Issues and patches written without announcing them, such as with
`rad issue comment --no-announce` or by another instance of the app,
never reached the open views. The cache also changes with every fetch and
write the app already heard of, and doesn't say which repo changed, so
right after those it refreshes once things are quiet instead of with each
one. The watchers now also notice a database that appears or comes back
after an error, and one locked database no longer holds up the others.
An update that arrived while a route was still loading was lost, since the
loader's result then counted as fresh, so the events now invalidate through
the shared helper that marks such fetches stale. A repo arriving or renamed
without the update naming its refs left the sidebar list as it was, and
updates that changed nothing still dropped the hover caches.
A test now checks the frontend's ref update type against the generated
one.
The repo list had its own TTL cache, a counter bumped by seeding and a
listener of its own, and release counts were fetched separately by the
sidebar and the releases page. Now they're queries like the rest, and
seeding invalidates the list like any other write. The head a loader
resolved is marked through `markLoading` too, so it needs no key of its
own.
Hover cards, references and mention candidates keep their own short-lived
caches, which went on showing an issue's or patch's old state for up to a
minute after the views around them had updated.
Every status chip of a repository shares one list of its jobs, which
refreshes when a job changes, at most every few seconds, since listing
them scans every job of the repository.