perf: Share the graph weak-reference across constructed sequences

[schochastics]

Split 1 of 4 out of #2696, per #2696 (comment).

Base: main. Next in the stack: #2887.

Why this one is not in the requested split

The review asked for three PRs (create_vs_list(), then C, then ALTREP). Profiling the split showed a fourth piece that belongs first, because it is where most of the speedup actually comes from and it is independent of all three: sequence construction was minting a graph weak-reference per object and setting attributes one at a time.

	max_cliques(sample_gnp(200, 0.16)), named graph
main	24.7 ms
this PR	7.8 ms

That is ~3.2× before a single line of create_vs_list() or C code exists.

What changed

1. One shared weak reference per graph instead of one per sequence.

add_vses_graph_ref() runs three .Call()s (Rx_igraph_copy_env, Rx_igraph_make_weak_ref, Rx_igraph_get_graph_id) for every object it touches. Functions returning thousands of sequences paid that per sequence — profiling put it at ~75% of construction time.

The weak reference's key is the graph's environment, which is identical for every sequence of a graph (Rf_duplicate() is a no-op on an environment, so get_vs_ref() hands back the same env each call). A single shared reference is therefore semantically identical to one per object: while the graph is alive the reference resolves; once the graph is released the reference reports it gone.

simple_es_index() already propagated env/graph from its input, so the edge path only needed the redundant per-object call removed. simple_vs_index() now propagates them the same way, so the existing lapply(res, unsafe_create_vs, graph = graph, verts = V(graph)) call sites share the one reference minted by V(graph) — no call-site or codegen changes.

2. Attributes set in one pass. Each incremental attr<-/class<- shallow-copies the vector. simple_vs_index() now sets names/class/env/graph in a single attributes<-.

3. unsafe_create_vs() skips a redundant copy. Its idx are vertex IDs straight from C and verts is always the full V(graph), so verts[idx] just reproduces idx. It now uses the IDs directly as the integer payload and subsets the names off verts, instead of copying V(graph) per object.

Correctness

Payload type (integer), names, NA handling and graph recovery are unchanged. Verified that get_vs_graph(seq) still returns NULL after rm(graph); gc() — the shared reference is still weak.

Full suite: FAIL 0 | WARN 0 | SKIP 7 | PASS 9299.

Also here

Touchstone group #5 covering the construction path on named graphs (max_cliques_named, head_of_named, V_named, E_named), written in the warm-up-then-loop style the script's header requires, with literal repeat counts targeting ~100 ms on main.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if 1220792 is merged into main:

• ✔️E_named: 234ms -> 234ms [-1.8%, +1.54%]
• ✔️V_named: 439ms -> 439ms [-1.15%, +0.76%]
• ✔️as_adjacency_matrix: 246ms -> 246ms [-1.08%, +0.97%]
• ✔️as_biadjacency_matrix: 265ms -> 263ms [-1.63%, +0.5%]
• ✔️as_data_frame_both: 267ms -> 268ms [-0.39%, +0.95%]
• ✔️as_long_data_frame: 212ms -> 214ms [-0.2%, +1.61%]
• ✔️es_attr_filter: 251ms -> 252ms [-0.41%, +1.74%]
• ✔️graph_from_adjacency_matrix: 280ms -> 279ms [-1.67%, +1.08%]
• ✔️graph_from_data_frame: 283ms -> 281ms [-1.58%, +0.31%]
• ✔️head_of_named: 186ms -> 187ms [-0.7%, +1.88%]
• 🚀max_cliques_named: 192ms -> 71ms [-65.03%, -61.11%]
• ❗🐌vs_attr_filter: 339ms -> 342ms [+0.15%, +1.95%]
• ✔️vs_by_name: 320ms -> 322ms [-0.34%, +1.51%]
  Further explanation regarding interpretation and methodology can be found in the documentation.