[tonyandrewmeyer]

A collection of framework edge cases, each of which surfaces a long way from its cause:

• StoredList.append, insert and extend stored the wrapped value, so putting one stored value inside another only failed in commit().
• observe() records the method's __name__ and looks it up again at dispatch, so a handler decorated without functools.wraps was registered but never called, and the deferred notice was silently dropped. The name is now checked when the observer is registered, with a warning if it doesn't find the method, and a missing method is logged at dispatch rather than ignored.
• A Handle key containing /, [ or ] was written into the path but couldn't be read back by from_path, failing in a later dispatch. It's now rejected when the handle is created.
• Handle.__eq__ raised AttributeError when compared with anything that wasn't a Handle.
• A deferred event whose type is no longer registered (because the charm only conditionally observes it) was dropped without any log message. It's now a warning.
• remove_unreferenced_events skipped event kinds containing digits, such as redis2_pebble_ready.
• PrefixedEvents and BoundStoredState recursed in __getattr__ when an attribute was looked up before __init__ had run, so copy.copy(charm.on['db']) raised RecursionError.
• A deferred StorageEvent whose storage had since been removed raised RuntimeError in restore() on every dispatch from then on. The Storage is now rebuilt from the snapshot (which already has the location) so the handler still runs.

One of these newly rejects input that was previously accepted. A handle key with /, [ or ] only failed before if an event was emitted through that object (a StoredState-only object with such a key worked, because load_snapshot takes a Handle rather than parsing a path). I think that's a bug rather than behaviour anyone relies on, and a Hyrum run found no charms that break. A handler without a matching attribute name still registers, but now logs a warning, because rejecting it broke the unit tests of oauth2-proxy-k8s-operator.

Two collateral changes to existing tests: test_reemit_ignores_unknown_event_type was using in-memory storage for both frameworks, so the second one had nothing to re-emit and the test passed vacuously - it now uses tmp_path. The removed-storage test resets model.storages._storage_map directly, to simulate a fresh dispatch without a second framework.

The last fix is the most significant, and the one I'm least sure about: the alternative would be to drop the notice, but running the handler with the recorded location seems closer to what the charm author would expect. The warning for the dropped deferred event could also reasonably be a debug message, if people think it'll be noisy.