The integration is moving to a new repository under a new domain. Home Assistant
keys recorder history and long-term statistics on entity_id, not on the domain,
so the move is free provided the registry entries are carried over instead of
recreated. A recreated entity gets a fresh entity_id and its history is orphaned.
predecessor.py does the carry-over with async_update_entity_platform, which
rewrites platform and config_entry_id and leaves entity_id alone. Everything the
user set by hand rides along, because the registry row itself survives: renames,
icons, areas, labels, categories, hidden and disabled state.
Guards, each one covered by a test that fails when it is removed:
- The startup placeholder state (unavailable + ATTR_RESTORED) is cleared first,
because async_update_entity_platform accepts only a missing or unknown state.
- A state that is not a placeholder means an integration is still driving the
entity, so it is refused rather than moved out from under a live platform.
- A unique_id already taken under our own domain is refused. Home Assistant does
not check this itself: the duplicate guard in _async_update_entity only runs
when new_unique_id is passed.
- Devices are repointed before the old entry is removed. A device that loses its
last config entry is deleted, and that takes its entities with it.
- The old config entry is removed only when every entity made it across.
Otherwise it stays, a retry stays possible, and nothing is lost.
Options are inherited in a separate, earlier pass. Adopting the registry alone
is not enough: SENSORS_TO_LOAD decides which entities the platform creates, so
an empty one leaves every adopted entry showing "no longer provided" until a
payload arrives, and derived sensors are never in a payload at all. The merge is
predecessor-fills-gaps, so anything just entered in the config flow wins. The
version 1 spelling of the Pocasi Meteo flag is translated on the way in, since
async_migrate_entry runs against this entry's own version and would never look
at a key copied in afterwards.
The two passes run at different points in async_setup_entry. Options first,
before the coordinator and routes read them, or an inherited WSLINK flag would
leave the entry listening on the wrong endpoint until a reload. Adoption after
the routes are proven but before async_forward_entry_setups, so the one
destructive step cannot run in a setup that then fails, and so our platforms
bind to the adopted registry entries rather than to freshly created ones.
Two problems the migration surfaced, fixed here as well:
- register_path now catches ValueError alongside RuntimeError and raises
ConfigEntryNotReady with a message naming HACS. aiohttp rejects a duplicate
route name with ValueError, and the route names are fixed rather than
domain-scoped, so a user who has not yet removed the old version hit an
unhandled traceback.
- add_new_sensors applies the derived-sensor expansion that async_setup_entry
already applied. Derived sensors are computed, never sent, so discovery can
only offer their inputs; the azimut unlocked by a newly discovered wind
direction went missing until the next restart. Only the sensors this batch
unlocks are added, so nothing is created twice.
PREDECESSOR_DOMAIN equals DOMAIN for now, which short-circuits the whole pass.
It stays correct until the day the domain changes.
SchiZzA