Found while testing crates/engine/src/cmd/object.rs:299-313 (object.stroke: `st.tint = v as f32`, `st.gap_tint = v as f32`; the no-selection path at :273 clamps, this one doesn't)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The conflict was in links.rs, where main's atomic relink work (b050343a) and
this branch's path gate both changed `update`. Kept main's mechanism (build the
replacement with relinked_asset, then insert it) and this branch's gate (a link
craft-docpath refuses is missing and never read).
relinked_asset now takes the path to read and the link to record separately.
Previously they were one argument, safe only because update passed the raw stored
link; passing the resolved path to the old signature would have rewritten every
relative link to an absolute one on each refresh.
Also passes None for the document in main's embedded-status test, where the asset
has no link so the document is irrelevant.
craft-libs#169 is merged, so the fork pin is no longer needed. Pinned to
storytold/craft-libs 82a24b00 (main after the merge) rather than the older
12281d07, which predates resolve_absolute and the Windows colon refusal.
Scratch exports from running the app and the CLI land in outputs/. It is
not ignored, so `cargo xtask assets` — which scans tracked and untracked
files alike — reports every PNG and PDF in it as an asset with no row in
ASSETS.md, and the check cannot pass on a machine that has ever run the
app with an output directory.
Merges upstream/main into the branch, resolving 12 conflicted files, and reconciles where the two sides meant different things — notably the format version, where both had independently defined version 2 (EPS sources here, pinned paragraph defaults on main), so a document without EPS was being written as version 1 and re-upgraded on load.
Verified locally against main b17efb2d: cargo test --workspace --no-fail-fast 2869 passed / 0 failed, cargo fmt --check clean, cargo clippy --workspace --all-targets -D warnings clean, xtask layers and wasm ok.
Hamza's diagnosis of the forgotten OpenGL failure, with the filter reordered so it covers any backend and does not hold another machine's record against this one, plus a regression test.
No workflow runs on this PR: every one is path-filtered and apps/designcraft/src/gpu/backend.rs matches none. Verified locally on 653f83b8 instead — cargo test --workspace --no-fail-fast: 2847 passed, 0 failed, 5 pre-existing ignores; cargo fmt --check clean.
Builds on the fix in this branch, which spotted that a failed OpenGL was
being forgotten because `failed` is filtered to this platform's
candidates and OpenGL is no candidate off Windows.
Rather than spare OpenGL from the filter, run the filter before the
previous start's backend is added. A stored entry that is no candidate
here came from another machine and still says nothing, so it is dropped
as before; the backend a start was actually running on is added
afterwards and stays, whichever backend it is. That keeps the same crash
out of the next instance without naming OpenGL or the platform, and
without holding a failure recorded on someone else's machine against
this one.
Adds the regression test: on macOS, where Metal is the only candidate, a
start left on OpenGL has to keep it out of the instance, while a stored
OpenGL entry that never ran here is dropped.
The merge compiled, but six tests showed where the two branches meant
different things. Fixes, in order of consequence:
- Format version: both branches had independently defined version 2 —
here for retained EPS sources, on main for pinned Keep Lines Together
and drop cap defaults, with an `upgrade` that runs below 2. Writing 1
for documents without EPS then re-ran that upgrade on load and rewrote
their paragraph styles, so saving and reloading changed the document.
EPS now asks for version 3 and everything else stays at 2.
- place.drop into a frame: main turned Auto-Fit off so that resizing a
frame crops rather than refits, and this branch had carried the older
FillProportionally along while moving the geometry out of the edit.
Main's behaviour is the intended one.
- save() compacts unreferenced assets now, so the format test has to
place its EPS for the version sniff to see it at all.
- EPUB entries are named from the asset's file stem (asset_path), not
its bare id: images/logo-<id>.png.
- Command parameters are validated strictly, so the PDF page test sends
each command only the parameters it takes.
- The relink batch's "commits once" check read a process-wide revision
counter, which a test running beside it can advance; the single undo
step is what actually shows it committed once.
The EPS place tests move to cmd/file/eps_place_tests.rs: they write real
files, and main's file-system guard reads src files without seeing inline
#[cfg(test)] modules, but exempts _tests.rs ones.
Four sites already on main that `cargo xtask ci` cannot get past with the
current toolchain; each rewrite is the one clippy suggests and changes no
behaviour. `!x.is_some_and(f)` becomes `x.is_none_or(!f)`, and pdf_bounds
folds `!(a && b) || !c` into `!(a && b && c)` by De Morgan.
Unrelated to the merge they are committed beside; separated so that merge
stays a merge.
Resolve 12 conflicted files. Where both sides changed the same code:
- format: keep main's bounded-read load_zip (MAX_ENTRIES, read_entry) and
this branch's u64 version reporting, so a version above u32::MAX is named
as written rather than clamped. Both sides had fixed the same truncation.
- render: main's natural_size supersedes image_natural_size (it already
reads an EPS bounding box and also handles PDF pages); callers in table
and data merge now use it.
- images: keep the DSC-aware header scan and (atend) trailer lookup, take
main's classic-Mac line endings, and carry its overflow guard into
parse_box. eps_proxy draws main's simple rectangles before falling back
to the placeholder, still over a retained EPS source.
- engine/file: take main's placement shape (add_asset, crop box geometry,
place cursor, finite_rect, object effects) and keep this branch's
validate_placement in both branches; edit() works on a clone and commits
only on Ok, so validating inside it stays all-or-nothing.
- engine/links: keep building the replacement before edit() so a file that
cannot be read leaves the document untouched, and move main's crop_pdf
refresh into relinked_asset, which both relink and update go through.
- engine/export: EPS preview warnings join the font-embedding warnings
rather than shadowing them.
- idml: an EPS element keeps ActualPpi (its pixels are a preview raster);
main's PDF exclusion is unchanged.
Cargo.lock regenerated. Tests and doc strings from both sides kept.
Links, data sources, IDML package links and book chapters go through
io::document_path (craft-docpath) before they reach the file system:
relative paths resolve beside the document, and paths to other computers,
devices and URLs are missing, with the reason in links.list and Preflight.
IDML and INDD packages read a linked file where its link points on this
computer, then from the link's trailing names beside the package, only
when it is a regular file up to 256 MB holding an image or a graphic. A
file URI that names a host keeps it as a network path.
`choose` pruned the failed list to this platform's candidates, which dropped an
OpenGL entry on macOS, where Metal is the only candidate. But OpenGL is not only
a candidate: off Windows it is the in-process fallback that `instance_backends`
adds whenever it has not failed. So a start that recorded OpenGL and never
showed a frame had the entry pruned a line later, and OpenGL went straight back
into the instance — the backend that had just failed was offered again on every
start, with nothing able to retire it.
An entry that can still be reached is kept now. A non-candidate entry still
cannot be chosen as a backend, since `pick` only looks at candidates, so which
backend starts is unchanged; it only stops a failed OpenGL being handed back to
the instance.
This is what `a_start_records_its_backend_until_a_frame_is_presented` asserts,
and that test failed on macOS: `cargo xtask ci` stopped in the `designcraft` bin
before the rest of the workspace ran, so the gate could not go green there at
all.
gaps.md's header asks for a ROADMAP progress-log line and a timestamp when a gap
is closed or shrunk. Gap 13 was shrunk in the first commit (four hyphenating
languages became five); this adds the two it was missing.
Docs only; the fix itself is unchanged.
gaps.md asks for it in its own header: "When you close a gap, delete its entry
(or shrink it), add a line to the ROADMAP progress log and update the
timestamps." The entry was still listed, which is how a second contributor picks
up work that is already done.
Docs only; the fix itself is unchanged.
