Follow-ups after merging PR #3128 (with the #2990 preview work):
- PDF preview failed on every browser without
Map.prototype.getOrInsertComputed (Chrome < 145, Firefox < 144,
older Safari): "This file cannot be previewed" for any PDF. Load
pdf.js's legacy build, which bundles the polyfills for page and
worker, and bundle the worker through Vite (?worker&url) so
build.target lowers its class static block for Safari 16.0-16.3.
The browser-baseline check only scanned .js output and never saw
the copied .mjs worker; it scans .mjs too.
- PDF grid thumbnails only existed after someone opened the preview.
Page one is now rendered server-side with PDFium (pypdfium2, new
dependency, prebuilt wheels for every shipped platform) on upload,
ZIP extraction and external-folder scans; Generate Thumbnails
backfills PDFs as well. Renders are serialised behind a lock
(PDFium is not thread-safe), run off the event loop, and scale
from the page size so a huge MediaBox cannot allocate a huge
bitmap. Unreadable PDFs land without a thumbnail and keep the
browser fallback.
- A large STEP file takes over a minute to mesh in the browser and
showed only a spinner. STEP loads now show "Converting STEP
model... N s" and a note that large files can take a minute or
more, in all 15 locales.
- Drop the two occt-import-js "externalized for browser
compatibility" build warnings (path/crypto are only required in
its Node branch); any other externalization still shows.
An iPhone on iOS 16 loaded nothing at all -- no error, no partial render,
just white, over LAN IP and over an HTTPS domain alike, while the same
install was fine on Android, macOS, Windows and Linux. remark-gfm, added
in v1.2.5 for the folder README panel, reaches
mdast-util-gfm-autolink-literal, whose module body carries a lookbehind
assertion. Safari did not support lookbehind until 16.4.
A regex literal is validated when its module is compiled, not when the
function holding it runs, so this was never going to fail as a broken
README panel. FolderReadmePanel -> FileManagerPage -> App is a plain
static import chain, the regex landed in the entry chunk, and the browser
refused to compile all 10 MB of it. Nothing executed, so nothing
rendered. v1.2.4 is the last release that loads on those iOS versions.
The panel now renders GFM through a locally composed plugin holding four
of remark-gfm's five sub-extensions -- tables, strikethrough, task lists,
footnotes -- and omitting autolink literals, the only one carrying the
lookbehind. Composing rather than configuring is forced by the bug:
importing remark-gfm at all is what breaks the page, so no runtime option
could have reached it.
Parity was measured rather than assumed. Serialized ASTs against real
remark-gfm over a 34-case corpus, position data included, are identical
in 29; the five that differ are exactly the autolink cases, where the
only change is link -> text with table and list structure intact. Across
26 hostile inputs -- NUL bytes, a BOM, an RTL override, a lone surrogate,
combining marks, a 200 KB line, 500 stacked tables, 60-deep nesting,
malformed and ragged tables -- neither implementation throws and none
diverge, and applying the plugin twice is idempotent for both.
The visible cost is that a bare https://example.com or foo@example.com
typed into a folder README no longer links itself; [text](url) and
<https://example.com> are core markdown and still do. The wiki claimed
"links all render" and now says which.
remark-gfm, mdast-util-gfm and micromark-extension-gfm leave the
dependency tree and their eight surviving sub-extensions are declared
directly, at ranges equal to or tighter than the ^2.0.0 those two
packages declared, so the resolution surface did not widen. The bundle is
23 KB smaller.
Vite's build.target governs syntax lowering and esbuild does not rewrite
regular expressions -- measured, a lookbehind builds silently under
safari15, safari16.0 and es2020 alike, which is how this shipped and then
sat unnoticed for two months. So the guard is a real check rather than a
compiler setting: npm run build now ends in check-browser-baseline.mjs,
which scans the emitted bundles for syntax Safari 16.0 cannot parse and
fails with the offending snippet. It is scoped to parse-time failures
only -- a missing runtime API breaks one feature, while one of these
takes down the whole app and has no graceful degradation to fall back on.
Verified firing on the stale bundle before the rebuild, and running
correctly inside the Docker frontend stage where only frontend/ is
copied.
Seven renderer tests pin both halves of the trade: each surviving GFM
feature still renders, and both forms of autolinking stay off on purpose
so a future dependency bump cannot quietly bring the lookbehind back.
Thomansky