ci: let pak ignore Bioconductor graph on R < 4.4

[krlmlr]

LLM disclosure: this PR was prepared with Claude Code (Claude Opus 4.8).

• By submitting this pull request, I assign the copyright of my contribution to The igraph development team.

Problem

graph is an optional Enhances: dependency published only on Bioconductor. The Bioconductor releases matching R 4.2 / 4.3 (graph 3.16 / 3.18) are archived and return HTTP 404, so pak dependency resolution aborts on the rcc: … (4.2) / (4.3) jobs, in the install step, before anything is built.

Fix

Use pak's documented downstream-package parameter, in DESCRIPTION — no workflow changes at all:

Config/gha/extra-packages: graph=?ignore-before-r=4.4.0

The install composite action already forwards Config/gha/extra-packages to setup-r-dependencies. Per ?pak_package_sources, <package>=?<params> is the special param reference type, whose purpose is "to add a parameter to a downstream dependency", and ignore-before-r "will be ignored on R versions that are older than the specified one". So the parameter attaches to the graph dependency pulled in via deps::..

Verified locally against pak's vendored pkgdepends:

• parse_pkg_refs("graph=?ignore-before-r=4.4.0") → type = "param", package = "graph", ignore-before-r = 4.4.0.
• Solver semantics (pkgplan_i_lp_dependencies): ignored when 4.4.0 > R → dropped on R 4.2 / 4.3, kept on 4.4 / 4.5 / 4.6.
• The action's own grep Config/gha/extra-packages DESCRIPTION | cut -d " " -f 2- extracts exactly graph=?ignore-before-r=4.4.0 (the adjacent Config/comment/gha/extra-packages note does not match the grep).
• DESCRIPTION remains valid DCF.

Why the previous approach was replaced

An earlier revision set PKG_USE_BIOCONDUCTOR=false from a workflow step. That did not work, as the CI showed:

• On R 4.2 / 4.3 the env var was set, but resolution still failed — it only turned the 404 into graph: Can't find package called graph. pak does not drop an unavailable soft dependency unless it is explicitly ignored (ignore, ignore-before-r, or ignore-unavailable).
• It also broke the Windows jobs (4.5 / 4.6 / devel), which died at the multi-line Rscript -e '…' invocation — those jobs had been passing.

Both problems disappear here: the workflow is untouched, and the ignore is explicit.

Scope

DESCRIPTION only. Related: #2754 (run R-CMD-check on claude/**, stacked on this), #2749 (the R_getRegisteredNamespace C23 sanitizer error — a stale CSAN Docker image, not a package bug). The structure() code fixes merged in #2755.

🤖 Generated with Claude Code

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 1d82032 is merged into main:

• ✔️as_adjacency_matrix: 685ms -> 683ms [-1.7%, +1.38%]
• ✔️as_biadjacency_matrix: 635ms -> 630ms [-2.43%, +0.75%]
• ✔️as_data_frame_both: 1.4ms -> 1.41ms [-1.11%, +2.39%]
• ✔️as_long_data_frame: 3.66ms -> 3.63ms [-2.3%, +0.82%]
• ✔️es_attr_filter: 2.65ms -> 2.66ms [-1.92%, +2.12%]
• ✔️graph_from_adjacency_matrix: 127ms -> 126ms [-5.39%, +3.65%]
• ✔️graph_from_data_frame: 3.14ms -> 3.1ms [-3.83%, +1.1%]
• ✔️vs_attr_filter: 1.4ms -> 1.39ms [-2.41%, +1.12%]
• ✔️vs_by_name: 910µs -> 909µs [-3.12%, +2.93%]
  Further explanation regarding interpretation and methodology can be found in the documentation.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 6566387 is merged into main:

• ✔️as_adjacency_matrix: 884ms -> 883ms [-2.86%, +2.72%]
• ✔️as_biadjacency_matrix: 814ms -> 799ms [-4.51%, +0.81%]
• ✔️as_data_frame_both: 1.7ms -> 1.7ms [-2.42%, +2.66%]
• ✔️as_long_data_frame: 4.22ms -> 4.24ms [-2.55%, +3.59%]
• ✔️es_attr_filter: 2.83ms -> 2.83ms [-2.59%, +2.44%]
• ✔️graph_from_adjacency_matrix: 147ms -> 148ms [-0.9%, +1.48%]
• ✔️graph_from_data_frame: 3.67ms -> 3.62ms [-3.07%, +0.27%]
• ✔️vs_attr_filter: 1.63ms -> 1.61ms [-7.47%, +4.37%]
• ✔️vs_by_name: 1.04ms -> 1.06ms [-1.22%, +4.7%]
  Further explanation regarding interpretation and methodology can be found in the documentation.