You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Selector or count rebuild after a sub-widget write replays differently by apply order #1101
A batch that writes a DynamicCombo sub-widget and then switches the selector away and back replays to different graphs depending on apply order.
On a Gen node already at model = max:
specs = [model.mode = quality, model = base, model = max]
Forward replay: the model = max rebuild refills model.mode with its default, so the node ends at balanced.
Reverse replay: model = max wins first and model = base loses the gate, so no rebuild happens. model.mode = quality still applies, so the node ends at quality.
The rebuild resets the sub-widget's value but leaves its _widget_stamps entry alone, so the earlier sub-widget write keeps winning its gate in one order and not the other. A DynamicGroup shrink-then-regrow in one batch has the same shape.
With #1007 the ops in a batch are minted in increasing order, so this shape diverges every time instead of depending on random op_ids. #1007 narrows its amendment's convergence claim and leaves this open.
Possible fixes: have a selector or count rebuild bump or reset the stamps of the sub-widgets it replaces, or refuse a batch that writes a sub-widget before rebuilding its selector or count.
Activity