Fast renderer: do not lose handler and cleanup errors

[maartenbreddels]

We want to turn on REACTON_FAST=1 in production, so we audited the fast renderer against the default renderer: a line-by-line review, a fuzzer that compares both renderers (about 18,000 random trees), solara 1.60.2's test suite in both modes, and a production app's test suite. This PR fixes the remaining fast-only bugs that matter. The default renderer is unchanged.

Rebased on #63, which fixed the crash for a replaced parent widget in the same way. This PR is now smaller.

Fixes

• Lost event handler exceptions. A handler (on_click, use_event) that raised while another thread was rendering lost its exception. force_update() only set a flag that the running render then cleared, and nothing marked the path to the component, so later renders skipped it. The handler now marks the ancestors, the same way a state setter does.
• Lost effect cleanup exceptions. A cleanup that raised during unmount (for example after use_exception replaced a subtree) lost its exception, and closed widgets stayed on screen. Marking the ancestors stopped at a stale flag inside the subtree being removed. It now always walks up to the root, which costs only the tree depth.
• A replaced parent with a shared element. The Fuzz the fast renderer against the default one, and fix the KeyError it finds #63 fix did not cover a widget replaced by a .shared() component element. Its arguments are rendered in the same context, so it still raised KeyError. Widget arguments and shared-element arguments now go through one helper (_render_arguments), which uses the _replacing counter from Fuzz the fast renderer against the default one, and fix the KeyError it finds #63.

Tests

• Three new tests in core_test.py. CI runs them with REACTON_FAST=0 and =1. All three pass on the default renderer. On current master they fail on the fast renderer.
• The threaded test checks that the click really happened during the render. It passed 30 of 30 runs in each mode.
• The fuzzer from the audit found no new differences, and the benchmarks do not change.

Astra (GPT-6) reviewed the version before the rebase and approved it, with nothing blocking.

Not in this PR

These fast-mode differences remain, and are deliberate or rare:

• Traits the frontend changed without an on_<trait> handler are no longer reset by unrelated renders.
• Effect cleanup order when one component replaces another.
• force_update() no longer re-applies element kwargs.

A keyed child that moves out of a parent that is being replaced shows a closed widget in both renderers. That is a separate bug.

🤖 Generated with Claude Code