BUG: Structure adjustments "repaired" orphaned Nodes which is not replayable

[mhsdesign]

A replay in beta 15 might fail with this error

Cannot update node with copy on write since no anchor point could be resolved for node d6cefdef-5aa4-4f27-9722-3f477c0d2c71 in content stream 092cd743-55dc-4c1e-b62c-9e4543704740

TLTR: we didn't correctly delete recursive child nodes in the past, which allowed events to exist that shoujldn't and cannot be replayed.

---

Denny and christian did some debugging for this, with the following noticeable results:

• the event in question appears to have just appeared in the live content stream, it's not related to any publish event at all, has no metadata, no causation or correlation and there is a whole block of them appearing together.
• it's a NodePropertyWasSet on the node in question, in which a property is unset! that might be a clue at how these events happened. our thought was maybe a node migration but that should have happened in a workspace with a publish, but we don't see such events...
• the node in question did exist correctly in live, BUT was indirectly removed by a grandparent removal, therefore no NodeWasRemoved appears anywhere yet the node is simply not there, we found the respective event removing the node but it is far before the offending event.
• how did the NodePropertyWasSet event ever get into the DB given that the node it applies to cannot have existed at the time the even was created and thus a constraint check should have prevented it.

---

So far we can explain the problem due to a bugfix in NodeAggregateWasRemoved and the use of structure adjustments before:

There is one known case which emits events without metadata: structure adjustments
Two of these use \Neos\ContentRepository\StructureAdjustment\Adjustment\PropertyAdjustment::removeProperty which would cause to immediately PUBLISH aNodePropertyWasSet without metadata and no constraint checks.

The query used for structure adjustments is findNodeAggregatesByType, but it behaves well if the projection is okay which does not seem to be:

We tested that a change in beta15 #5320 which fixes NodeAggregateWasRemoved to correctly delete previously orphaned nodes (commit 9df5948) causes this problem to be noticed.

This also fixes a bug in NodeRemoval where only one child node was removed per removed parent

---

As a migration to fix this, we can introduce possibly a NoopEvent that is will be inserted at the places of the malformed events.
To automate this (super expensive but i see no other way) we need to replay the content graph one event at a time, or as much until we get the errror? And then get the event, see if it has no metadata, and then replace it with a NoopEvent and continue the replay.

Alternatively we would need to have the graph projections logic in memory just to keep track which node was removed and which not.

The NoopEvent could actually reduce complexity as WorkspaceWasRenamed and other legacy events that are just noops could be transformed into these as well and thus removing legacy code.

---

That we dont set metadata or correlation infos always MUST be addressed before release:

• see 9.0: Correctly set Correlation and Causation IDs #3887
• wip pr: BUGFIX: Fix and test tethered node structure adjustment #4969

---

See also our slack discussion: https://neos-project.slack.com/archives/C04PYL8H3/p1730799833465089