`document.json` was read under the 512 MB cap for any archive member, so a
large imported document saved but didn't open again: 900 000 layers save
as a 1.48 GB document.json, packed into 244 MB. It may now unpack to 16
times its packed size (counted up to the archive's own size), within
512 MB and 4 GB; images and other members keep 512 MB. A small archive
still can't unpack into gigabytes.
That saved file reopens in 13 s with a 5.9 GB peak.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YQxjkNSUJhTkTfY74aNeki
A real 705 MB .fig file (401 644 nodes, 894 004 layers once instances are
expanded) didn't open: "message: too many values". Lifting that cap alone
made it import only its first 200 000 layers, and lifting that one took
41 GB of memory.
The global caps stay as they are. The .fig import gets budgets from the
size of its compressed canvas instead (counted up to the file's own size),
not from the message it unpacks to, since chunks inflate up to 512 MB:
- the kiwi value budget is one value per 2 bytes of the canvas, at least
the old 20 million (kiwi is only read for .fig files);
- the layer cap is one layer per 32 bytes of the canvas, at least the REST
import's 200 000 and at most 2 000 000. A small or crafted file can't
expand its instances into millions of layers, however far it unpacks; the
REST JSON import keeps 200 000.
Memory:
- The mapper no longer passes the document through `json!`, which serialises
its arguments: the whole tree was deep-copied twice.
- Vector networks are built once per distinct network and handed to the REST
importer beside the JSON (`vectorNetworkRef`), not written into it as JSON
objects for every instance that shows them.
- Decoded kiwi objects are trimmed, and each import stage's input is dropped
as soon as the next stage is built.
On that file: peak footprint 41 GB -> 11.3 GB, 67 s -> 25 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YQxjkNSUJhTkTfY74aNeki
A home-screen drop that carries only bytes queued file.open with data and name, which file.open rejected because it requires a path. file.open now also takes base64 data with a name and opens it as a new document without a save path.
Found while testing crates/engine/src/cmd/file.rs:1283
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Regression test for #586: with the colour picker open on a red rectangle or
pen path, duplicating it (edit.duplicate, or copy and paste) and recolouring
the copy leaves the original red. The cause was the picker staying bound to
the original after the selection moved to the copy, fixed by #523; without
that fix this test fails. The duplicated paint is already independent data.
- Files dragged over the window show a "Drop to open" (home screen) or
"Drop to import" (canvas) overlay.
- On the home screen each dropped file the app opens becomes a document of its
own; any other file is turned away with a toast.
- On the canvas dropped files import as before, at the drop point.
- A small card ("Opening as a new file", "Adding to {document}") is on screen
before a large file blocks the window; the files then open a frame later.
- A canvas import stays with the document and page that received the drop,
keeps its placement, restores the page and selection chosen meanwhile, and is
cancelled with a message if that document or page is gone.
- ui.inspect reports pendingOpens; every new string, including the cancel
messages, goes through tr() with en, de and es entries.
page.duplicate copies every layer with duplicate_all_into (navigation between
copied frames points at the copies) and builds the old-to-new list for flows
from that copy, instead of copying each layer separately.
Main's case-insensitive replace loop (from an earlier PR) took the place of the
loop this change edited, so the replacement text is normalised once, where
the parameter is read.
Czech counts take three integer forms: one (1), few (2-4) and other
(0 and 5+). Since #257 catalogs can supply `few`, so `plural()` gets a
Czech rule and the 38 plural entries in cs.toml now read naturally
("1 vrstva", "3 vrstvy", "5 vrstev") instead of the count-neutral
"Vrstvy: {n}" that #469 used for everything but 1.
Test czech_integer_plurals covers the rule (1; 2-4; 0, 5, 11, 12, 21,
22, 100, 1 000 000), a real key through counted(), and that every Czech
plural entry has one and few. Docs no longer say Czech has only one/other.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The control server answered a line that wasn't JSON with an error and read
on. A web page can make the browser send an HTTP request to the loopback
port; its request and header lines were skipped that way and the JSON lines
of its body ran as commands. The first line that isn't JSON now closes the
connection before anything after it is read. Clients that send JSON lines
(MCP --attach, scripts) are unaffected.
