Keep keyed children when their wrapper container goes away

[maartenbreddels]

Why

When a container is no longer in the new tree, the stale-key sweep removes it together with all its children, by key. An explicit .key() is the same in any container of a component. So a keyed child that is still in the tree at another level was removed too. For example:

child = Child().key("x")
w.VBox(children=[w.VBox(children=[child]) if wrapped else child])

When wrapped becomes False:

• default renderer: crashes with AttributeError: 'NoneType' object has no attribute 'comm_id'
• fast renderer: shows a closed widget in the tree, or leaves elements behind

This is the same kind of bug that #65 fixed for a container that is replaced by one of another type. The crossreview of #65 found this case, where the container is removed.

What

Two commits:

1. The fuzz test finds it. An Unwrapper component puts a child with an explicit key inside 0, 1 or 2 wrapper containers, depending on its state. The child is sometimes the same object (made once with use_memo) and sometimes new. The keyed children that move (in Mover and Unwrapper) now have an effect, and the effect log at close() is compared too. On 1.10.4, 2 of the 12 seeds fail, and 73 of 300.
2. The fix, in both renderers.
   • The stale-key sweeps now remove through the same path as Keep keyed children that move out of a replaced container #65. It is renamed from _remove_replaced to _remove_outgoing. Children with an explicit key that the new tree still uses stay, and reconciliation updates them where the new tree has them.
   • The fast renderer's early cleanup of stale effects (from Fast renderer: clean up removed components before new effects run #68) follows the same rule, so a kept child keeps its effects.
   • Shared elements keep the old behavior, and close() still removes everything.

Visible change: a keyed child whose wrapper goes away now keeps its widget and its state.

Tests

• Fuzz test, including effects and the close log, 1000 seeds with 25 steps each (a local run; the PR keeps 12): all pass. On 1.10.4, 73 of 300 fail.
• test_keyed_child_out_of_removed_wrapper (child as the same object, and as a new object): fails on 1.10.4 in both renderers.
• test_keyed_child_out_of_removed_wrapper_keeps_its_effects: the effect runs once and is cleaned up once, at close().
• pytest reacton/ with REACTON_FAST=0 and =1: all pass.
• Solara's unit suite: all 496 pass, in both renderers.

Review

Crossreview by three reviewers (astra, opus, glm), two rounds.

• Round 1: all three found the same bug in the first version. The fast renderer's early stale-effect cleanup still cleaned up the kept child's effects, so the child stayed mounted with dead effects. The fuzz test missed it, because the moved child had no effect. This was fixed with the rule above, and the fuzz test and a unit test now cover it.
• Round 2: all three approve, with no new findings. They checked:
  • that self.context owns the removed element at each sweep
  • that used_keys is the right "still in the tree" signal
  • key reuse by another type
  • moves into and out of child components
  • a second render pass
  • error boundaries
  • that close() still cleans up everything
• Not fixed here, older bugs: .shared() elements have several bugs, and they fail the same way on master.

🤖 Generated with Claude Code