A11y tree is empty in a Python trace: no per-action element snapshots #341

Part of #330

A Python trace has no A11y tab content and no element overlay. Measured on two
archives from the same login flow:

The DOM snapshots themselves are there — the tab is empty because nothing
produces the per-action element data beside them. core/action-snapshot.ts
captures it for the JS adapters through core/element-scripts.ts, and
selenium-devtools/src/action-snapshot.ts wires it in; the Python adapter has
no equivalent and never sends one. grep -rn "a11y\|elements.json" over
selenium_devtools/ returns nothing.

What this needs

The element scripts run in the page, so they are not a per-language port —
packages/script is already common to every adapter and the Python adapter
already injects it. What is missing is the adapter-side capture: run the
element-snapshot script per action, and send the result under a scope the
exporter reads. actionSnapshots is a defined TraceLog scope that no adapter
sends today (core synthesizes from command screenshots when it is absent), so
this and #290 are the same shape of gap — a defined scope with no producer.

Whether the capture belongs in the Python adapter or behind the wire follows
the #298 decision: transforms are shared, capture is per-language. Reading the
page needs the live driver, so it stays in-language.

The overlay is a separate, larger gap

The A11y tree is cross-framework in player mode. The overlay — drawing
boxes on the replayed page — is still WDIO-only, and needs per-command locators
plus XPath rather than an elements.json reader. Out of scope here; this issue
covers only the tree.

Acceptance criteria

A Python trace carries *-elements.json beside its frame snapshots, and the
A11y tab renders a tree for a run, matching what the JS Selenium adapter
produces for the same flow.

Size: M. Depends on #288.