The problem, mentionned first, with Bidi can be summarized in the two following situations. Sometime you have two (inline) different directional runs, A and B say, which are logically adjacent (thus to be displayed B after A), but:

1- Visually, A breaks B in two parts, and shouldn't.

By UBA, a 1st part of B, the same directional run, sticks to A, the adjacent (previous) directional run.

Example (your example-2): A is "PIZZA" and B is "- 3 review"

In LTR context

    PIZZA - 3 review

results in

    3 - AZZIP review

i.e.,  "- 3" (part of entity B) sticks to "PIZZA" (A). However, You want "- 3" not to separate from the rest of B.  

The problem here is:

Some directional run breaks in two, and shouldn't.

Yes, you can argue that one may consider "- 3" as part of directional run A, so there is no B to break. This falls in the second situation just bellow.

2- Visually, you want, B say, to be broken in two parts and it doesn't do.

By UBA, B is displayed at once, continuously. Let A be "ABC" and B be "def ghi"

Example:  In RTL context, you want "ABC def ghi" to be displayed

    ghi def CBA

instead of actually

    def ghi CBA

i.e., where "ghi" sticks to it's (original) directional run.

The problem here is:

The visual display becomes a bit complicated, when punctuation or mirrored symbols interfere with the affair.

"ABC (def] ghi" results "def] ghi) CBA"

Examples 3 and 4 fall in this categories. Example 4 is a little more subtle, since the > sign falls in between the two same directional  runs that we want to separate.

Figurative summary of the two problems

1- Some directional run "breaks" in two parts, and shouldn't. Consider a LTR context.

logical:
           A        B
        -------> =======>
desired visual:
           A       B
        <------- =======>
obtained :
         B1      A     B2
        <=== <------- ====>

2- Some directional run should "break" in two parts, and  it doesn't. Consider a RTL context.

logical:
           A        B
        -------> =======>
desired visual:
         B2    B1       A
        ====> ====> <--------
what you get:
         B1  B2       A
        =======> <--------

I found that the notion of "self-contained entity" SCE is what gives some semantic to the notion of "directional run" DR.

In the first example above, there is clearly a SCE, which is "- 3 review", and thus should be kept joined into one and same DR. And in the second above, there are two SCEs in the same DR, which are "def" and "ghi", and thus should give rise to two DRs to be displayed one after the other in the context of the base direction.

Hence the current solution  to add LRM / RLM such that:

1. to have a DR not to sticks to the wrong one (e.g. add a LRM after a RTL DR)
2. to have a DR "break" in two DRs by inserting a mark (the opposite) between them (e.g. RLM between two LTR DRs)