When update-main pushes a swap into a nested copy in the main, and the
position there holds the same component, the pushed head becomes a plain
copy of that position. Its content differences become overrides while its
placement follows the position, but its descendants kept the places of the
pushed copy with no geometry override, so a later sync would move them.
They now take the places the position gives them, except the subtrees that
author their own geometry.
Generated regression: an eleven-step add-variant, switch and update-main
program that failed I4 (seed 256567394, round 8) now passes, as does the
original program.
Benchmark (scripts/bench-components, ms, lower 2.5% quantile):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+---------+---------+--------|
| deepest-edit-through-three-levels | 35.215 | 23.622 | 27.685 | 0.67 |
| edit-through-a-chain-of-dependents | 68.931 | 48.952 | 50.286 | 0.71 |
| library-update-into-copies | 83.083 | 54.647 | 59.195 | 0.66 |
| main-edit-among-many-shapes | 4.837 | 3.23 | 3.376 | 0.67 |
| main-edit-into-10-copies | 1.79 | 1.152 | 1.196 | 0.64 |
| main-edit-into-100-copies | 15.415 | 11.605 | 11.945 | 0.75 |
| main-edit-into-1000-copies | 153.086 | 110.919 | 114.366 | 0.72 |
| variant-switch-into-copies | 32.052 | 24.905 | 26.59 | 0.78 |
AI-assisted-by: claude-opus-5-5