Allow CSV 1.0 and add stdlib compat entries in docs

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

What changed and why

This is the non-NeuralPDE half of the Dependabot group update in #371. It takes the CSV = "0.10" → "0.10, 1.0" widening and the six stdlib compat entries (LibGit2, LinearAlgebra, Pkg, Printf, Random, Statistics), and leaves out that PR's NeuralPDE = "5.15, 6" → "5.15, 6, 7.0".

The diff here is byte-identical to #371 minus the NeuralPDE line.

Why NeuralPDE 7.0 is held back

NeuralPDE 7.0 is a breaking major release, and #371's CI never actually exercised it. docs/src/showcase/pinngpu.md is the only page in docs/src/ that executes NeuralPDE code — the other three mentions are prose and index links. In the #371 run, that page failed at line 146:

ps = ps |> ComponentArray |> gpud .|> Float64
   exception =
    Failed to compile PTX code (ptxas exited with code 255)
    ptxas fatal   : Value 'sm_70' is not defined for option 'gpu-name'

The first NeuralPDE call is at line 156 (NeuralPDE.GridTraining), ten lines later, so it was never reached. That sm_70 failure is a runner issue, not a dependency one: #371 drew self-hosted-gpu-v100-... (sm_70) while the last main build drew self-hosted-gpu-t4-... (sm_75), with the same CUDA.jl v6.2.2 on both.

The names pinngpu.md uses (GridTraining, PhysicsInformedNN, discretize) do still exist and are exported in NeuralPDE 7.0.0, so the widening may well be fine — but PhysicsInformedNN's init_params kwarg and the discretize signature are the kind of thing a major bump changes, and there is currently no run that says either way. It is worth revisiting once the V100/sm_70 runner problem is resolved and pinngpu.md can build again.

For context, the same NeuralPDE 7.0 release breaks NeuralLyapunov.jl (SciML/NeuralLyapunov.jl#234) — it removed logscalar, which that package extends.

Verification

docs/Project.toml resolves cleanly on Julia 1.12.7 with this change:

$ julia +1 --project=docs -e 'using Pkg; Pkg.instantiate()'
DOCS_RESOLVE_OK
  VER Statistics 1.11.5
  VER CSV 0.10.17
  VER NeuralPDE 6.4.1
  VER CUDA 6.2.2

typos docs/Project.toml is clean. Runic is not applicable — no .jl files are touched.

Note that CSV still resolves to 0.10.17 in this environment even though 1.0.0 is registered and its julia = "1.10.0 - 1" bound is satisfied. Something else in the docs environment is still holding it at 0.10; I did not chase down what. The 1.0 allowance is therefore latent here — it is not wrong, but it does not change what gets installed today.

What I did not verify

• The docs build itself. I did not run docs/make.jl locally; it is a multi-hour build and several showcase pages require a GPU that this machine does not have.
• Whether Build Documentation will go green. It will almost certainly still fail on :linkcheck — main's scheduled build already fails there with 403s on docs.sciml.ai/DiffEqDocs/..., independent of this PR. It may also still fail on the GPU example blocks if it draws a V100 runner.

Should a reviewer push back

Reasonable alternative: merge #371 as-is and take NeuralPDE 7.0 on the theory that the API names still exist. I opted against that because the evidence for it is absent rather than positive.

🤖 Generated with Claude Code (model: claude-opus-5)

https://claude.ai/code/session_01E571asmVWqyBRDHiR7PGFE