WASM: JupyterLite proof-of-concept deployment and browser smoke suite

[!NOTE]
Updated 2026-08-21. Phase 0 was overtaken by ad-hoc browser runs: the _LAPACK question (#927) was answered on the public numba-ecosystem demo, the intp question (#929) was settled in the #942 review, and PR #938 (merged as 481947d) has landed the deployment config and smoke suite under ci/wasm/. The checklist below is reworded to match those findings and the upstream Numba change that removed the expected gini_coefficient failure (patch 0010, emscripten-forge/recipes#6293); items with recorded browser evidence are ticked and a results table is maintained in the comments. The original text is preserved in the edit history.

Part of #925 (Phase 0). This issue was meant to gate #927 and calibrate #929 and #930; #927 was resolved by ad-hoc browser runs and #929 by the #942 review instead (see the checklist), so its remaining gating value is the cold/warm import timing for #930 and the first full-suite browser run via #933.

Goal

Prove the stack end-to-end with zero code changes, and convert every "should work" in the audit into a known. Because quantecon is noarch: python on conda-forge, there is a real chance it installs and imports in JupyterLite today with no packaging work.

Setup

Build a minimal JupyterLite deployment with jupyterlite-xeus:

name: qe-lite
channels:
  - https://prefix.dev/emscripten-forge-4x
  - https://prefix.dev/conda-forge
dependencies:
  - xeus-python
  - numba
  - quantecon

This file is now committed as ci/wasm/environment.yml (PR #938). numba is unpinned, so a fresh build resolves the current emscripten-forge-4x build — numba 0.67.0 build 1 as of 2026-08-18, which carries patch 0010 (emscripten-forge/recipes#6293) forcing @njit(parallel=True) to the serial pipeline on Emscripten. Note that the browser results recorded so far did not come from this deployment: they came from the public numba-ecosystem demo (Numba 0.66.0) and the #942 review, so the versions per result are recorded in the table.

Smoke checklist

One representative function per Numba feature class used in the library:

• import quantecon — works (implicitly confirmed by the _numba_linalg_solve reproducer in #927 and the simplex_grid reproducer in #942; no timing recorded)
• import quantecon — timed, cold cache and warm cache (feeds #930; #930's "measure first" gate and PR #943 are currently proceeding without this number)
• qe.tauchen — plain lazy @njit (_fill_tauchen); qe.rouwenhorst is pure NumPy with no Numba at all, kept as a no-JIT control case
• MarkovChain.simulate — lazy @njit kernel writing into a Python-allocated array (NRT allocation inside jitted code is exercised by simplex_grid and nelder_mead below)
• qe.random.probvec — exercises the parallel-ufunc → cpu fallback (emscripten patch 0007)
• qe.random.sample_without_replacement — eager guvectorize, explicit i8 signature
• optimize: nelder_mead, brent_max, newton
• game_theory.lemke_howson
• game_theory.vertex_enumeration — numba.typed.Dict; also a jitted generator (_vertex_enumeration_gen), so expected to hang for the same reason as support_enumeration (#927) — not yet run
• game_theory.support_enumeration — FAILS: first call never returns. The _LAPACK question is answered (LAPACK works, see the next two items); the cause is the jitted generator _support_enumeration_gen — emscripten-forge patch 0002 exports only Numba_make_generator, not the lowercase numba_make_generator the generator wrapper imports (#927, not yet reported upstream)
• minimal @njit generator (for i in range(n): yield i) — FAILS (hangs on first call), recorded 2026-08-19 in #927; isolates the generator mechanism from quantecon's code
• np.linalg.solve inside a scratch @njit — works (correct result on Numba 0.66.0; _numba_linalg_solve also works), recorded in #927 on 2026-08-15 and 2026-08-19
• qe.gini_coefficient — expected to work, serially, on emscripten-forge numba ≥ 0.67.0 build 1 (patch 0010, emscripten-forge/recipes#6293, 2026-08-18); the @njit(parallel=True) failure mode this item originally expected no longer exists. Confirm it runs and time it — the serial O(n²) path is why the O(n log n) rewrite in #926 / PR #937 still matters. No browser run of gini is recorded anywhere yet
• simplex_grid — FAILS on a warm persistent cache: RuntimeError: no compiled object yet for <Library 'comb_jit'> at qe.simplex_grid(3, 4) (#944; upstream emscripten-forge/recipes#6309, open). The 32-bit intp boundary is now a documentation question (#929 is doc-only); left unticked because the boundary check itself was never reached and stays blocked behind #944
• quantecon.util.array.searchsorted — the deprecated objmode() shim

Deliverables

• Results recorded in this issue as a table: function → works / fails / notes — started in the comments on 2026-08-21 with the rows that already had evidence (import, np.linalg.solve, support_enumeration, bare generator, simplex_grid); the remaining items are not yet run
• The deployment config committed somewhere reusable — done: ci/wasm/environment.yml plus the suite ci/wasm/smoke_test.py (16 tests mapping onto the 13 checklist items, all passing natively), the Playwright harness ci/wasm/test_jupyterlite.py and a native gate in .github/workflows/ci_wasm_smoke.yml (PR #938, merged 2026-08-21 as 481947d). The Emscripten runner job is parked in #933

Coverage of the browser harness

What actually runs on Emscripten today is test_jupyterlite.py (driven by the job parked in #933), and it covers 5 of the 13 checklist items: import quantecon (untimed), tauchen, bare np.linalg.solve in @njit, support_enumeration and gini_coefficient. smoke_test.py is not yet shipped into the kernel, so rouwenhorst, MarkovChain.simulate, probvec, sample_without_replacement, the three optimize routines, lemke_howson, vertex_enumeration, simplex_grid, searchsorted and the import timing have no browser test until #933 wires in-kernel execution of smoke_test.py. A first green run of the parked job therefore fills at most 5 rows of the table, not 13.

Several of the committed tests also encode expectations that are already out of date; #933 (the runner job) currently notes only the test_support_enumeration hang, so the rest are being handed to it as follow-ups: test_gini_coefficient (xfail(IS_EMSCRIPTEN, strict=True)) and test_gini_fails_on_emscripten expect a failure that patch 0010 has removed; test_support_enumeration and test_vertex_enumeration have no hang guard; test_simplex_grid cites #929 where it should cite #944.

Related

• #927 — jitted-generator hang (support_enumeration, vertex_enumeration); _LAPACK verified working
• #944 — simplex_grid fails with no compiled object yet for comb_jit on a warm cache; decision is to wait for upstream (emscripten-forge/recipes#6309)
• #926 / PR #937 — gini_coefficient O(n log n) rewrite; the browser confirmation of gini belongs here, not in #926's acceptance criteria
• #930 / PR #943 — import-time eager compilation; #930's gate still calls for the cold/warm import timing from this checklist, and PR #943 is proceeding without it
• #933 — Emscripten runner job; the test fixes listed above are to be recorded there
• PR #938 — merged 481947d; delivered the config, suite and harness

Start with ci/wasm/test_jupyterlite.py, ci/wasm/smoke_test.py, ci/wasm/environment.yml, and .github/workflows/ci_wasm_smoke.yml, then follow the parked runner work in #933. The deployment and native suite are already delivered by merged PR #938; remaining browser coverage, timing, results, and outdated expectations are split across #933, #930, and #944.