The base class carried a lazy_views flag, a _views_source field, a _build_views split
and an introspection check for subclasses that decorate their views - all of it shared
by 37 plugins to serve one. Overriding update_views/get_views in processlist does the
same with no change to model.py, and set_views/reset_views clear the pending flag so a
reset stays reset.
update_views() called get_raw() twice and returned an empty dict from a separate exit;
one read and one exit instead, with the deferred case named in the docstring.
The test fixture built its own ProcesslistPlugin, whose load() writes Mock attributes
into the glances_processes singleton and broke seven tests in test_core.py whenever the
two files ran together. It now takes the shared plugin and restores what it changes.
One rationale in one place instead of three, and the class attribute no longer
describes the per-key build that is gone. test_unknown_key_raises only exercised
dict lookup and passed unchanged on develop.
A dict subclass that built entries on access turned out to be more machinery than the
problem needs: nothing outside the plugin reads .views directly, and every per-item
caller belongs to a plugin that builds eagerly. Remembering the stats and building once,
on the first read, gives the same result with a plain dict.
dipetm