Relied on the existing vendored partial-update library as an integration point. The new script cancels speech on its pre-swap event and on page hide so audio does not continue across content swaps. No new dependency was added.
- What worked
- Documented pre-swap event made cleanup straightforward without backend changes.