the runners’ check-run annotations and job logs:

1. publish / {linux,macos,windows} / Java — :core:test, 2267 tests, 1 failed: WindowResizeTest > what the manager decided arrives through onResize, and a frame follows, with UnsatisfiedLinkError: libgoldberry not found. Those jobs build with -Pgoldberry.skipNative=true, so there is no library by design. The test came in with ADR-0342. It is the only one of the five that runs the frame loop, and a frame rasterizes. It never called RendererRequirement.enforce(), which is the fifth instance of the defect ADR-0338 fixed in four tests.
2. publish / linux / Verify layouts (linux-aarch64) — :core:prepareAssets failed with Server returned HTTP response code: 500 for github.com/rsms/inter/releases/download/v4.1/Inter-4.1.zip. The other three verify legs downloaded the same archive in the same minute. It was GitHub’s failure, and AssetCache made one attempt.

Because the Java job stops at the first failing task, :widgets, :html, :example and :natives never ran there. A local run of every module’s tests with -Dgoldberry.native.library=/nonexistent/libgoldberry.so found no other test in that state: 0 failures across 5263 tests.

Decision

The painting test skips without a library. A download retries a failure that can pass on its own, and only that.

• WindowResizeTest.arrivesThroughTheHandler calls RendererRequirement.enforce(). The other four tests there only move a headless window and keep running without the library.
• io.github.digitalsmile.goldberry.assets.download.Downloader makes the request through java.net.http with redirects followed. It tries four times, waiting 2, 4 and 8 seconds, on HTTP 408, 429, any 5xx or an IOException from the connection. Any other status fails on the first attempt: a 404 is a pin that names nothing. Each failed attempt is printed to stderr, so a slow asset step says why.
• AssetCache.fetch and fetchText go through it. fetchText is now an instance method, so the licence download shares the cache’s downloader.
• The checksum is not retried. An archive that arrives whole and hashes wrong is a changed upstream, and AssetCache’s message says so.

Consequences

• A GitHub outage that lasts longer than about 14 seconds still fails the asset step, and should: a longer wait makes a stuck run look like a slow one.
• The transport, the sleeper and the attempt count are constructor arguments, so DownloaderTest covers the retry policy with no network and no waiting.
• Any new :core test that opens a window and runs Goldberry.run() with a paint handler needs the same guard. The no-library run above is how to check: ./gradlew test --continue -Dgoldberry.native.library=/nonexistent/libgoldberry.so.