[server/sources/native] deal with connections, not sessions

[server] eschema_eid needs a connection, not a session

[repository] operations get a connection instead of a session Left extid2eid unchanged because that function scares the shit out of me.

[server/hook] operations get a connection instead of a session

[repository] Use an internal connection in register_user

avoid to internally raise 3.19 warnings

[sources] ensure we have a cnxset in undoable_transactions

[server] move security/integrity hook management away from InternalSession.__init__ and to Repository.internal_{session,cnx} instead. Lets internal_cnx avoid deprecation warnings.

[migration] make 'session' object be a server-side Connection, not ClientConnection I think this better corresponds to the previous behaviour. However, that Connection does not have a cnxset, which should probably be changed (for both interactive_shell and cmd_process_script).

[sources] Skip problematic sources when starting shell instead of crashing. Closes #3670208 At least this gives an opportunity to fix the problem in using c-c shell, which is currently not possible because it crashes during initialization. Notice that since sources are copy based, it should not be a problem at all to have some disabled, beside for services such as authentication sources as ldapfeed.

[web] fix language negotiation Regression caused by commit 6fd0ac6506cb "[web-request] handle default language earlier". We would handle language negotiation at request init time, but overwrite the selected language with the site default when calling set_cnx.

[merge] bring the 3.17.14 fixes to 3.18

[test/utils] repair concat-resources test That was forgotten in 240a620b9cd3. Related to #3670503.

[pkg] bump cubicweb.spec

[dataimport] Correctly call rschema(rtype) in SqlGenObjectStore, closes #3694139

[pkg] prepare 3.18.4

[web] Disable 'concat-resources' by default (closes #3670503) It makes debugging harder and it can actually *break* working JS or CSS files. This is the first step towards complete removal.

[testlib] Drop override of assertItemsEqual method The override is not compatible with underlying (standard) method because it builds sets and rely on eid. Closes #3641111.

[rewrite] Fix crash when the main variable doesn't appear in the snippet's vargraph Also, the added test discover another bug once the first has been fixed (erroneous argument given to .rewrite recursive call). Closes #3661918

[ext/rest] Catch SystemExit raised by docutils docutils publisher may raise SystemExit which is not caught by Exception. Closes #3648763.

[entity] User-defined relation to skip for copy has precedence Otherwise permission problems can occur on rtypes not yet skipped Closes #3653459

Make EditController edit_entity method always return an eid In cases the entity was to be created or copied, eid was None hence the method returned None despite the dosctring says the method should always return an eid. Now if the eid variable is None, it is assigned to the newly created entity eid. Closes #3593029.

[migration] Improve update of in-memory schema during 3.18 CWAttribute.defaultval change We weren't updating the schema properly so a bunch of structures still had the old rdef with 'String' as object. That caused failures to add new attributes later in the migration as their default value would be a Binary object but the schema would expect a String. Closes #3683640

[migration] always rebuild infered relation This was skipped for some bad reason (see 12ad88615a12 which introduced the change). Fix for #231956 in Yams is necessary to allow this cset: during a migration, we want to always reinfer relations while allowing explicit redefinition of an infered relation later. Else, we may run the migration with missing parts of the schema (the one that should have been infered). Closes #3685463.

Backout "[web/navigation] use add_onload instead of inline javascript href" This reverts changeset 170e1437948d, for at least two reasons: the PageNavigationSelect component was not updated for the new type of navigation links, and external cubes defining View.page_navigation_url exist in the wild and also got broken navigation links. This reopens #3501626 and closes #3677945.

[test] those are rql tests, not cubicweb (and they break with rql 0.31.5). Remove them.

[source/native] allow many eid creation per .create_eid call (closes #3526594) [jcr: move migration code to bootstrapmigration_repository.py to make sure we stop using the old sequence ASAP]

[pkg] bump the dependency on logilab.database This will allow use of an extended API. Related to #3526594.

[source/native] refactor eid generation code This set of related methods clutters the native source API. It will be better served with a dedicated eid genrator object. Related to #3526594. [jcr: allocate the eid generator in __init__]

Move entity cache from web.request.ConnectionCubicWebRequestBase to repoapi.ClientConnection ResultSet._build_entity relies on an entity cache to avoid going round in circles (stack overflow due to infinite recursion) e.g. in a postcreate script. Clear the cache at commit/rollback time to avoid unexpected side effects; and explicitly clear the cache in a few tests where we mix ORM and RQL using the same connection.