MUST vs is/are

In the previous meeting we had (once again) discussed when to use MUST 
and the other 2119 terms [1]. Looking through our use of these terms I 
find it rather difficult to make this 100% consistent. Could I get some 
input on how to reformulate

- Each value of <code>sh:targetClass</code> must be an <a>IRI</a>.
- The values of <code>sh:minCount</code> must be literals with datatype 
<code>xsd:integer</code>
- Validation results must have a single value for the property 
<code>sh:focusNode</code>

In the latter case it could either be understood to be implementation 
advice (an engine must produce a single value...) or a language 
description. And the same applies to all other cases. I find this whole 
topic completely arbitrary and unhelpful.

The SPARQL 1.1 spec doesn't use MUST at all. Could we just do the same?

Holger

[1] https://tools.ietf.org/html/rfc2119