perf: Build the whole vertex-sequence list in one C pass

[schochastics]

Split 3 of 4 out of #2696, per #2696 (comment) — this is the requested "C implementation of create_vs_list() without ALTREP" step.

Base: #2887 (review #2886 and #2887 first; this PR's diff is only the top commit). Last in the stack: #2696.

What changed

After #2887, create_vs_list() still drove the per-element work from R — an lapply closure call, as.integer(), a name subset and attributes<-, once per sequence. That is the dominant remaining cost when a function returns thousands of igraph.vs objects.

Rx_igraph_vs_list() in rinterface_extra.c does the whole loop in one pass. R only builds V(graph) once (that is what mints the shared weak reference and graph id) and hands the pieces to C:

create_vs_list <- function(graph, idx_list) {
  verts <- V(graph)
  .Call(
    Rx_igraph_vs_list,
    idx_list,
    if (is_named(graph)) vertex_attr(graph)$name else NULL,
    attr(verts, "env"),
    attr(verts, "graph")
  )
}

Each element gets a fresh integer payload, the corresponding names when the graph is named, the shared env/graph attributes, and the igraph.vs class. Names are built eagerly here — no ALTREP, as requested. That is the only thing #2696 changes on top.

Note the payload is duplicated when Rf_coerceVector() is a no-op, so a caller-owned integer vector is never mutated in place.

max_cliques(sample_gnp(200, 0.16)) on a named graph: 5.6 ms -> 2.65 ms.

Where that leaves the stack

	median	vs main
main	24.7 ms	—
#2886 (shared weakref)	7.8 ms	3.2×
#2887 (create_vs_list())	5.6 ms	4.4×
this PR (C loop)	2.65 ms	9.3×

So the three merged PRs get essentially the whole win without any ALTREP. This is the point at which #2696's "if needed" can be judged on its own merits.

Correctness

Verified against the return.vs.es = FALSE numeric path:

• integer payloads with identical values and names;
• no names attribute at all on unnamed graphs;
• NA and out-of-range IDs map to NA_STRING;
• double ID vectors coerced correctly;
• caller's input vectors unmutated;
• the shared weak reference still releases the graph: get_vs_graph(seq) is NULL after rm(graph); gc().

Compiles without new warnings, and clean under gctorture(TRUE).

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

Also here

cpp11::cpp_register() added only the Rx_igraph_vs_list registration in src/cpp11.cpp. Two touchstone benchmarks for the batch path: ego_order2_named and all_simple_paths_named.

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if dd0e928 is merged into perf/vs-create-vs-list:

• ✔️E_named: 223ms -> 222ms [-2.11%, +1.26%]
• ✔️V_named: 504ms -> 509ms [-1.39%, +3.23%]
• 🚀all_simple_paths_named: 208ms -> 133ms [-36.53%, -35.9%]
• ✔️as_adjacency_matrix: 244ms -> 245ms [-0.38%, +1.2%]
• ✔️as_biadjacency_matrix: 261ms -> 262ms [-0.42%, +0.92%]
• ✔️as_data_frame_both: 251ms -> 252ms [-0.57%, +0.65%]
• ✔️as_long_data_frame: 194ms -> 194ms [-0.51%, +0.54%]
• 🚀ego_order2_named: 221ms -> 146ms [-34.33%, -33.68%]
• 🚀es_attr_filter: 234ms -> 233ms [-1.29%, -0.07%]
• ✔️graph_from_adjacency_matrix: 280ms -> 280ms [-0.73%, +0.47%]
• ✔️graph_from_data_frame: 262ms -> 262ms [-0.69%, +0.62%]
• 🚀head_of_named: 174ms -> 172ms [-1.73%, -0.82%]
• 🚀max_cliques_named: 54.8ms -> 25.2ms [-54.37%, -53.69%]
• ✔️vs_attr_filter: 301ms -> 301ms [-0.61%, +0.59%]
• ✔️vs_by_name: 254ms -> 254ms [-0.26%, +0.71%]
  Further explanation regarding interpretation and methodology can be found in the documentation.