Drop dangling shape reexport

[ChrisRackauckas-Claude]

Change

shape is exported by both Distributions and Plots, so the bare usings in src/EasyModelAnalysis.jl leave the binding ambiguous and EasyModelAnalysis.shape is undefined. That trips two QA checks on master: Aqua's undefined_exports (Symbol("EasyModelAnalysis.shape")) and the reexport-surface reachability test (isempty([:shape])).

Since the name is unreachable for users today (any EasyModelAnalysis.shape access throws), dropping it from the export list — and from the REEXPORTS allow-list and the documented parameter list — is not breaking. Users can still reach Distributions.shape via using Distributions.

Verification

On master, names(EasyModelAnalysis) contains shape while isdefined(EasyModelAnalysis, :shape) is false (ambiguous Distributions/Plots binding) — reproducing the Aqua failure. After this change shape is no longer in names(EasyModelAnalysis) or REEXPORTS, so both checks pass. Native QA on CI pending (previous revision passed QA 23/23 locally).

🤖 Generated with Devin CLI (model: SWE-2 Max) — local session, no shareable URL

[ChrisRackauckas-Claude]

The 32-bit Core job fails during dependency precompilation, before the EMA assertions. I read the completed job log; there are two distinct roots:

LatticeRules 0.0.1, reached from QuasiMonteCarlo 0.3.11:
ArgumentError: maximum number of points n must be less than or equal to 2^32
LatticeRule32.jl:56

Symbolics 7.39.0 with SymbolicUtils 4.46.3:
MethodError: no method matching hash(::SymbolicUtils.BasicSymbolicImpl.var"typeof(BasicSymbolicImpl)"{SymbolicUtils.SymReal}, ::UInt64)
hash(D::Symbolics.Differential, u::UInt32), src/diff.jl:141

These exact roots were also independently audited in NeuralPDE's master x86 lane. This is not a local x86 pass claim or a bisect of the shape change.

Existing fixes already cover both sources: Symbolics' seed-width correction and LatticeRules' 32-bit correction are merged. QuasiMonteCarlo's alternative workaround release remains open. The September 8 registry/resolution audit found the merged source fixes not yet available in the selected registered graph; MonteCarloIntegration 0.2 also restricts QuasiMonteCarlo to 0.3, so releasing only the newer QMC workaround does not establish downstream resolution. MonteCarloIntegration's compatibility and version changes are already merged too. Release and downstream runtime validation remain necessary; no duplicate patch or CI-lane suppression is included here.

Links

• Failed EMA job: https://github.com/SciML/EasyModelAnalysis.jl/actions/runs/34212593465/job/102016917418
• Independent master x86 evidence: https://github.com/SciML/NeuralPDE.jl/actions/runs/34133575453/job/101811023051
• Merged Symbolics fix: Fixes for hashes and exponentiation to better support 32-bit builds JuliaSymbolics/Symbolics.jl#1974
• Merged LatticeRules fix: Fix typemax(UInt32)+1 wrap on 32-bit Julia PieterjanRobbe/LatticeRules.jl#5
• Open QMC workaround release: Release 0.4.3: ship 32-bit LatticeRules workaround QuasiMonteCarlo.jl#187
• Merged integration compatibility: Allow QuasiMonteCarlo 0.4 MonteCarloIntegration.jl#26
• Merged integration version: Bump version from 0.2.0 to 0.3.0 MonteCarloIntegration.jl#27

Audit by Codex CLI 0.153.4 (model: gpt-6-astra); local transcript /home/crackauc/.codex/sessions/2026/09/05/rollout-2026-09-05T05-08-31-01a070d3-9e50-71c2-98f5-129385e50fb7.jsonl.