TopbarSweeper (`app/services/topbar_sweeper.rb`) uses `include Singleton` with an `attr_accessor :controller` that is shared across all Puma threads within each worker process. Under concurrent requests, one thread's `ensure` block in `set_topbar_sweeper_controller` nils the controller while another thread is actively using it, causing:

```
NoMethodError: undefined method `expire_fragment` for nil:NilClass
at topbar_sweeper.rb:6
```

The existing `if controller.present?` guard has a TOCTOU (time-of-check to time-of-use) race: the check passes, then another thread nils the value before `expire_fragment` is called.

The bug is hard to reproduce under typical load because the Singleton is per-process (each Puma worker has its own), so only threads within a single worker compete. With the default 5 threads per worker, the race window is extremely narrow. It becomes observable under high concurrency (many simultaneous registrations) or with increased threads per worker, where more threads share the same Singleton and the interleaving becomes likely.

Introduced in commit 0e52ebb70f57 (Redmine #19035) when migrating from rails-observers for Rails 5 compatibility.

Fix: replace the Singleton with `Thread.current` storage, following the same pattern already used by `ThreadSession` for `User.current` / `Organization.current`.