WASM: Host a WASM testbed site under the QuantEcon org

Proposal: host a WASM testbed site under the QuantEcon org

Motivation

Work on #925 has surfaced two bugs in the emscripten-forge Numba build (the jitted-generator hang, #927; and a crash when saving to the persistent cache), with diagnoses, reproducers, and in one case a monkeypatch-verified fix. We should have our own place to try candidate fixes in a real browser kernel and to check that the affected quantecon functions work, both for these two issues and for whatever comes next in the campaign.

Upstream CI (fork + PR) can build a patched package and run the recipe's own tests in wasm, but only as part of a PR, and it cannot run the end-to-end check with quantecon on the xeus-python/JupyterLite stack that #925 targets. A C-level fix such as the generator alias cannot be tested by monkeypatching a live kernel at all; it needs a rebuilt package running in the browser.

Proposal

A small repo (e.g. QuantEcon/wasm-testbed) publishing a JupyterLite site to GitHub Pages whose environment can include locally patched packages. The result is our own counterpart of https://notebook.link/@anutosh491/numba-ecosystem — a URL that opens an xeus-python notebook in the browser with the smoke notebooks preloaded, nothing to install — except that we choose what goes into the environment, and every deployment corresponds to a commit.

1. A fork of emscripten-forge/recipes with a branch per candidate fix to recipes_emscripten/numba; build numba for emscripten-wasm32 either locally (pixi run build-emscripten-wasm32-pkg) or via a draft PR on the fork (CI uploads the built .conda as an artifact).
2. A GitHub Actions workflow that indexes the patched .conda into a static conda channel with rattler-index, builds the site with jupyterlite-xeus from an environment.yml listing that channel first, then emscripten-forge and conda-forge, and deploys to Pages. Swapping the package under test is a one-line change.
3. Smoke notebooks shipped with the site: the environment-check cell, the per-issue MWEs (currently the five-line generator MWE and the persistent-cache MWE), and the quantecon checks (support_enumeration / vertex_enumeration on small games with known equilibria; simplex_grid under a warm cache), extended as new issues appear.
4. Results go into #928's table. If we open upstream PRs, the browser evidence comes from here.

Implementation notes

• Make the patched package win the solve. Listing the custom channel first is not enough: at equal version the solver prefers the higher build number, and the stock package is build 1. Bump build: number in the patched recipe (or pin the build in environment.yml); otherwise the site silently uses stock numba.
• The channel is only needed at build time. The browser never contacts a channel; jupyter lite build bakes the packages into the site. So a local directory in the CI workspace suffices (channels: - ./channel, with emscripten-wasm32/ and noarch/ subdirs indexed by rattler-index fs channel/). If the xeus build's solver turns out not to accept local paths, the fallback is a quantecon channel on prefix.dev (rattler-build upload prefix), the mechanism emscripten-forge itself uses.
• Commit the .conda into the repo (channel/emscripten-wasm32/, a few MB) rather than fetching CI artifacts, which expire and need tokens across repos.
• Prove the patched build is in use. The environment-check cell should print the numba build string, and the ctypes symbol check from #927 (numba_make_generator resolves only under the patch) is the definitive marker.

Notes

• Patched packages must be built from the recipe so they carry the channel's pins (llvmlite ==0.49.0, numpy 2.4.*); mixing them with stock emscripten-forge packages is then safe.
• Once a fix lands upstream, the corresponding package is dropped from the channel; with an empty custom channel the site is just the #928 deployment.
• Nothing runs server-side: the site is static files on Pages.
• The site's notebooks/?path=smoke_test.ipynb URL deep-links straight into the smoke notebook; that is the link to cite next to results in #928.

Refs: #925, #927, #928

🤖 Generated with Claude Code (Claude Fable 5.1)