The strategy is to copy the template from the the reader to wrapped stream
until a tag is found. Depending on its value, the appropriate content is copied
to the wrapped stream.
The copy from template is then resumed, repeating this sequence until reading
the complete template.
Tags arg given inside curly brackets:
Valid tags are:
* `{bin}` - Binary name.
* `{version}` - Version number.
* `{author}` - Author information.
* `{usage}` - Automatically generated or given usage string.
* `{all-args}` - Help for all arguments (options, flags, positionals arguments,
and subcommands) including titles.
* `{unified}` - Unified help for options and flags.
* `{flags}` - Help for flags.
* `{options}` - Help for options.
* `{positionals}` - Help for positionals arguments.
* `{subcommands}` - Help for subcommands.
* `{after-help}` - Help for flags.
The largest organizational change is that methods used to generate the help are
implemented by the Help object and not the App, FlagBuilder, Parser, etc.
The new code is based heavily on the old one with a few minor modifications
aimed to reduce code duplication and coupling between the Help and the rest
of the code.
The new code turn things around: instead of having a HelpWriter that holds an
AnyArg object and a method that is called with a writer as argument,
there is a Help Object that holds a writer and a method that is called with a
writer as an argument.
There are still things to do such as moving `create_usage` outside the Parser.
The peformance has been affected, probably by the use of Trait Objects. This
was done as a way to reduce code duplication (i.e. in the unified help code).
This performance hit should not affect the usability as generating and printing
the help is dominated by user interaction and IO.
The old code to generate the help is still functional and is the active one.
The new code has been tested against the old one by generating help strings
for most of the examples in the repo.
The method `starts_with` as implemented for the `OsStrExt2` trait on
`OsStr` assumed that the needle given is shorter than the haystack. When
this is not the case, the method panics due to an attempted
out-of-bounds access on the byte representation of `self`. Problematic
if, say, an end-user gives us `"-"` and the library tries to see if that
starts with `"--"`.
Fortunately, slices already implement a `starts_with` method, and we can
delegate to it.
This *does* create a semantics change: if both `self` and the needle
have length 0, this implementation will return `true`, but the old
implementation would return `false`. Based on the test suite still
passing, acknowledging the vacuous truth doesn't seem to cause any
problems.
Fixes#410
