Fix get_widget for a child element of a memoized component

[maartenbreddels]

Supersedes #48, fixes widgetti/solara#927.

Why

A component creates an element, passes it as a child into a function component, and calls get_widget(el) from an effect. On a rerender where the creating component re-executes (new element object) but the receiving component is skipped because its arguments compare equal by value, the child context still maps the old element object, and get_widget fails with "not found in all known widgets".

Why #48 did not work

Two concrete gaps, not a design problem:

1. fix: get_widget could fail with 'x not found in all known widgets' #48 predates the fast renderer. At its own head its suite is green. Rebased onto master, its _render(..., old_to_new) signature change is not mirrored in _RenderContextFast._render, so REACTON_FAST=1 breaks wholesale.
2. Its substitution stopped at the first element boundary: _visit_children_values does not descend into an element's own args, so HBox(children=[VBox(children=[arg])]) left arg unmapped. New test test_get_widget_fail_on_rerender_nested_widgets covers it.

What

#48's approach, an old-to-new element map threaded through _render, applied to both renderers, with the substitution descending through nested elements, the map dropped once a component actually re-executes, and the fast renderer's el is el_prev fast path disabled while a map is active. The helpers skip when nothing changed and when a component takes no element props.

Tests: #48's three tests, plus the nested-widgets one. 190 passed on both renderers.

Please review with care

• The reused root element is mutated in place (root_element.args/kwargs), as in fix: get_widget could fail with 'x not found in all known widgets' #48. A component that holds an element across renders in use_memo or state and re-emits it could get substituted. No test covers that.
• old_to_new is identity-keyed; equal-but-distinct siblings map by traversal position, which relies on utils.equals implying structural alignment.
• Benchmarks, REACTON_FAST=1, interleaved A/B: force_update_wide, memo_subtree_skip, list_reorder, teardown at parity; root_update about 5% slower, the element-prop walk on each skipped component.

🤖 Generated with Claude Code

[maartenbreddels]

Parking this as a draft for now. Context for picking it up later:

• The bug and the three original tests come from fix: get_widget could fail with 'x not found in all known widgets' #48; the fourth test, test_get_widget_fail_on_rerender_nested_widgets, covers the case fix: get_widget could fail with 'x not found in all known widgets' #48 missed (an element nested one container deeper than the memoized component's direct children).
• What fix: get_widget could fail with 'x not found in all known widgets' #48 was missing, concretely: it predates _RenderContextFast, so its _render signature change is not mirrored there, and its substitution stopped at the first element boundary. Both are addressed here, and the tests pass on both renderers (190 passed, 4 skipped).
• Alternatives considered and rejected: a value-equality fallback inside get_widget (leaves elements / element_to_widget keyed on stale objects, so rc.find() and reconciliation disagree, and it masks real stale-element bugs), and re-executing a memoized component whenever an element argument changes identity (defeats the memoization).
• Open questions before merging: the reused root element is mutated in place (root_element.args / kwargs), as in fix: get_widget could fail with 'x not found in all known widgets' #48; a component that keeps an element across renders in use_memo or state and re-emits it could get substituted, and no test covers that. root_update is about 5% slower on the fast renderer (interleaved A/B; the other benchmarks at parity), the cost of walking element props on each skipped component.

[maartenbreddels]

GitHub did not start the pull_request run for this branch; the full unit-test matrix ran via workflow_dispatch instead and passed, 23 jobs: https://github.com/widgetti/reacton/actions/runs/33753168419