Register the broker wait before the first poll (closes #157)

GET /api/v1/messages read the client output queue and only then
registered with the broker. Notify wakes only the channels registered
at that moment, so a message queued between the empty read and the
registration waited out the whole long-poll timeout. The handler now
registers first and removes its channel on every exit path; the IRC
relay loop registers before each drain in the same way.

The handlers read the queue through a function field that defaults to
Database.PollMessages, so tests can act just after a read: one checks
that a message queued there comes back at once, others that no wait
channel stays registered however a request ends.

Model: opus-5-5

internal/handlers/handlers.go

@@ -64,6 +64,16 @@ type Handlers struct {
 	loginLimiter  *ratelimit.Limiter
 	stats         *stats.Tracker
 	cancelCleanup context.CancelFunc
+
+	// pollMessages is Database.PollMessages, with which
+	// HandleGetMessages reads the client output queue. Tests
+	// replace it to act just after a read, or to make a read
+	// fail.
+	pollMessages func(
+		ctx context.Context,
+		clientID, afterQueueID int64,
+		limit int,
+	) ([]db.IRCMessage, int64, error)
 }
 
 // New creates a new Handlers instance.
@@ -95,6 +105,7 @@ func New(
 		hashcashVal:  hashcash.NewValidator(resource),
 		loginLimiter: ratelimit.New(loginRate, loginBurst),
 		stats:        params.Stats,
+		pollMessages: params.Database.PollMessages,
 	}
 
 	lifecycle.Append(fx.Hook{