lyx2lyx sets all new math packages to off when converting from old formats.
This is correct if any command which would cause an automatic package loading
exists in the document. However, it is wrong if no command exists: This leads
to problems if later a command is added (bug #9069), which is especially
annoying for templates.
The fix consists of two parts:
1) convert_use_package() considers now the used commands like
revert_use_package(), and uses them to decide whether to set the package to
auto or off.
2) convert_undertilde() and revert_undertilde() use a slightly adjusted copy of
convert_use_package() and revert_use_package(), so that the bug is also
fixed for undertilde. We cannot use the latter functions directly, because
of "\usepackage undertilde" vs. "\use_undertilde".
When doing the lyx2lyx round trip of the 2.0.8 user guide from format
413 -> 474 -> 413 you get an invalid document, because the math packages
mhchem and undertilde are incorrectly converted.
It turned out that the old reversion code did only work for particular
package oderings. The new one works for abitrary sorted packages.
When doing the lyx2lyx round trip of the 2.0.8 user guide from format
413 -> 474 -> 413 you do not get a zero diff. The most important problem is
caused by the conversion of the argument insets to the old syntax: This
conversion adds an additional empty line (harmless), and it destroys the
document structure if the first inset in e.g. a subsection is not an argument
inset, but e.g. an index or label inset.
The fix is quite easy: Ensure that the paragraph begin is set to the first
argument inset.
revert_justification() issues a warning if the \justification parameter does
not exist, and LyX itself always writes it unconditionally as well, so add it
also in lyx2lyx when converting old documents.
Currently this does not have any user visible effect, but it decouples the
default value of BufferParams::justification from the conversion of old
documents: Now it is possible to set the default to false in LyX, and old
documents will still be converted correctly.
In the current code each paragraph contains a map<Language,
WordList*>, which means that it contains a full copy of the language
object. Since these objects contain translation tables nowadays, this
is a very bad idea.
This patch simply replaces the Language key by a string.
When loading the Userguide on linux/x86_64, the total memory
consumption decreases from 36.27MB to 31.50MB.
