You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 35c7081
Browse filesBrowse the repository at this point in the historyBrowse files
feat(wasm): add the fix-candidate output pane to five more recipes
Layer 1 pages showed one of two different output sections depending on
when the recipe was written. `dateutil-1478` and `lark-1585` render
"Baseline output" above "Fix-candidate output"; the other six render a
single "Output". That split was an accident of authoring order, not a
decision, and it left the numpy page looking a generation behind the
dateutil one.
Seeing the wrong result next to the right one is the point of a
reproduction page. One pane only tells a visitor that something is
broken; two tell them what "fixed" looks like. So the two-pane block
becomes the Layer 1 default, and this commit moves five recipes onto it:
cpython-137205, numpy-28287, pandas-56679, php-12167, ruby-21709.
regex-779 lands separately because its second pane runs a real build.
None of those five has a corrected build this page can execute, so the
second pane says exactly that and names the upstream status. It never
shows a hand-written "expected" output — Vivarium reproduces, it does
not assert. Verified upstream state, rather than trusting the recipe
READMEs:
numpy/numpy#28287 open. The only merged response is a
deprecation of the generic unit (PR 29619,
milestone 2.5.0), not a behaviour fix.
python/cpython#137205 open.
pandas-dev/pandas#56679 open.
bugs.ruby-lang.org#21709 open.
php/php-src#12167 fixed in PHP 8.2.12 — but see below.
php-12167 is the interesting one, and its README said something that is
no longer true. The recipe was documented as a "sentinel" that would
flip to `unreproduced` once php-wasm shipped a build past 8.2.12. It
would not. [email protected] does ship PHP 8.4.1, but its default build
has no SimpleXML: `simplexml_load_string()` is undefined, so bumping
the pin does not flip the verdict, it breaks the reproduction — and
reports a false fix signal while doing it. The one prerelease between
0.0.8 and 0.1.0 never finished initialising. Both verified in-browser;
the README now records that instead of the old premise.
`.vh-variant-output--note` is new in the shared stylesheet. `<pre>`
defaults to `white-space: pre`, which is right for JSON but turns a
one-line sentence into a horizontal scroll that clips mid-word —
observed on the live dateutil page before writing this.
The panes also carry `data-fix-status`, so the state is readable
without parsing prose. The follow-up that makes the two-pane block
mandatory can assert on it, in both locales, instead of matching text.
Not in this commit: regex-779's real fixed build, and the validator
change that requires the block on every recipe. The validator has to
land after every recipe conforms or it red-lights main — so the docs
added here say plainly that today's check covers only recipes shipping
a `fix-candidate.json`, rather than promising enforcement that does not
exist yet.
Also folded in: regex-779, php-12167 and ruby-21709 each carried three
`<link rel="preload">` tags for Pyodide's `pyodide.asm.wasm`,
`python_stdlib.zip` and `pyodide-lock.json` — copy-paste from the
Pyodide template. None of the three loads Pyodide (they run a Rust
wasm32-wasip1 binary, php-wasm and Ruby.wasm respectively), so every
visitor to those pages was fetching roughly 10 MB of assets nothing
ever read. The `preconnect` to jsDelivr stays; all three still pull
their own runtime from it.
Drive-by: `.claude/launch.json` pointed at bun 1.3.14 and uv 0.11.14,
neither of which is installed any more, so `preview_start` failed with
ENOENT. Repointed at the current mise installs.
Co-Authored-By: Claude Opus 5 <[email protected]>
0 commit comments