Version 0.5.0

This release adds port and edge value inspection, connection checks
and canvas drops for interactive graph editing. Node and Edge properties now sync to
the browser in place, batches of changes render as a single update, and
an error boundary keeps a malformed graph from blanking the canvas.

Highlights

• Port and edge inspection: handle and edge click events can open a
  Panel popup with flow.show_popup(content, position). Set
  popup_trigger="hover" to use hover events instead; use
  flow.close_popup() to dismiss it. Applications supply the value to
  display, and examples/port_value_inspection.py shows both interactions
  (#74).

• Connection validation: opt-in browser policies reject connections
  with the wrong direction or handle type, full inputs, duplicate edges,
  and cycles. Register Python rules with add_connection_validator() to
  show rejection reasons on candidate handles during a drag. Python
  validators check interactive drags only; they do not restrict
  flow.add_edge() or direct changes to flow.edges. See
  Validate Connections and
  examples/connection_validation.py
  (#75).

• Error recovery: the canvas is now wrapped in an error boundary,
  controlled by the new error_recovery parameter ("auto" by default,
  or "manual" / "off"). In "auto" mode a React render error
  remounts the canvas, then remounts again in a view-only safe mode that
  repairs or hides elements it cannot render (invalid positions, unknown
  node/edge types, dangling edges, duplicate or missing ids) without
  mutating the server-side graph. If retries are exhausted a recovery
  panel offers Try again, Reload page and Copy details. Every
  error is logged to the panel.reactflow logger and emitted as a
  client_error event, so browser-side failures are no longer invisible
  to the server
  (#70).

• Base property sync: the top-level React Flow fields on Node and
  Edge (label, type, style, className, draggable,
  connectable, deletable, ...) are now synced to the frontend, so
  assigning node.label = "Start (running)" patches the browser in place
  instead of requiring flow.nodes to be replaced. Parameters declared
  on a subclass continue to sync into data. New patch_node_props()
  and patch_edge_props() methods do the same for dict-based nodes and
  edges, where passing None clears a field back to the CSS/theme
  default, and new on_props_change hooks fire on Node/Edge
  subclasses when the frontend changes a property. position and
  selected remain browser-owned during drag and selection and are only
  pushed via patch_node_props()
  (#71).

Enhancements

• Batch removal: remove_node() and remove_edge() now accept
  several ids, either as separate arguments
  (flow.remove_node("n1", "n2")) or as a sequence
  (flow.remove_node(["n1", "n2"])), and remove them in a single update.
  For any other batch of changes made from Python, wrap them in
  pn.io.hold() to render them at once
  (#72).

• Drop data onto the canvas: drop_types lists MIME types the canvas
  accepts, and each drop of one of them emits a drop event with the
  dropped data, its flow position, and the node or handle under the
  pointer, so a handler can add and wire what was dropped. Draggable
  panel_material_ui.MenuList items work as a source; see
  React to Events
  (#77).

Bug fixes

• Nodes remounted when node types change: any change to
  node_types rebuilt the component for every node type, so React
  remounted every node and detached its embedded view, which could leave
  Bokeh plots blank. Each type's component is now reused until its own
  spec or the popup settings change
  (#77).

• Progressive re-render when deleting multiple elements: deleting a
  multi-node selection removed the nodes one at a time, syncing an
  intermediate graph to the browser per node so the nodes visibly
  disappeared one by one. Updates triggered by a frontend message are now
  held and combined into a single patch, and node/edge deletion assigns
  nodes and edges once
  (#72).

• Serialization error when updating node_types: assigning
  NodeType or EdgeType descriptors with a schema class after
  construction synced them to the browser before they were converted to
  JSON descriptors, so the server failed with can't serialize ParameterizedMetaclass. They are now converted before syncing, and
  updating node_types and edge_types together no longer raises a
  TypeError
  (#76).