CHANGELOG: map session and churn fixes

CHANGELOG.md

@@ -147,6 +147,11 @@ the ACME DNS-01 challenge through a built-in DNS provider: `acmedns`,
 - Fix the unix socket being created `0700` instead of the documented `0770` when `unix_socket_permission` is unset. The socket grants full admin access without authentication, so members of its group gain that access; check who is in it before upgrading [#3540](https://github.com/juanfont/headscale/pull/3540)
 - A policy write builds the peer map once instead of twice [#3557](https://github.com/juanfont/headscale/pull/3557)
 - Speed up the peer map build on large tailnets: each node's matchers are found once instead of checking every pair of nodes against every rule [#3556](https://github.com/juanfont/headscale/pull/3556)
+- Map sessions that fall behind during update bursts are no longer dropped; a session is only closed once it stops reading or a newer session of the same node replaces it [#3575](https://github.com/juanfont/headscale/pull/3575)
+- Node joins and deletions under churn, especially in tailnets with many tags, no longer stall registrations, map updates and API requests; ephemeral nodes are cleaned up within seconds instead of minutes [#3575](https://github.com/juanfont/headscale/pull/3575)
+- Clients no longer report an `INTERNAL_ERROR` stream error when their node logs out or is deleted, or when headscale shuts down [#3575](https://github.com/juanfont/headscale/pull/3575)
+- Shutdown keeps the LastSeen of connected nodes: it waits for their map sessions to close before closing the database [#3575](https://github.com/juanfont/headscale/pull/3575)
+- Deleting an ephemeral node no longer logs a spurious `ephemeral node deletion failed` error later [#3575](https://github.com/juanfont/headscale/pull/3575)
 
 ## 0.29.5 (2026-10-13)