test: stop the e2e stack broker deadlocking on close

[frahlg]

go/test/e2e/stack_test.go carries the same MQTT-broker shutdown race that
#706 fixes in go/cmd/sim-ferroamp/e2e_test.go, and this is the same fix.

The race

mochi-mqtt v2.7.9 Clients.GetByListener (clients.go:91) takes RLock and then
calls Clients.Len, which takes RLock again. sync.RWMutex is not reentrant:
a Clients.Delete write lock landing between the two read locks blocks the
second one, while the writer waits for the first. Both goroutines stop for good.

Server.Close walks GetByListener, and the broker runs Clients.Delete the
moment a client drops. stack.Close calls reg.ShutdownAll(), which disconnects
the Ferroamp driver's MQTT client, and then closes the broker — exactly that
race. Confirmed from a CI goroutine dump on the sibling test:
https://github.com/srcfl/ftw/actions/runs/30378575736 (attempt 1, job "core (Go)").

Nothing bounds it here but make e2e's -timeout 180s, so a hit would burn
three minutes and fail an unrelated pull request.

The fix

Capture the broker's client count as a baseline right after it starts, and close
through closeBroker, which waits (bounded, 2s) for the count to fall back to
that baseline before calling Close. Close then runs behind a 5s deadline that
fails the test with a clear message if it still wedges.

Both writers on Clients are quiet at baseline, which is what makes the wait
sufficient rather than merely likely to help:

• server.go:491 (disconnect path) — Len() only drops once Delete has
  returned, so baseline means no disconnect is in flight. Nothing reconnects
  after ShutdownAll.
• server.go:1711 (clearExpiredClients, a 1s ticker in eventLoop) — skips
  clients whose StopTime() is zero. At baseline only the inline client
  remains, and it is never stopped, so the sweeper issues no Delete at all.

New() registers the inline client (server.go:200-202), so the baseline is
already settled where it is captured and does not race Serve().

v2.7.9 is the newest release, so there is no upgrade that fixes this.

Verification

make e2e passes; forced uncached run at 25.7s. make verify clean.

An instrumented run reported baseline=1, already at baseline on entry to
teardown, settling in 24µs — the guard costs nothing on the happy path. The
instrumentation is not part of this change.

Notes for review

• This deliberately does not touch go/cmd/sim-ferroamp/e2e_test.go, which
  belongs to open test: stop the sim-ferroamp broker deadlocking on close #706. The helper therefore exists in two copies, each
  cross-referencing the other in its comment. There is no shared test-helper
  package to host it, and creating one would mean editing the file test: stop the sim-ferroamp broker deadlocking on close #706 owns.
• No changeset: test-only, nothing user-visible ships. test: stop the sim-ferroamp broker deadlocking on close #706 has the same shape
  and used the no-changeset label, applied here too.

🤖 Generated with Claude Code