Add opt-in fast render context (REACTON_FAST) + benchmark harness

The default renderer is unchanged: _RenderContext keeps the original
tree-walking (_render/_reconsolidate/_remove_element/_visit_children*),
byte-for-byte. A _RenderContextFast(_RenderContext) subclass overrides only
those methods with a reimplementation that does work proportional to what
changed instead of a full tree walk:

- state setters mark needs_render_descendant up the parent chain, so a render
  pass only descends into subtrees that can contain work; an identical,
  fully-reconciled, exception-free subtree is skipped in both phases
  (clean_subtree) and keeps its widgets;
- force_update/update/first render walk fully (rc._walk_all);
- stale-element removal runs once per context instead of an O(n) set-diff per
  element;
- widget updates are skipped when an identical element reconciles to identical
  child widgets.

Selected with REACTON_FAST=1 (via _render_context_class()); off by default.
Side-effect (Layout/Style) widget tracking also moved to ipywidgets'
on_widget_constructed hook (shared by both renderers) instead of a per-creation
global-dict diff that was O(live widgets).

benchmarks/ holds the harness, baseline-vs-fast result JSONs, and README.md
documenting the design, the renderer contract both implementations satisfy, the
deliberate test-invisible deltas, and known issues found during the work. On a
300-row tree the fast renderer takes a single-leaf update from 13.7ms to 0.62ms
(22x) and a memoized-subtree skip from 14.4ms to 1.0ms (14x), behavior-identical
to the default.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: maartenbreddels

benchmarks/README.md

@@ -0,0 +1,182 @@
+# Reacton render benchmarks & the opt-in fast reconciler
+
+This directory holds the performance harness for a faster reimplementation of
+the renderer (the tree-walking downstream of `render()`), plus its design notes.
+
+The fast renderer is **opt-in and off by default.** The default `_RenderContext`
+is the original, unchanged implementation; the rewrite lives in a
+`_RenderContextFast(_RenderContext)` subclass that overrides only the
+tree-walking methods. Set `REACTON_FAST=1` to select it. CI runs the full test
+suite against both (a `reacton-fast: ["0", "1"]` matrix dimension), so the two
+stay behavior-identical.
+
+## Running the benchmarks
+
+```bash
+# default renderer, then the fast one:
+python benchmarks/bench.py --label baseline
+REACTON_FAST=1 python benchmarks/bench.py --label fast
+python benchmarks/compare.py baseline fast
+```
+
+`bench.py` writes `benchmarks/results/<label>.json`. `--scenario <name>` runs a
+subset; `--profile` also writes a cProfile `.prof` per scenario. Scenarios cover
+initial render (wide/deep), single-leaf update, root update, memoized-subtree
+skip, keyed list reorder, burst updates, full `force_update`, and teardown — so
+a change can be checked against both the headline case (one leaf updating in a
+large tree) and the no-op/​full-walk cases.
+
+Use the cheapest widgets possible in scenarios (bare `Button`/`Label`): real
+ipywidget construction cost can dominate and mask reconciler cost, so confirm
+with a profile that you are measuring tree-walking, not traitlets.
+
+## What the fast renderer changes
+
+Goal: an update should cost work proportional to what actually changed, not a
+full tree walk. With the default renderer, a single-leaf state change, a
+memoized-subtree "skip", and a full `force_update` all cost the same (~14ms on a
+300-row tree), and stale-element removal is accidentally quadratic. The fast
+renderer (`_RenderContextFast`, `REACTON_FAST=1`) addresses both:
+
+- **Dirty-subtree skipping.** State setters mark `needs_render_descendant` up
+  the parent chain, so a render pass only descends into subtrees that can
+  contain work. A component subtree whose element is identical to the previous
+  render (`el is el_prev`), is fully reconciled, and has no dirty/excepted
+  contexts is skipped in *both* phases and keeps its previous widgets
+  (`clean_subtree`).
+- **Forced full walks** (`force_update()`, `update()`, the first render) set
+  `rc._walk_all`, disabling skipping for that pass — faithful to the old
+  behavior.
+- **Stale-element removal** runs once per component context (and once for the
+  root context in `render()`) instead of an O(elements) set-difference per
+  element.
+- **Widget updates** are skipped when an identical element reconciles to
+  identical child widget objects (`_values_identical`), avoiding pointless
+  traitlets assignments.
+- **Side-effect ("orphan") widgets** (Layout/Style created during construction)
+  are tracked via ipywidgets' `on_widget_constructed` hook instead of diffing
+  the global widgets dict per creation — the old diff was O(live widgets) per
+  widget, so it degraded as an app grew.
+
+## Renderer contract
+
+What both renderers must preserve (derived from `core.py` + the test suite).
+The fast renderer overrides only `_render`, `_reconsolidate`, `_remove_element`,
+`_visit_children`, `_visit_children_values`; everything else (Element widget
+create/update/close, hooks storage, `ComponentContext`, exception plumbing, the
+render loop) is shared with the default renderer.
+
+**Phases** (inside one `rc.render()` call, under `thread_lock`, with `local.rc` set):
+
+1. *Render phase*: execute component functions depth-first, build
+   `elements_next`/`children_next`. Repeat while `_rerender_needed` (max 50,
+   then `RuntimeError`) and no exception has bubbled up.
+2. *Reconsolidate phase*: create/update/close widgets, run effects, move
+   `*_next` → current. Loop back to phase 1 if state was set during reconcile
+   (e.g. in an effect). `render()` returns the root widget.
+
+**Keys.** `el._key` or a positional default. A context's root key is `"/"`.
+Children of an element with key `K`: list → `f"{K}{i}/"`, dict → `f"{K}{k}/"`.
+Component child contexts live in `context.children[key]`. A duplicate key in one
+context raises `KeyError`. `el._key_frozen` is set once an element is rendered.
+
+**Render phase, per element:**
+- `el._render_count += 1` (a shared element is visited once; a non-shared one
+  per use — see `test_render_count_element`).
+- Element arguments (args/kwargs) are walked only for `ComponentWidget`
+  elements or shared elements; a `ComponentFunction` element does not walk its
+  arguments.
+- `ComponentFunction`: reuse the `ComponentContext` when
+  `same_component(invoke_element.component, el.component)`, or when a context
+  exists with no `root_element` (pre-created by `state_set`). A different
+  component → fresh context, old one removed during reconcile.
+- `needs_render` = `context.needs_render` (set by setters/force) OR
+  `el._arguments_changed(el_prev)` OR `context.exceptions_children`. If false,
+  the body is *not* executed (component `render_count` stays put), the previous
+  `root_element` is reused, but it is still walked.
+- Body execution resets `state_index`/`effect_index`/`memo_index`,
+  `user_contexts={}`, `exception_handler=False`, `needs_render=False` before the
+  call; wraps in `context_managers` (the solara `ContextManager` hook) and the
+  `_default_container` handling (implicit containers via
+  `ContainerAdder`/`el.__enter__`).
+- Body exceptions → `context.exceptions_self`, set `_rerender_needed`; a
+  hook-count mismatch is itself an error (fewer calls only allowed if an
+  exception interrupted the body); a component returning `None` → `ValueError`.
+- After: `parent.children_next[key] = context`; prune `children_next`/
+  `elements_next` to `used_keys`; `user_contexts_prev = user_contexts`.
+  Exceptions bubble to `parent.exceptions_children` unless `exception_handler`
+  (set by `use_exception` during the body).
+
+**Reconsolidate phase, per element** (depth-first over the element tree,
+entering child contexts):
+- *Widget element*: visit children → new kwargs (Elements replaced by widgets;
+  `FragmentWidget` children flattened into lists); then create / update-in-place
+  (same component) / replace (remove old first). Created via `el._create_widget`
+  (orphan side-effect widgets tracked into `rc._orphans[widget.model_id]`);
+  updated via `el._update_widget` (dropped kwargs restored to trait defaults,
+  listeners removed) under `hold_sync` + `suppress_events`.
+- *Component element*: recurse into the child context root, then process
+  effects: an `effect.next` with equal deps drops `next`; otherwise clean up the
+  old effect and run the next; never-executed effects run now; all are skipped
+  (cleanups still run) if the context has exceptions.
+- `el.meta` / root-widget meta merge → `widget._react_meta`.
+- Bookkeeping: `elements_next[key]` → `elements[key]`; `children_next[key]` →
+  `children[key]`; shared elements move `_shared_elements_next` →
+  `_shared_elements` with the widget in `rc._shared_widgets` (one widget per rc);
+  non-shared widget in `context.widgets[key]`; `context.element_to_widget[el]`
+  set (used by `get_widget`).
+- Stale keys (old elements whose key is not in `used_keys`) → `_remove_element`,
+  in sorted order for reproducibility.
+- After reconciling a context, `elements_next` must be empty (shared leftovers
+  are moved), else `RuntimeError`. `root_element = root_element_next`;
+  `root_element_next = None` (asserted by `test_internals`).
+
+**Removal.**
+- *Component element*: clean up all effects, recurse into `root_element`, assert
+  the context is fully empty, `del parent.children[key]`; cleanup exceptions
+  bubble (`close()` raises the first).
+- *Widget element*: recurse children first, close orphans,
+  `el._cleanup_callbacks(widget)`, `el._close_widget(widget)`. Shared widgets are
+  removed only on the last reference.
+- Removing an already-removed key is a no-op (`test_remove_element_twice`).
+
+`close()`: `_remove_element(root)`, container closed (+layout), asserts
+`_shared_elements`/`_orphans` empty, raises any pending exception.
+
+**Other pinned behavior:** `rc.render_count` increments per `render()` call;
+root widget type change without a container → `ValueError`; container gets
+`.children = [widget]` (`[]` on error). `handle_error=True` renders an HTML
+traceback widget instead of raising. Batching: `with rc:` defers re-render, and
+`hold_trait_notifications` is wrapped so frontend-triggered trait updates batch
+into a single render. Setters use `eq`/`utils.equals`, warn on mutated
+list/dict/DataFrame state, are ignored while `_closing`, and trigger a render
+unless already rendering or batching. `use_context`/`provide` walk parent
+contexts; `provide` notifies listeners only when the value changed.
+
+## Deliberate behavior deltas (not observable by the test suite)
+
+- Stale elements are now closed *after* the parent widget gets its new children
+  assigned (the old code closed them mid-walk, so a closed widget could briefly
+  remain in a container's `children`).
+- Per-element debug logging in the hot paths was dropped.
+
+## Known issues worth revisiting (found during the rewrite, not fixed here)
+
+- **Effect/cleanup exceptions** are recorded on the *parent* context's
+  `exceptions_self`, while render exceptions go to the component's own context.
+  So an effect exception is caught by `use_exception` one level higher than a
+  render exception would be. Behavior preserved from the old code; looks
+  unintentional.
+- **pandas 3.0**: setting state to an equal DataFrame triggers a re-render and
+  leaks widgets (`test_set_state_with_dataframe`). The reacton equality logic
+  likely needs a pandas-3 fix; the test environment pins `pandas<3` for now.
+- `ComponentContext.owns` is dead — never written, only asserted empty in
+  `_remove_element`.
+
+## Where initial-render time goes
+
+After the rewrite, initial render is dominated by ipywidget *construction*, not
+reconciler tree-walking — most of it is traitlets default materialization and
+`get_state` during `Widget.open()`. Reacton-side reconciliation of a large tree
+is now a small fraction of that. Future initial-render wins are mostly in
+ipywidgets/traitlets, not here.