Fixes#10161.
Includes transition of existing options from registry to options.json,
and once options.json exists, will always read from that and ignore
registry. Was able to eliminate some of the legacy datatype usage for
several options. Reader/writer includes naive file lock retry to avoid
contention, which I am hopeful will be sufficient for our needs, given
the low chance of contention.
First part of addressing #10161 is to refactor the existing usage of the
registry values into the common KeymanDeveloperOptions unit. This was
mostly straightforward except for the changes to KeymanSentryClient,
which happens to be the most important part of this particular fix.
Relates to #9473.
Now that each compiler has a common interface, we can consolidate the
common code into BuildActivity. This makes it very easy to see where the
compiler calls have special cases, such as the model_info and
keyboard_info compilers (these are primarily for internal use, so the
special case work should not be a problem).
`outFile` is not a compiler option, and so it should not be a part of
the `CompilerBaseOptions` interface. This formally separates the
command-line options from the compiler options, which eliminates a
number of restrictive structures.
Fixes#10200.
This is a two-part fix. First, with our normal `set -e` build
environment, the existing loop never ran more than once, so the 5
second delay pattern we used to use in the dim dark past was no longer
being applied. It is likely that in most cases, this will address the
failure to detach.
The second part is to escalate our detach with a `-force` parameter,
after 5 failed polite attempts. This is new and is addressing the
suggestion in #10200, so we'll have to see if it has any impact on
build stability.
Relates to #9473.
Refactors the public API of ModelInfoCompiler to meet KeymanCompiler,
including moving file write responsibilities into the class itself.
`sources` has become a property of `options`, which is perhaps a slight
violation of the intent of the `init()` function being a one-time call,
as it means we'll need to instantiate an new class for each file we
compile. Given the use-case for the model_info compiler is almost
exclusively internal, I think this is acceptable. Tell me if you think
otherwise!
Relates to #9473.
Refactors the public API of KeyboardInfoCompiler to meet KeymanCompiler,
including moving file write responsibilities into the class itself.
`sources` has become a property of `options`, which is perhaps a slight
violation of the intent of the `init()` function being a one-time call,
as it means we'll need to instantiate an new class for each file we
compile. Given the use-case for the keyboard_info compiler is almost
exclusively internal, I think this is acceptable. Tell me if you think
otherwise!
Relates to #9473.
Refactors the public API of PackageCompiler and
WindowsPackageInstallerCompiler to meet KeymanCompiler, including
asyncing a bunch of functions, and moving file write responsibilities
into the classes themselves.
The classes were updated together in a single commit because
WindowsPackageInstallerCompiler depends on PackageCompiler, and needed
refactoring for the updated API access there anyway.
Most of the test cases needed only minor patching, but introducing the
`async init()` function has caused them to all be async. The
test-messages module has a slight functional change with the compile
process running completely rather than stopping on first message, which
meant we needed to add extra dummy fixtures to avoid other errors that
arose later in the compile process. The keyman.exe fixture is a text
file, not a Windows executable.
