Building a paragraph through insert_text/insert_object at its end walked
every run (and insert_object counted every object before the offset) on
each call, so n appends cost O(n^2): one paragraph of 64k runs took
8.5 minutes. push_text and push_object extend the last run or start a
new one, giving the same paragraph as inserting at len().
* ci: run the Windows gate's tests with nextest and link with rust-lld
The Windows leg of the Workspace gate takes 55-100 minutes. About half of that is
cargo test running ~370 test binaries one after another, twice (default and all
features), and another 15-17 minutes per feature set is the uncached test build,
where link.exe links every test executable.
- Windows runs the tests with cargo-nextest (parallel across binaries) and the
doctests with cargo test --doc. The same tests run; .config/nextest.toml puts
the tests that relied on in-process mutexes (craft-ui GPU, craft-video h264
fixtures, craft-print) in one-at-a-time groups.
- Windows links with rust-lld via CARGO_TARGET_X86_64_PC_WINDOWS_MSVC_LINKER.
Linux and macOS legs are unchanged.
* ci: keep link.exe on the Windows gate
The first run with rust-lld built the tests no faster than link.exe at the same
runner speed (14.4 min against 15.1 min on a runner 8% slower at clippy), so the
linker change is not worth its risk. nextest stays.
* svg: floor declared() name window to a character boundary
If the MAX_NAME+1 cut split a UTF-8 character, get(..bound).unwrap_or(rest)
fell back to scanning the entire suffix. Floor the window instead, and
when recording an overlong name extend to the next character boundary so
the key is still longer than MAX_NAME.
* test(svg): assert declared() keys stay bounded after a multibyte cut
The overlong-name refusal alone also passed on the old unwrap_or(rest)
fallback, which recorded the whole suffix. Bound recorded keys to
MAX_NAME+4 so the test fails if the suffix is stored again.
camera_raw_mode read a missing SubIFDs offset as 0, so it walked the TIFF
header as the pointer array and could classify a malformed TIFF as RAW.
Skip the entry instead, matching LightCraft's classifier.
* wordmodel: keep the kashida length of a justified paragraph
ST_Jc lowKashida, mediumKashida and highKashida were all read as
Align::Justify, so the length was lost. ParaProps and ResolvedPara now
keep it; a direct both/justify cancels an inherited length. The DOCX
reader and writer round-trip the three values.
* docx: keep a style's kashida length when writing justify
w:jc encodes alignment and kashida together. A paragraph or style
with Align::Justify and kashida unset still inherits a length from
the style chain; writing "both" would read back as Some(Kashida::None)
and cancel it. Use the resolved length in that case.
A defined name's formula may use a structured reference (PtgList). Names were decoded inside
xlsb::workbook, before Package::workbook indexed the table parts, so such a name was left out
and the cells using it kept only their saved results. xlsb::workbook now returns the names
undecoded (Names), and Package::workbook decodes them with xlsb::read_names after
index_tables. Regression: a_name_using_a_structured_reference_is_decoded_after_tables_are_indexed
fails without the reorder.
from_bytes still requires the slice to be fully consumed. parse_prefix
returns the descriptor and the number of bytes used, matching
VersionedDescriptor, so apps can read a descriptor that is followed by
more data (layer effects, fill and adjustment blocks).
Part of #778 and effectcraft#786.
file() only checked length when reopening after eviction, so a prefix
read could still succeed if the handle was pooled. Check metadata on
the cached handle too, and drop the entry if the file is shorter or no
longer a regular file. MAX_OPEN_FILES is documented as a bound on
retained pool entries, not in-flight Arc handles.