Reworked per review direction: no grandfather mechanism.
Two commits:

1. refactor: move the 23 non-constant defaults into the bodies

All non-constant defaults across the 22 migrated functions on main
become constant NULL in the signature,
resolved in the body once all arguments are available:

• V(graph) / E(graph) selectors (19 arguments):
  alpha_centrality(), betweenness(), closeness(), diversity(),
  edge_betweenness(), harmonic_centrality(), page_rank(),
  power_centrality(), strength(), all_simple_paths(),
  all_shortest_paths(), constraint(), degree(), distances()
  (v and to), ego(), ego_size(), knn(), make_ego_graph(),
  shortest_paths(), which_mutual();
• as_adjacency_matrix(sparse = igraph_opt("sparsematrices"));
• max_bipartite_match(eps = .Machine$double.eps).

Body pattern, inserted right after the ARG_HANDLE block
(after ensure_igraph() where present):

if (is.null(vids)) {
  vids <- V(graph)
}

Behavior notes for review:

• Passing the selector NULL explicitly now means the full vertex/edge
  set, same as omitting the argument.
  Previously an explicit NULL (including a computed c() that came out
  empty) was coerced to an empty selection — an undocumented, untested
  accident of as_igraph_vs(). The new semantics are pinned in
  test-constant-defaults.R; if you'd rather keep the accidental
  empty-selection meaning, say so and I'll switch the resolution to a
  missing() check instead.
• Watch the interplay with Standardize the handling of optional vertex parameters #1572, which settled on NULL meaning
  "no vertex" for optional single-vertex parameters
  (random_spanning_tree(vid = NULL)). Plural selectors saying
  "NULL = all" vs. singular saying "NULL = none" is a distinction
  worth an explicit line in the docs policy.
• This fixes a live bug on main: legacy positional recovery of
  diversity() died with a spurious "supplied more than once" error
  because re-evaluating its V(graph) default never compares
  identical(). With a constant default the recovery machinery needs no
  further changes — the defect class fix(migrations): detect supplied tail args with missing() #2783 addressed disappears at the
  root, per "avoid nonconst defaults at all".

2. feat: the generator enforces the rule — no escape hatch

is_constant_default() (recursive AST check): constant = literals,
NULL/TRUE/FALSE/NA*/Inf/NaN, c()/list() of constants,
typed empty vectors (integer(), numeric(), ...),
unary sign, parentheses, the deprecated() sentinel.
Everything else errors at generation time —
option lookups, V(graph), cross-references (bins * 2), RNG draws,
rep(), .Machine$…, bare symbols including T/F.
A leftover nonconst_defaults field errors with its own message.
Documented in tools/migrations/README.md
(and the matching rule prose is on #2757).

Round 2: the whole package

After merging main (which migrated the games/layout/make/… topics
behind the ellipsis), the same treatment now covers every exported,
non-deprecated, hand-written function — the remaining 106 non-constant
defaults across 67 functions, migrated or not, plus the two
plot_dendrogram() S3 methods for coherence with their generic.
The registry new literals of the 32 affected migrated functions carry
the same NULL defaults, so the regenerated recovery blocks stay
constant.

Special cases, handled deliberately: cross-referencing defaults resolve
in signature order (area before repulserad/cellsize, types
before pref, agebins before pref); layout_with_drl()'s random
default seed is now drawn inside the function with unchanged RNG stream
consumption for defaulted calls; options = arpack_defaults() resolves
in the body; layout_with_gem(temp.min = 1/10) and
plot.igraph(mark.shape = 1/2) became the literals 0.1/0.5 instead
of NULL; sample_motifs(cut.prob = NULL) needs no resolution because
the C layer already treats NULL as "no cuts" (matching motifs());
and two places keep NULL as a legal value and therefore resolve via
missing() instead of is.null(): print.igraph(max.lines = NULL)
still prints everything, and normalize(ymin = NULL) still disables
normalization along that axis (as in norm_coords()). The
explicit-NULL flip described above now also applies to the
option-backed combination specs (edge.attr.comb = NULL etc. used to
mean an empty combination spec, dropping attributes).

c() defaults were NULL in disguise

c() evaluates to NULL. The five empty-HRG fields that fit_hrg()
builds when no starting model is given (formerly hrg()'s signature
defaults left = c(), right = c(), …) therefore declared NULL values
that historically meant "empty sequence". Under the new doctrine NULL
is the resolve-in-body sentinel, so NULL-as-empty and NULL-as-default
cannot coexist: empty-sequence defaults must be spelled as typed
empty vectors (numeric() here — the fields are as.numeric()-coerced
on the spot; layout_as_tree()'s root = numeric() was already
spelled right and the checker now accepts it as a literal). Tests pass
integer() where an empty id selection is meant
(max_degree(g, v = integer()) is 0). Passing NULL explicitly to
mean "empty" is no longer supported for treated arguments — a
deliberate breaking change: max_degree(g, v = NULL) now means all
vertices, pinned in test-constant-defaults.R.

Remaining worklist (not this PR)

The exported hand-written remainder is zero: scanning every
function definition against the classifier finds no non-constant
default outside R/aaa-*.R. What is left is the Stimulus-generated
_impl layer — 103 defaults across 78 internal wrappers — and handling
those in the generator (constant-ness check there, or dropping defaults
from _impl signatures entirely) is still under discussion.

Validation: generator idempotent (generate → air → generate is a
byte-identical fixed point); full suite FAIL 0 · PASS 9250 (the only
error is the known sandbox network fetch in test-foreign.R);
gate-on IGRAPH_LIFECYCLE_ERRORS=true migration/constant-defaults
filter FAIL 0 · PASS 151; 85 Rd files regenerated via
devtools::document().