The rate limit, the version compare and the download check were all wrapped
in Android — a Context, a socket, a file — so none of them had a test, and
all three shipped broken: the remote thrown out after five reconnects
(0.11.1), a truncated APK handed to the installer (0.11.5), and a version
compare that would read 0.11.10 as older than 0.11.9 the moment we get
there.
They are now plain Kotlin in AttemptBudget and UpdateRules, with the clock
injected, and 20 tests covering the failures themselves. AttemptBudget also
prunes expired entries before evicting: dropping only empty queues let the
map grow without bound while every tracked address held one fresh failure.
tests/smoke.sh installs the debug APK on the headless emulator, walks the
first screen with the D-pad and fails on a crash or ANR — the class of bug
unit tests cannot see, and the reason 0.11.0 crashed on startup. Forty
seconds end to end.
tools/release.sh does the bump, tests, signed build, tag, apk branch and
Gitea release in one command. It refuses an unsigned build and never writes
the release notes itself; commit subjects are offered as a draft.
tests/helpers/jdk.sh picks a JDK Gradle can run on: Ubuntu moved
default-java to 25, which fails the build with a bare
"IllegalArgumentException: 25.0.4".
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three fixes.
The pulse ring never showed. It was drawn on a 30dp canvas inside a box
sized to the 20dp crest, so the parent's constraints shrank the canvas
to 20dp and the crest image covered what was left. The box is now wider
than the crest and the ring has room.
"läuft" was drawn straight onto the row, and a focused row is filled
with the accent colour — red on turquoise. The signal now sits on a
dark chip, which reads on an unfocused row, a selected one and a
focused one alike.
The updater handed the installer whatever had been written so far. It
now checks the response code, downloads to a .part file, compares the
byte count against Content-Length before renaming, and refuses a second
concurrent attempt instead of letting two writers share one file. A
half-written APK leaves the installer spinning with nothing to report,
which is what a hang looks like from the sofa.
Crests are also resampled to 128px with a proper filter when fetched;
Android's 8x downscale from the 330px originals left ragged edges.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Fetching 56 images on every run earned nothing but rate limits, and the
CI artefact is a test build that does not need the badges — the shield
fallback covers it. The resolved image URLs are cached in
tools/crest-sources.json, which is plain text and safe to commit, so a
local run needs one request per club instead of two and can resume after
a rate limit. The artwork itself stays out of the repository.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Resolving crests from Wikipedia at runtime never produced a visible badge
on the TV and left nothing in the log to debug, so the shield fallback was
all anyone ever saw. They are now downloaded once at build time by
tools/fetch-crests.py into assets/crests (git-ignored, so no trademarked
artwork is committed) and loaded from there — instant, and independent of
the network.
CI: accepting SDK licences by piping "yes" into sdkmanager died of SIGPIPE
(exit 141) under pipefail; the licence hashes are written directly now.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
