This fixes an issue where mdbook would panic if a non-draft chapter has
a None source_path when generating the search index. The code was
assuming that only draft chapters would have that behavior. However, API
users can inject synthetic chapters that have no path on disk.
This updates it to fall back to the path, or skip if neither is set.
This config setting provides the ability to disable search indexing on a
per-chapter (or sub-path) basis.
This is structured to possibly add additional settings, such as perhaps
a score multiplier or other settings.
Uses an iframe instead. The downside of iframes comes from them
not necessarily being same-origin as the main page (particularly
with `file:///` URLs), which can cause themes to fall out of sync,
but that's not a problem here since themes don't work without JS
anyway.
Before this change, the Rust `unstable-book` is 88MiB.
With this change, it becomes 15MiB. Other pages might not be
as extreme, but it's expected to help any book like this.
This change is so drastic because, if every chapter has a link to
every other chapter, the result is *O*(n<sup>2</sup>) text output.
on rust reference book this reduces total allocs from 490mb to 474mb:
==23272== Total: 490,538,699 bytes in 1,760,117 blocks
==23272== At t-gmax: 13,872,954 bytes in 4,655 blocks
==23272== At t-end: 488,516 bytes in 884 blocks
==23272== Reads: 830,509,060 bytes
==23272== Writes: 522,290,614 bytes
to
==40876== Total: 474,156,323 bytes in 1,521,025 blocks
==40876== At t-gmax: 13,872,954 bytes in 4,655 blocks
==40876== At t-end: 488,516 bytes in 884 blocks
==40876== Reads: 820,933,434 bytes
==40876== Writes: 514,838,350 bytes
Before, a code block would always end with a final newline. The
newline was added unconditionally by `hide_lines`.
When the code block is syntax highlighted by highlight.js, this is not
a problem, no empty line is added for a final trailing `\n` character.
However, when the code block is editable and thus handled by the ACE
editor, a trailing newline _is_ significant. I believe this issue is
most closely described by https://github.com/ajaxorg/ace/issues/2083
in the upstream repository.
The effect of the way ACE handles newlines is that a code block like
<pre>
Some code
</pre>
will create an editor with _two_ lines, not just one.
By trimming trailing whitespace, we ensure that we don’t accidentally
create more lines in the ACE editor than necessary.
Surprisingly, this fixes the error filed at #1860!
This seems suspicious, perhaps indicative of a bug in Rust's non-lexical
lifetime handling?
The lifetimes in the `handlebars::Renderable::render` method signature
are quite complicated, and its unclear to me whether or not Rust is
catching some new safety edge-case that wasn't previously handled
correctly...
Possibly related to `drop` order, which I *think* is related to the
order of binding statements?
