Resolves the FIX-ω obligation D-093 §7 and D-094 §5/§7 both named as
binding on "whoever implements delay/Later next" and explicitly
declined to make — resolved here, not deferred a third time.

Tested and falsified a plausible hypothesis before designing anything
new: reg <- expr is NOT an informal Later judgment. Checked directly
against StatementKind::RegisterUpdate — it type-checks the update
value against the register's plain element type T with ordinary
same-cycle equality; the only "guarded-feedback vocabulary" D-092/
D-093 referenced turns out to be a single boolean bypass in the
unrelated mem-taint pass, not a modal type. Later<Clock,T> is
therefore a real new judgment, not existing machinery to formalize.

v0 needs zero new syntax — delay(init, next) already parses as an
ordinary Call expression — but does need a new CheckedExprKind::Delay
node and a new TypeKind::Later variant (investigated and rejected
reusing TypeKind::Stamped: that's D-014 epoch-generativity, a
different axis from clock-cycle timing). Scoped to one non-nested
delay call per component, forced by the committed signature itself
since no fixture in the corpus uses delay/feedback at all — the first
doc in this series with no forcing example to scope against.

FIX-ω resolved as branch (b): any linear-tracked binding referenced
inside a delay's next argument is rejected outright, not
ω-promoted. Branch (a) would require first deciding what "consumed
via feedback" means for a grade-1 binding's end-of-scope obligation —
a real separate design question, not a good bundle for this v0. (b)
is sound, fully specified, needs no new UseCount variant, and closes
OP-7's exact gap; (a) is named north star, not another punt.

Lowering is unchanged from today's reg/seq.firreg path — Later<'d,T>
is a purely compile-time causality proof with zero runtime
representation once checked sound.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>