fix: as_adjacency_matrix(attr = NULL) is unweighted again

[krlmlr]

attr and weights spell "unweighted" differently, and the deprecation shim forwards one to the other unchanged.

• attr = NULL was the documented way to ask for a plain 0/1 matrix. CRAN 2.3.3's ?as_adjacency_matrix says: "attr: Either NULL or a character string ... If NULL a traditional adjacency matrix is returned."
• weights = NULL means the opposite: use the weight edge attribute if the graph has one.

resolve_edge_weights() did weights <- attr, so every call that spelled out attr = NULL started returning a weighted matrix — silently, with no warning beyond the generic "use weights instead". A graph having a weight attribute is exactly the case where a caller bothers to ask for an unweighted matrix, so the inversion hits precisely the calls that meant it.

lifecycle::is_present(NULL) is TRUE — only the deprecated() sentinel counts as absent — so the branch really does see an explicit attr = NULL and has to translate it.

What changed

• attr = NULL now maps to weights = NA (the new vocabulary for "explicitly unweighted"). A character attribute name is still forwarded unchanged.
• The deprecation warning gains a detail line for the NULL case: `attr = NULL` becomes `weights = NA`, not `weights = NULL`. Following the bare advice literally would reintroduce the same inversion in user code.
• @param attr documents the translation instead of claiming the value is always "a character edge attribute name".
• Tests for both callers of resolve_edge_weights() — as_adjacency_matrix() and as_biadjacency_matrix() — plus one for the warning text.

Reproducer

g <- make_graph(c(1, 2, 2, 3), directed = FALSE)
E(g)$weight <- c(0.5, 0.7)
as_adjacency_matrix(g, attr = NULL, sparse = FALSE)

CRAN 2.3.3 and this branch return a 0/1 matrix; main returns the 0.5/0.7 one.

Where this came from

Found while triaging the reverse-dependency failures in #2646. bnstruct calls as_adjacency_matrix(ig.mst, type = "both", attr = NULL) in buildJunctionTree-methods.R, on a minimum spanning tree that carries weights — the whole point of the call is to get the topology without them. Introduced by a6ace15 (#2677).

What this does not fix

The attr = NULL spelling is unambiguously a bug: the shim inverts an explicit request. The default is a separate, deliberate-looking question that this PR leaves alone.

as_adjacency_matrix(g) with no weight argument at all also changed meaning — weights = NULL picks up weight, where attr's default NULL did not. That is the item #2646 raises for sfclust, and it is also what breaks MetaNet: within_module_deg_z_score() and part_coeff() call as_adj(g, sparse = FALSE) in their weighted = FALSE branch and as_adj(g, sparse = FALSE, attr = "weight") in the other, so the unweighted branch now returns the weighted matrix and the Zi/Pi values land outside every role band.

Whether the new default is the intended 3.0.0 behaviour is a maintainer call — it is the same shape as the modularity() item in #2646, and it needs a NEWS entry either way. Flagging it here rather than changing it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01D1xpHRV7yVfgtJg4vp9P7z

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 0fa7f52 is merged into main:

• ✔️as_adjacency_matrix: 807ms -> 797ms [-3.48%, +0.93%]
• ✔️as_biadjacency_matrix: 802ms -> 803ms [-2.42%, +2.69%]
• ✔️as_data_frame_both: 1.74ms -> 1.74ms [-2.08%, +2.15%]
• ✔️as_long_data_frame: 4.09ms -> 4.16ms [-3.64%, +7.39%]
• ✔️es_attr_filter: 2.76ms -> 2.76ms [-1.73%, +1.63%]
• ✔️graph_from_adjacency_matrix: 146ms -> 145ms [-1.44%, +1.12%]
• ✔️graph_from_data_frame: 3.76ms -> 3.92ms [-3.48%, +11.6%]
• ✔️vs_attr_filter: 1.57ms -> 1.56ms [-2.16%, +1.19%]
• ✔️vs_by_name: 1.03ms -> 1.1ms [-3.16%, +16.72%]
  Further explanation regarding interpretation and methodology can be found in the documentation.