fix(agent): race in readWithDeadline clobbers conn deadline (#7)

* refactor(docs): inline pandoc build into Makefile, drop scripts dir

* fix(agent): wait for deadline watcher before returning

readWithDeadline's watcher goroutine could clobber the conn's read
deadline with time.Unix(1, 0) after the main function reset it to zero.
When the Accept ctx was canceled shortly after Accept returned, the
watcher raced with close(done) in select and sometimes picked ctx.Done()
even though we were already done reading, leaving the conn unusable for
the next read (instant i/o timeout on step 1 snapshot).

Synchronize on the watcher's exit before resetting the deadline so it
can never override the reset.

* test(agent): cover readWithDeadline race on Accept ctx cancel

Drives Accept with a short-timeout ctx, cancels it right after Accept
returns, then does a Snapshot. Reliably fails without the readWithDeadline
synchronization fix (watcher goroutine overwrites the deadline to past).

internal/agent/server.go

@@ -92,11 +92,11 @@ func (c *Conn) Close() error {
 func readWithDeadline(ctx context.Context, conn net.Conn) (Message, error) {
 	if deadline, ok := ctx.Deadline(); ok {
 		_ = conn.SetReadDeadline(deadline)
-		defer conn.SetReadDeadline(time.Time{})
 	}
 	done := make(chan struct{})
-	defer close(done)
+	exited := make(chan struct{})
 	go func() {
+		defer close(exited)
 		select {
 		case <-ctx.Done():
 			_ = conn.SetReadDeadline(time.Unix(1, 0))
@@ -104,6 +104,9 @@ func readWithDeadline(ctx context.Context, conn net.Conn) (Message, error) {
 		}
 	}()
 	message, err := ReadMessage(conn)
+	close(done)
+	<-exited
+	_ = conn.SetReadDeadline(time.Time{})
 	if err != nil && ctx.Err() != nil {
 		return Message{}, ctx.Err()
 	}