Fix Burgers docs tutorials: Hugging Face dataset mirror + Reactant-safe GridEmbedding

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

What changed and why

The docs build on main is red (https://github.com/SciML/NeuralOperators.jl/actions/runs/35161521723). Two independent causes:

1. Burgers dataset download (docs tutorials). Both Burgers tutorials fetched burgers_data_R10.mat from Google Drive via gdown. Drive rate-limits shared files, so the build died with Python: FileURLRetrievalError: ... Too many users have viewed or downloaded this file recently, and every later @example block on those pages cascaded into UndefVarError: x_data / x_data_dev / y_data_dev. The tutorials now fetch the same file from the kks32/sciml-dataset Hugging Face mirror over plain HTTPS with the pinned sha256 d1a04567..., pinned to an immutable mirror revision (982685ff) rather than main. This also removes the PythonCall/CondaPkg/gdown stack and docs/CondaPkg.toml from the docs environment. The DataDep is renamed Burgers -> BurgersR10 so machines with a stale Drive-zip cache refetch instead of reusing the old unzipped layout.
2. GridEmbedding vs Reactant (src/layers.jl). fix: transfer GridEmbedding grid to input device (fixes CUDA scalar indexing #125) #132 placed the CPU-built grid via Lux.get_device(x)(grid), but get_device throws on TracedRArray inside Reactant.@compile, breaking the FNO tutorial training step. Allocate like x and broadcast-assign instead: no runtime device query, grid still lands on x's device, preserving the FNO invokes scalar indexing #125 CUDA fix.

Verification (local, Julia 1.12.4)

Full docs build, julia --project=docs docs/make.jl from the repo root (after julia --project=docs -e 'using Pkg; Pkg.develop(PackageSpec(path=pwd())); Pkg.instantiate()'):

[ Info: SetupBuildDirectory: setting up build directory.
[ Info: Doctest: running doctests.
[ Info: ExpandTemplates: expanding markdown templates.
[ Info: CrossReferences: building cross-references.
[ Info: CheckDocument: running document checks.
[ Info: Populate: populating indices.
[ Info: RenderDocument: rendering document.
[ Info: HTMLWriter: rendering HTML pages.
...
[ Info: Automatic `version="0.7.3"` for inventory from ../Project.toml
┌ Warning: Documenter could not auto-detect the building environment. Skipping deployment.

Zero failed @example blocks and no makedocs encountered an error [:example_block]; both tutorial pages rendered (docs/build/tutorials/burgers_deeponet-*.png 258827 B, burgers_fno-*.png 247647 B). The output is byte-identical to the pre-pinning build, so the pinned-revision commit changes no behavior.

Pinned mirror fetched through DataDeps' own fetch+checksum pipeline into a fresh DataDep (BurgersR10Pinned, no cache): PATH=.../datadeps/BurgersR10Pinned/burgers_data_R10.mat, SIZE=644427710, sha256 d1a0456776255a4bd841dbc18951d3f468266d945d96e24ae531a12f18bb5a1a (matches the pin).

Fail-before evidence on the same machine/paths: full build failed at burgers_deeponet.md:5-50 with FileURLRetrievalError, cascading to UndefVarError: x_data / x_data_dev on the following blocks; and before the GridEmbedding commit the FNO training step failed inside Reactant.@compile with get_device "isn't meant to be called inside Reactant.@compile context".

Not verified

• GPU/CUDA path for the new similar+broadcast grid placement (no GPU locally). The repo GPU CI lane passed on this branch and should confirm FNO invokes scalar indexing #125 stays fixed.
• tests / Core (julia 1, ubuntu-latest, x86) is red on this PR. It fails in a dependency __init__ (TypeError: in keyword argument priority, expected Int64, got Int32) and this PR does not touch the test environment; I did not trace it to a main run from here.
• The full local @example run reused the on-disk BurgersR10 dataset cache; the fresh-fetch path was verified separately (above) and CI's Documentation run on this branch (https://github.com/SciML/NeuralOperators.jl/actions/runs/35211262876) performed the fresh download.

🤖 Generated with OpenCode (model: muse-spark-1.3-contributor-free); no public session URL (local session at /home/crackauc/sandbox/tmp_20260917_040834_91074).

CI triage (2026-09-25)

Head commit a534b3c; base main @ b42958f. The branch is up to date with the base (git merge-base HEAD upstream/main == base HEAD), so no merge commit was needed; GitHub reports MERGEABLE. gh pr checks 174 shows exactly one failing check on the head.

check	classification	evidence
tests / Core (julia 1, ubuntu-latest, x86)	master-red — identical failure on base	PR head run: https://github.com/SciML/NeuralOperators.jl/actions/runs/35291773241/job/105437248640 — Base (b42958f) run: https://github.com/SciML/NeuralOperators.jl/actions/runs/35161522453/job/105022031954 — Both fail in Reactant.Accelerators.CPU.__init__ (CPU.jl:50) with TypeError: in keyword argument priority, expected Int64, got a value of type Int32 during package __init__/precompile, before any NeuralOperators test code runs. 32-bit-Julia dependency issue; this PR touches only docs/** (Burgers tutorials) and src/layers.jl (GridEmbedding) and does not touch the test environment.
all other checks (Documentation, Downgrade, Runic, Typos, benchmark, remaining Core/QA/GPU lanes)	pass on head	gh pr checks 174 — all green

Notes:

• Documentation was red on base (https://github.com/SciML/NeuralOperators.jl/actions/runs/35161521723) and is green on this head (https://github.com/SciML/NeuralOperators.jl/actions/runs/35291773216/job/105435948470) — the docs fix is confirmed by CI.
• Local verification of the head tree: typos on changed files clean; Runic --check src/layers.jl clean (matches the green CI lanes). The Core suite was not re-run locally: the only red lane is 32-bit-only and fails identically on base, and all x86_64 Core lanes already pass in CI on this exact head commit.
• No code changes made by this triage (nothing PR-caused is red), so no new commits were pushed.

---

🤖 Posted by an AI agent — harness: opencode 1.18.31 · model: opencode/muse-spark-1.3-contributor-free
Conversation: /home/crackauc/sandbox/fleet-master-jobs/nw-neuraloperators-174/log.txt

Risk assessment

• Risk: medium
• Blast radius: GridEmbedding and FNO models using positional grids, particularly CUDA/custom arrays; Burgers tutorial downloads. No new dependency internals, weakened tests, or apparent SemVer-breaking API changes.
• Evidence: Head a534b3c has 19 passing checks and one failure: tests / Core (julia 1, ubuntu-latest, x86) / Tests - Core (Julia 1). Head and main b42958f, plus recent main commit fed7f48, fail identically in Reactant.Accelerators.CPU.__init__: priority expects Int64, receives Int32. Pre-existing; no observed CI regression. Documentation, GPU, QA, remaining Core, formatting, spelling, downgrade, and benchmarks pass. No review comments exist.
• Independent review: Codex CLI 0.156.1 (Mac) / gpt-6-astra rated it medium, high confidence: the CI failure is pre-existing, but the description overstates GPU validation because GPU FNO tests omit positional grids and existing grid checks use CPU arrays.
• Merge: needs human review (no focused regression test was added for the allocation/broadcast change; passing GPU CI does not establish CUDA GridEmbedding compatibility, an undisclosed coverage gap).

---

🤖 Risk assessment posted by an AI agent (fleet master) — harness: Codex CLI 0.156.1 (Mac) / gpt-6-astra; dispatched by Claude Code head, model claude-opus-5-5[1m]
Conversation: local Claude Code session 3cd6500a-1f81-46b5-ac0b-c466e15b6a53 on Chris's Mac (session ID, no URL)