fix(web-app-suitecrm): raise the stored system user around the external-auth save

495d23dbde raised an admin context with a fresh in-memory bean, which fixed
hasSaveAccess() and moved the failure one guard further. User::verify_data()
compares $this->id == $current_user->id, and between two unsaved beans that is
'' == '', so a deployment with a single administrator takes the last-admin
branch for the very account CreateUser is in the middle of creating. The branch
reads the legacy $mod_strings, which this request never loads.

That is why the two SuiteCRM jobs of CI run 38818071194 looked like two
failures: the LDAP variant ended in ErrorException "Trying to access array
offset on value of type null" at modules/Users/User.php:1678 and answered
HTTP 500 to POST /login, while the SAML variant took the same verify_data()
failure, avoided the 500 because $mod_strings happened to be loaded, redirected
to index.php?action=Error&module=Users and left no session, so its spec timed
out waiting for the SPA instead. Both are the first login of a persona that has
no local row yet; administrator passes in both because it already has one.

Use the stored system user, which has a real id, so hasSaveAccess() passes,
restrictAdminOnlyFields() early-returns, setIsAdmin()'s reloaded bean resolves
and '' == '1' skips the last-admin branch.

Assign $current_user before calling getSystemUser(), not from its return value.
User::retrieve() runs loadPreferences() after filling the bean, and both
UserPreference::loadPreferences() and reloadPreferences() dereference
$current_user->id unguarded, so retrieving while the caller's anonymous
$current_user is still in place raises "Attempt to read property id on null"
and answers HTTP 500 one line further on. Assigning first also means the only
session preferences key written is the system user's own, never the key of the
account signing in.

The patch header records both failure modes, because the patch outlives this
commit and is applied at image build time with -F0, so whoever reads it next
when a hunk stops matching needs to know what it was for.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 207cdc70ff)

roles/web-app-suitecrm/README.md

@@ -145,10 +145,10 @@ docker run --rm -it \
 - **Auto-created accounts** reach `User::save()` through SuiteCRM's own `CreateUser`
   while the request is still anonymous. Since 8.10.2 that save answers
   `RuntimeException: Not authorized` and forces `status = Inactive`, so
-  `files/patch/external-auth-user-create.patch` raises an admin context around the save
-  and restores the caller's afterwards. [`files/Dockerfile`](./files/Dockerfile) applies
-  it with `patch -F0` on the unpacked release; the image build fails once the hunk stops
-  matching.
+  `files/patch/external-auth-user-create.patch` makes the stored system user the current
+  user around the save and restores the caller's afterwards.
+  [`files/Dockerfile`](./files/Dockerfile) applies it with `patch -F0` on the unpacked
+  release; the image build fails once the hunk stops matching.
 
 ## Persona contract opt-outs
 

roles/web-app-suitecrm/files/patch/external-auth-user-create.patch

@@ -13,6 +13,25 @@ record from both guards; upstream has no release and no hotfix carrying a fix.
 installer, so the admin context is raised here instead of relaxing the guards,
 which keeps them closed for every request-driven save.
 
+The raised context is the stored system user, never a fresh in-memory bean
+carrying `is_admin = 1`. User::verify_data() compares `$this->id ==
+$current_user->id`, which is `'' == ''` between two unsaved beans, so a
+deployment holding a single administrator takes the last-admin branch for the
+very account it is creating. That branch reads the legacy `$mod_strings`, which
+this request never loads, so an LDAP login ends in `ErrorException: Trying to
+access array offset on value of type null` at modules/Users/User.php:1678 and
+HTTP 500 on POST /login, while a SAML assertion redirects to
+`index.php?action=Error&module=Users` and leaves no session.
+
+`$current_user` is assigned before `getSystemUser()` runs, not from its return
+value. User::retrieve() calls loadPreferences() after it has filled the bean,
+and both UserPreference::loadPreferences() and reloadPreferences() dereference
+`$current_user->id` unguarded, so retrieving while the caller's anonymous
+`$current_user` is still in place would raise `Attempt to read property "id" on
+null` and answer HTTP 500 again. Assigning first makes the global the very bean
+being filled, so those reads see an object throughout and the only session key
+written is the system user's own.
+
 Generated with `diff -u` against the v8.10.2 tarball and applied with `-F0`, so
 the build fails instead of letting GNU patch's default fuzz factor land the
 hunk somewhere upstream moved it to.
@@ -37,7 +56,7 @@ hunk somewhere upstream moved it to.
 +
 +        if (!is_admin($current_user)) {
 +            $current_user = BeanFactory::newBean('Users');
-+            $current_user->is_admin = 1;
++            $current_user->getSystemUser();
 +        }
 +
 +        try {