v2.47.0

Fixed

• A calendar event's object name no longer decides whether it arrives (#883). A CalDAV event was
  never picked up: no error, no warning, the sync reported success with an unchanged event count. The
  report pointed at the iCal parser - DURATION instead of DTEND, a TZID without an inline
  VTIMEZONE. The parser handles both, and the reproduction from the report now ships as a test so
  the false trail does not come back.

  The cause sat one layer above it. tsdav filters the hrefs of a calendar-query response on
  .ics in the path by default, but the extension is pure convention - RFC 4791 prescribes no name
  for the object resource, and a server may assign its own. Stalwart does exactly that for everything
  created over JMAP (NZtPkIOMoK), while objects written by a CalDAV PUT keep the client-chosen
  <uid>.ics. In the same calendar part of the events synced and part did not - and because the
  filtered-out ones were never fetched, no log line could mention them. The two conspicuous
  properties and the object name all come from the same JMAP origin; that is what the diagnosis
  tripped over.

  Yuvomi now uses the rule tsdav itself applies on the CardDAV side - let everything through except
  the collection itself - and applies it at the client rather than at the call sites, since five
  places fetch calendar objects and a sixth would lose the rule again. Two files that were building
  their own CalDAV client turned up in the process and are now held by a guard. The same filter also
  covers the outbound path, where the same objects were being dropped a second time.

  Per-event visibility comes with it, as the report asked: anything the parser discards is now
  reported with its UID and the reason instead of being swallowed. A skipped event was
  indistinguishable from one the server never sent, and that is precisely what made this impossible
  to diagnose from the outside.

• Deleting a series from a synced calendar says so first (#880). Tapping one occurrence in the
  month view and pressing delete made every occurrence disappear without a word. The scope was right

  • Yuvomi cannot split a series that belongs to another calendar, an excluded occurrence would come
    back on the next sync - but doing it silently was not.

  A foreign series now asks first, and the question names the reach rather than offering a choice;
  a dialog with only one selectable answer would be a prop. The wording follows what actually
  happens, because a promise that does not hold is worse than none: a birthday event mirrors a
  birthday entry and is recreated on the next run, so it points at the birthday itself; an ICS
  subscription event cannot be deleted at the source at all and returns with the next fetch;
  everything else is told that the whole series falls, with all its occurrences. Single events are
  unchanged - the undo toast carries those.

• The module head is one row again on desktop (#882). Tasks, Contacts, Budget and Birthdays broke
  their head onto two rows, the Calendar in its week, day and agenda views. The rule that pulls the
  end of the head row back to the reading measure did it with a margin on the last slot - and a
  margin counts towards the flex container's line occupancy while never yielding. Measured at
  1960px: the actions slot claimed 406px of content plus 560px of margin out of a 1280px row, leaving
  315px for seal, title and search where they needed 441px. The wrap was arithmetically unavoidable
  rather than content-dependent.

  That offset is now a shrinkable slot, and from 1024px up a head no longer wraps: flex splits rows by
  the hypothetical sizes, i.e. before anything has shrunk, so the head broke while yielding slots
  stood right next to it. The page title is the last to give way, being the only one that cannot come
  back. The Calendar's layer chips give up their label before the head wraps, but never their tap
  target - until now, how many layers you had switched on decided how many rows your calendar head
  had. At exactly 1024px the head still wraps, and deliberately so: the sidebar leaves it 740px of
  inner width, of which the reading measure alone claims 720px.