Actors, scenes, and lipsync

A scene-launch failure can occur before any subtitle, audio, or visualizer appears. Diagnose actor readiness, acquisition identity, scene graph entry, and indexed resource tables in that order.

Locate the launch boundary

Record the last confirmed event and the first missing event:

community Activate
  -> CharacterSpawned passed
  -> broad setup trigger passed
  -> scene node entered
  -> scene entry resolved
  -> actor acquisition
  -> section/event start
  -> subtitle / WEM
  -> named scene exit

If the game crashes before dialogue visualization and the crash follows the scene launch trigger when that trigger moves, focus on scene initialization. If bypassing the scene lets the world/community route continue, do not rebuild the world sector first.

Actor readiness workflow

For each required actor, compare all identity fields:

Then test this order:

1. activate the community;
2. wait for CharacterSpawned on the exact community NodeRef/entry scope;
3. enter a broad setup boundary;
4. checkpoint if the tested pattern requires it;
5. start the scene;
6. let narrower scene conditions control approach beats.

A single fan-out that activates the community, waits, and starts the scene does not enforce that order. Starting only at a narrow engage boundary can also make several already-true scene conditions cascade in one tick.

Readiness and lipsync are independent. Legacy research repaired a fast-approach readiness gap, yet a deterministic launch crash remained until the lipsync table was corrected.

Audit actor and performer joins

For every spoken line or section:

1. confirm the screenplay item’s speaker and addressee are valid actor IDs;
2. confirm the section’s actor behavior rows use the intended performer IDs;
3. confirm debug-symbol performer mappings agree with the actor table;
4. confirm a community actor’s area NodeRef and entry name match the active registry entry exactly;
5. confirm the scene marker exists and the quest scene node uses the correct entry point CName;
6. confirm the quest expects an actual named scene exit, not a guessed end.

Actor ID, performer ID, screenplay item ID, localization RUID, event ID, and CR2W handle ID are different domains. A numerically plausible reuse is not a join.

Audit lipsync cardinality

For each actor, record lipsyncAnimSet.id. Then inspect the cooked scene’s native resouresReferences.lipsyncAnimSets array:

actor slot IDs: 0, 1
addressable rows after cooking: 1
result: invalid index 1

Require every selected index to be addressable. Two identical depot paths in the authoring view may collapse to one import; inspect the round trip and cooked imports, not only the source list.

The safest production correction is to provide the intended distinct, valid NPC and V resources and prove both rows remain addressable after cooking. A shared slot can isolate a cardinality fault, but it leaves facial and gesture quality Experimental unless separately accepted.

Head deformation during speech

If the actor looks normal before a line but stretches while speaking, inspect the selected facial clip before rebuilding the appearance. A clip can import and match the audio duration while lacking the skeletal reference its additive animation needs. Check facial keys, neutral skeletal channels, rig/track ordering, and the export from the cooked asset. See Custom clip checks for the fields to compare.

If the defect is also present at idle, investigate appearance, skinning, parent bindings, and rig compatibility. Treat a faint surface mark separately from severe animated mesh deformation; screenshots alone may not distinguish hair or cyberware clipping from a rendering artifact.

Reduce a crashing scene without losing causality

Use staged candidates, one change each:

1. Keep world/community flow and replace the scene invocation with direct phase continuation. If that passes, the earlier world path is not the immediate crash boundary.
2. Restore scene invocation with Start-to-End only. If it passes, root scene loading and entry/exit can work; actor readiness under full dialogue is not yet proved.
3. Restore one actor-acquired section without audio/choices.
4. Verify resource-table cardinalities and actor joins.
5. Restore the line and external localization/audio path.
6. Add choices, animation, interruption, and cleanup as separate surfaces.

Hash each candidate and reuse the same eligible starting-save lineage. Do not call a bypass “fixed”; it only locates a boundary.

Symptom routing

Previous: NodeRefs, streaming, and placement. Next: Journal, localization, and audio.