Gaps between specification and implementation in typed_value::composing()

## Description and Reproduction

Consider [Example 63.2](https://theboostcpplibraries.com/boost.program_options#ex.program_options_02) ([Execution on Wandbox](https://wandbox.org/permlink/UCyNgHUbG5egtL6S))

> test --phone 123 --phone 456

In this case, both phone numbers, 123 and 456, are parsed. The call to composing() makes it possible to use a command-line option multiple times – the values are composed.

From this description, it would seem that, conversely, without the call to `composing()`, only 123 would be assigned to the `--phone` option but 456 would be ignored. (*I assume this is the specification. Is my understanding correct?*)
- It seems that there are others who have a similar understanding:
- https://boost-users.boost.narkive.com/oyPRDQxi/program-options-documentation-dead-end#post2
- https://stackoverflow.com/questions/8175723/vector-arguments-in-boost-program-options

However, even after removing `composing()`, both 123 and 456 are assigned to `--phone`. ([Execution on Wandbox](https://wandbox.org/permlink/jCoYvofxP4GsMsDE))

On the other hand, in order to confirm the effect of `composing()` in the current implementation, when `store()` is called twice, we can see that the result changes depending on whether `composing()` is called or not.
- Twice `store()` with `composing()`: [Execution on Wandbox](https://wandbox.org/permlink/3FghGNuoJwwQBWYW)
- Twice `store()` without `composing()`: [Execution on Wandbox](https://wandbox.org/permlink/L9JAcE0OxFzXjLGg)

## Implementation Analysis

From what I've seen of the implementation of Boost 1.81.0, these lines are suspect.
- In the first for loop to proceed the parsed options ([L48-L93](https://github.com/boostorg/program_options/blob/boost-1.81.0/src/variables_map.cpp#L48-L93))
- The condition for skipping assignment depends on `xm.m_final`. ([L67-L68](https://github.com/boostorg/program_options/blob/boost-1.81.0/src/variables_map.cpp#L67-L68))
- If `is_composing()` is false, `new_final` is modified. ([L85-L91](https://github.com/boostorg/program_options/blob/boost-1.81.0/src/variables_map.cpp#L85-L91))
- The comment here clearly states that `composing()` affects subsequent calls to `store()`. I think this statement has a gap with that in Example 63.2.
- After the first for loop, `xm.m_final` is modified by `new_final` ([L102](https://github.com/boostorg/program_options/blob/boost-1.81.0/src/variables_map.cpp#L102))

The intuitive image from the description of Example 63.2 is an implementation in which:
- Deleting L102.
- Modifying L91 to `xm.m_final.insert(option_name)`

## Summary

- There seems to be a gap between the specification (which I can intuitively read from Example 63.2) and the implementation.
- I would like to clarify if my understanding is wrong or if either the description of Example 63.2 or implementation is incorrect.

## Version

- Boost 1.81.0

Start with src/variables_map.cpp in Boost 1.81.0, especially the first parsed-options loop at lines 48-93 and the xm.m_final update at line 102. Reproduce the single-store and repeated-store cases using the linked Wandbox examples, with and without composing(). Done means determining whether Example 63.2, its documentation, or the implementation is inconsistent and clearly documenting the conclusion.

Written by the indexing model from the issue text.