0.2.3

No database migration, so going back to 0.2.2 is safe. One new optional setting,
DATABASE_STATEMENT_TIMEOUT_MS, whose default needs no action. On a Swarm
deployment Caddy restarts once, for a few seconds, and its certificates are not
reissued.

The reason to take it is the first fix below: the Checks page could stall for
minutes in a desktop browser while the server sat idle. The interface on a phone
is the other half of the release.

Fixed

• A page waiting on a server that never answers now says so instead of
  spinning. Every read had no time limit, so a connection that went silent —
  a proxy that lost its upstream, a socket the network dropped without a word —
  left the Checks page on its loading bar for as long as the browser cared to
  wait, which is minutes. A read now gives up after 30 seconds, aborts the
  request, and shows "The server is taking too long to answer". Writes are left
  alone: one that timed out in the browser may still have succeeded.
• The Checks page asks for one thing at a time. Its background refresh fires
  every 15 seconds whether or not the last answer had come back, so against a
  stalled server it stacked a request behind every stalled one and released them
  all together when the server recovered.
• The Swarm stack no longer offers HTTP/3. Caddy advertised it to every
  browser (Alt-Svc: h3=":443"; ma=2592000, which a browser remembers for thirty
  days), but QUIC is UDP and the routing mesh in front of Caddy is the one place
  a UDP flow is not dependable. The symptom matches what was reported: on one
  desktop browser the Checks page stalled with its requests unsent — no protocol,
  nothing transferred — and recovered after about five minutes, which is how long
  Chrome keeps QUIC marked broken before trying it again; the same site on a
  phone, and curl, which never tries QUIC, were fine throughout. Caddy now
  serves HTTP/1.1 and HTTP/2 only, answers Alt-Svc: clear so browsers that
  already learned the old advertisement forget it on their next visit, and
  443/udp is no longer published. Nothing to configure. Caddy restarts once on
  the deploy that carries this, a few seconds of interruption; its certificates
  are in a volume and are not reissued. Turning QUIC off in the affected
  browser (chrome://flags/#enable-quic) made the stalls stop, which is what
  points at this; if they continue after this release, the logging added below is
  how to find out what it was instead. Docker Compose is unchanged — it publishes
  UDP through Docker's own proxy, not the routing mesh.

Changed

• The interface works on a phone. It used to come apart: the Checks table
  showed two columns and scrolled the state — the one that matters — off the
  edge, forms were clipped, and the three destinations were a second row of small
  links. Under 820px (which includes a tablet held upright) the header is one
  row with the project at full width and the three destinations move to a tab bar
  under the thumb; tables become lists of cards, the state beside the name and
  the facts below it; the check form takes the whole screen with its buttons
  pinned at the foot, while confirmations stay small; and everything you press is
  44px, with 16px fields so iOS does not zoom in on focus. Sorting is a column
  header, so on a phone each list keeps the order it was designed to be read in.
  Wider screens are unchanged. Checked in a real browser at 360, 390, 768, 844×390
  and 1280 pixels on every screen and dialog, not yet on a physical device.
• A hint longer than one line no longer overlaps the next field. It did on
  every screen, not only a phone: the form field reserved one line for it.
• Database queries are bounded, and idle connections are probed. A query on
  the API's connection pool is cancelled after 30 seconds
  (DATABASE_STATEMENT_TIMEOUT_MS), and all three database connections — API,
  ingestion, and the listener that invalidates the ingest cache — now send TCP
  keep-alive probes. Before this, a connection the network had dropped never
  failed on its own: the request using it waited for TCP to give up, and the
  listener never noticed it was deaf.
• The server logs what it cannot otherwise explain afterwards. A request that
  takes two seconds or more is logged by route pattern — never by URL, since
  heartbeat URLs are secrets — and so is any stretch of 500 ms or more in which
  the event loop did not turn. A healthy server writes neither. When the
  application seems to freeze, these say whether the process was stuck or
  something around it was.

Container image

ghcr.io/liliang-dev/silencewatch:0.2.3