refactor: Purge legacy alias formals via the migration registry

[krlmlr]

Part of #2788 (PR 5 of the stack, on top of #2862); answers the survey's "migrate the existing hand-rolled alias formals into the registry" question for the aliases that are ready, so the signatures shrink to canonical names.

What this does

• layout_with_fr(): the trailing maxiter = deprecated() formal disappears; the registry now recovers maxiter = as a rename to niter. This alias was bridged silently before (no deprecation signal at all) — it now gets the standard 3.0.0 soft-deprecation.
• layout_with_kk(): start (alias of coords) cannot ride the registry because coords is a head formal, so the sentinel formal stays — but the bridge now emits the previously missing deprecate_soft() (it was a silent alias too), and the @param gains its deprecated badge.
• cluster_leiden(): gains its missing registry entry (it was the only cluster_*() without one, still on strict check_dots_empty()). The resolution_parameter = deprecated() formal and hand bridge are removed; the registry recovers it as a rename to resolution. The old entry reproduces the true v2.0.0 positional signature (verified against the v2.0.0 tag, commit 8102f30, where resolution_parameter was canonical slot 4), so both 2.0-era positional calls and 2.1-era keyword calls recover correctly.
• bfs() / dfs(): the neimode/father trailing formals — hard deprecate_stop() errors since 1.3.0/2.2.0 — are removed, together with the res$neimode <- res$mode and res$father <- res$parent return-value aliases (marked "Remove in 1.4.0" in the source). Calls using the dead names now fail check_dots_empty() (`...` must be empty) instead of the bespoke stop message. The graph.bfs()/graph.dfs() wrappers needed their unconditional neimode forwards dropped to keep working; their still-functional father → parent mapping is untouched.

Deliberately left for a follow-up

• make_lattice()/lattice() circular → periodic: these functions have no ellipsis migration yet; the alias should ride that migration when it happens (note the known dim/dimvector prefix hazard documented in the generator fixture).
• eigen_centrality()/centr_eigen()/centr_eigen_tmax() scale = deprecated(): dropped functionality (normalization is always on), not aliases — nothing to recover; removal is a later deprecation-advancement step (write code to advance lifecycle deprecations #1121).
• layout_with_fr()/layout_with_kk() zombie formals (coolexp, maxdelta, area, repulserad, sigma, initemp, niter-on-kk) — hard-stopped since 0.8.0; same advancement track. The niter/maxiter role swap between fr and kk remains a naming wart for the iterations-family decision in Comprehensive plan: argument renames for 3.0.0 #2788.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RTPj4qNv2FWui6etixZeuR

---

Generated by Claude Code

[github-actions]

This is how benchmark results would change (along with a 95% confidence interval in relative change) if e79a7b8 is merged into claude/issue-2788-subagents-faw1t1-4:

• ✔️as_adjacency_matrix: 842ms -> 843ms [-1.58%, +1.73%]
• ✔️as_biadjacency_matrix: 861ms -> 860ms [-1.19%, +0.87%]
• ✔️as_data_frame_both: 1.91ms -> 1.87ms [-6.52%, +2.03%]
• ✔️as_long_data_frame: 4.72ms -> 4.62ms [-4.11%, +0.2%]
• ✔️es_attr_filter: 2.95ms -> 2.95ms [-3.96%, +3.54%]
• ✔️graph_from_adjacency_matrix: 134ms -> 134ms [-1.71%, +1.77%]
• ✔️graph_from_data_frame: 4.17ms -> 4.15ms [-2.27%, +0.95%]
• ✔️vs_attr_filter: 1.8ms -> 1.8ms [-2.86%, +2.66%]
• ✔️vs_by_name: 1.12ms -> 1.13ms [-3.06%, +4.04%]
  Further explanation regarding interpretation and methodology can be found in the documentation.