gloryfromca ce2713e9a5 fix(*): ship the cloud memory plugin with raven, not beside the release
The release builds everos-cloud-memory's wheel and attaches it, but no
installer and no upgrade named it, so the only way in was `uv tool install
--with`, which the next `raven upgrade` undoes: uv replaces the tool's
requirement set with what the command names, and the upgrade names the
release's plugin list alone. A user who picked the cloud backend was told to
install a plugin that had no install path.

The live-service gate that kept it out is met: the six key-gated tests in
tests/integration/test_everos_cloud_real_cloud.py ran against api.evermind.ai
on 2026-10-09. So the distribution joins the list a complete install is made
of -- release.yml, publish_beta.py for the beta channel, both installers'
local-clone lists, and the Docker image's three plugin defaults. Installed is
not selected: memory.backend still decides, doctor and the gateway build only
the backend it names, and the one visible change is that `raven onboard` asks
which memory backend to use rather than taking the only one installed.

The rung that drops the product engines kept only everos-memory. That rung
exists so long-term memory outlives a native build a platform refuses, so on
such a machine every upgrade uninstalled the cloud backend from under a host
configured to use it. install.sh, install.ps1 and the upgrade helper now keep
both memory distributions, and the contract test reads which those are off the
backend/distribution pairing rather than naming one by hand.

Co-authored-by: Claude (claude-opus-5) <noreply@anthropic.com>
2026-10-20 21:54:35 +08:00
..