save deflated every asset that isn't PNG, JPEG, GIF or WebP, with no size
cap, while opening refuses a deflated asset over 1 GiB or deflated entries
over 4 GiB together. An asset past either cap is now stored as is, which
opening reads at any size. Smaller assets are deflated as before, and the
zip bomb guard on load is unchanged.
* Automation roots: refuse COM¹ to LPT³ and a device name before a colon
is_device_name missed COM¹, COM², COM³ and LPT¹, LPT², LPT³, which Windows reserves as it does COM1.
It also cut the name only at a period, so COM1: and NUL:x were not seen as device names.
* Credit my name in the About window
* Automation roots: document the device-name tests
CodeRabbit's docstring check counts the touched tests. same_component stays without a doc comment, because #2205 deletes that function and a doc line above it conflicts.
* Automation roots: record the device-name refusal
The MCP roots description and the roadmap now name COM¹ to LPT³ and a device name before a colon.
* fix(engine): index lookups that were quadratic in a document's styles, tags, swatches and runs
* address review: first swatch per name in book sync; bound spelling input before collecting
* address review: keep the last source style per name in book sync; bound GREP replacement bytes while generating them
* address review: write GREP replacements into a byte-bounded buffer; gate keep_last test off wasm
* merge main into fix/sec-engine-linear-lookups
* address review: make the suggestion length-cap test deterministic
* fix(export): leave out text hidden by inherited conditions or deletion in every exporter
* address review: skip hidden frames in page text and hidden article items; first visible paragraph as nav label; read table anchor's own format
* address review: skip nonprinting frames and non-printable layers in fixed-layout page text
* Apply review fixes
- crates/epub/src/html.rs
- crates/engine/src/cmd/export/formats.rs
* Apply review fixes
- crates/epub/src/html.rs
- crates/engine/src/cmd/export/formats.rs
* address review: fixed-layout page text skips frames in inactive object states
* address review: place a threaded story at its first shown frame; skip it only when every frame is hidden
---------
Co-authored-by: autocraft-bot[bot] <354215646+autocraft-bot[bot]@users.noreply.github.com>
* format: bytes a failed entry read inflated count against the archive budget
* address review: mutate the payload bytes explicitly in the checksum test
* address review: charge a failed read to the budget it belongs to (stored vs deflated)
* idml: story runs and marker offsets found once, object-style overlay by position, embedded image budget
* address review: saturating embedded-bytes counter
* address review: IDML export past the embed limit succeeds and warns for each image written as a link
* address review: an unlinked image is always embedded, only linked ones fall back to their link
* xtask: refuse a corpus archive that expands past the cap before extracting it
* address review: a within-cap answer needs tar to exit successfully
* Apply review fixes
- xtask/src/main.rs
---------
Co-authored-by: autocraft-bot[bot] <354215646+autocraft-bot[bot]@users.noreply.github.com>
A background write that is running, or that finished before the session took its result, is not in the document's entries yet. Opening recovery from that folder reopened it as a second copy of the live document. It is now skipped. The PUBLISH doc comment names the functions that take the lock.
`recovery::save` and the tracked save cancelled every background write of the
document's folder and uid, including one another session in the same process
had queued. Both now cancel only their own session's writes, refuse while
another session's write of that uid is queued or running, and keep other
sessions' background writes of it from queuing until the write now is done.
Discarding a document's recovery entries cancelled every background write of that uid in the folder, so another session's write of a document with the same uid was dropped without a replacement. The cancel now covers only writes the session itself queued.
Entries written or recovered in a recovery folder that is not the run's own were tracked per folder and uid only, so another session in the same process with a document of that uid could replace or remove them; a recovered entry keeps the uid it was written under. Each entry now carries the session that wrote or recovered it, and only that session may replace or remove it.
A write queued while its entry was the session's is recorded again when it lands, so a cleanup that dropped the record in between no longer leaves the snapshot on disk after Save.
- recovery::save cancels the background writes the session queued for a clean document and removes what a finished one left, so an older snapshot is not published after Undo returns the document to its saved state.
- Cancelling a document's writes leaves a write another session queued to follow in place.
- When the folder sync fails after the rename, the snapshot is reported as published with the error, so it is tracked and Save or Close removes it.
A save, revert or close during a background recovery write now stops
the write at its next piece of output (at most 1 MiB) instead of after
the whole package, and a stopped write is not reported as a failure.
Follows #2645. Its test said a misspelled key used to apply Align Left.
Only empty params and a bare string did; the parameter check already
refused a misspelled key, and the value parser refused an unknown
value. Each case now asserts the refusal message it gets.
A recovery write queued while a cancelled write of the same document
and folder is still running now follows it instead of being refused
until the next timer tick. A write in a folder that is not the run's
own is recorded as the run's in the same step as its rename, so a save
right after it can remove it.
A synchronous recovery write now cancels a background write to the same
folder named another way, and a cancel also matches the folder as the
session named it. Background writes whose results have not been taken
count toward the per-document checkpoint limit, cancelled ones excluded.
Save, Revert and Close clean up a cancelled write's folder under the
name the session used for it.
Save, Revert and Close now cancel the document's background writes to
every folder, including one named with `file.recovery.save {dir}` that
is not in its entries yet, and remove a snapshot such a write already
published even when its result has not been taken. A cancel that comes
after the write published retires its result, so the removed snapshot
is not recorded as a checkpoint after a later edit. Cancels apply only
to writes the same session queued.
The test that closes a run's lock and expects it free no longer runs
while a test starts its child process, which held a copy of the lock
until the child ran its program.
A save, revert or close that cancels a background write now waits for a
rename already under way, and a later rename sees the cancel, so a stale
snapshot or its metadata cannot be published after the cleanup. The
writer's results go back to the session that queued them: another
session's poll no longer removes their snapshots, and results for a
session that has been dropped are not kept.
The timer test now holds the recovery writer at the start of its write
through a handshake (`recovery::testing`, behind the engine's `testing`
feature), checks that frames and edits go on while the write is held,
and fails when the write runs on the interface thread. The engine tests
wait for the writer to go idle instead of polling with sleeps.
`file.recovery.save` rejects a `background` that is not true or false.
type.align with no `align` key applied Align Left, so a call that
misspelled the key, or passed a bare string, changed the paragraphs to
left alignment instead of failing. The command now returns an error,
leaves the text unchanged and records no undo step. A named value still
applies, and `align: null` still removes the override.
* Render: a blur size is clamped to 1,000,000 pt before the filter planner squares it
* address review: bound effect extents and sigma before spread, padding and reach arithmetic
release.yml also runs nightly on main with craft-common-actions' dry-run: every release build
except macOS, with nothing signed, pushed or drafted. Breakage that only the release build hits
(riscv64, Flatpak, FreeBSD, packaging) then shows up the next morning instead of on release
day, as it did on 11 Nov 2026.
* fix: index style, font, class and layer name lookups in exporters and importers
* Apply review fixes
- crates/epub/src/css.rs
---------
Co-authored-by: aiongg <67543387+aiongg@users.noreply.github.com>
* fix(engine): whole-file reads take regular files only, up to the format limit
* Apply review fixes
- crates/engine/src/io.rs
* Apply review fixes
- crates/engine/src/io.rs
* Fix failing CI check
- crates/engine/src/io.rs
* address review: refuse to save an encoded document longer than the open limit
* address review: open files for reading without blocking (FIFOs, stream-backed pseudo-files)
---------
Co-authored-by: aiongg <67543387+aiongg@users.noreply.github.com>
* fix(engine): package fonts from their loaded bytes under the read roots
* address review: a font's own file is neither overwritten nor packaged when it changed since loading
* Apply review fixes
- crates/engine/src/cmd/package.rs
* address review: compare packaged font file in fixed-size chunks
---------
Co-authored-by: aiongg <67543387+aiongg@users.noreply.github.com>
* Strokes: cap the points a wavy or hashed stroke places along a path
* address review: zero-length quadratic contributes no length in stroke samples
* Fix failing CI check
- crates/geom/src/stroke_style.rs
* Fix failing CI check
- crates/geom/src/stroke_style.rs
---------
Co-authored-by: autocraft-bot[bot] <354215646+autocraft-bot[bot]@users.noreply.github.com>
* engine: Erase/Smooth look at each distinct drag sample once, thinned past 4096
* Apply review fixes
- crates/engine/src/cmd/anchors.rs
* address review: decide whether a drag touched the path on every distinct sample, then thin
---------
Co-authored-by: autocraft-bot[bot] <354215646+autocraft-bot[bot]@users.noreply.github.com>
* doc: appending to a story extends the last run, no scans
* Apply review fixes
- crates/doc/src/story.rs
---------
Co-authored-by: aiongg <67543387+aiongg@users.noreply.github.com>
* Tables: commands given a table id run with nothing in a table selected
Table commands whose params list `story?` and `table?` (table.insertRow,
table.deleteColumn, table.options, the Convert Rows commands and the
others) were refused with "place the insertion point in a table" when the
call named the table but the selection was not in one. A call that passes
`table` now acts on that table whatever is selected. A table that is not
in the document, or not in the given `story`, is refused, and a `story`
that is not a non-negative integer is an error.
The menus are unchanged: with nothing in a table selected, the table
items stay disabled, and calls that name no table are still refused.
* Tables: a bad story is refused whether or not a table is selected
A call that names a table with a story that is not a non-negative integer (a string, a negative number, a fraction or null) was refused only when nothing in a table was selected. With a table selected, the command ignored the story and acted on the named table. Both paths now read story with one helper and refuse it the same way.
* Tables: an edit to a named table keeps a caret in footnote text
With the caret in footnote text, a table command that names a table
runs, and the edit cleared the caret's footnote address, so the next
typing went into the story. clamp_selection now leaves footnote
addresses alone, as in_cell does.