Retire the twelve hand registrations the generated registrar shadows

[mparrett]

Child of #531 (S2). Fast-follow from #686.

#686 landed TestNoNewShadowedHandRegistrations, which pins the set of names registered twice: once by hand in installLangNS, once by a //lg:native decl in native_prims.go. The generated registrar drains after installLangNS, so it takes the var root and the hand-written closure never runs. That list is a ratchet against new duplicates; it does not resolve the existing ones.

Twelve names are on it:

name	hand registration in lang.go	generated decl
conj	ns.Def("conj", conj)	Conj
deref	ns.Def("deref", deref)	Deref, Deref3
get	ns.Def("get", get)	Get, Get3
int	ns.Def("int", intf)	Int
name	ns.Def("name", name)	Name
namespace	ns.Def("namespace", namespace)	Namespace
nth	ns.Def("nth", nthf)	Nth, Nth3
pop-binding!	ns.Def("pop-binding!", popBinding)	PopBinding
push-binding!	ns.Def("push-binding!", pushBinding)	PushBinding
some	ns.Def("some", some)	Some
str	ns.Def("str", str)	Str
subs	ns.Def("subs", subs)	Subs, Subs3

Why close them

reduce was the thirteenth. Its two bodies had drifted: the closure carried an ArrayVector fast path, a Range fast path, and a reused argument buffer that the //lg:native decl never got. When #639's namespace-alias fix made the generated registration resolve for the first time, reduce silently switched to the naive body and got 1.88x slower, with no diff to reduce and no failing test. TestReduceRangeFastPath stayed green: it asserts results, not which path produced them (#656).

A second copy that nothing executes has no pressure keeping it honest.

What the audit found

All twelve pairs were compared body for body during #686. The //lg:native decls are faithful ports, several carrying the closure's comments verbatim. No name is losing behavior today, so this is deletion work rather than porting work, and it can go at whatever pace suits.

Shape of the work

Per name, or per small family:

1. Delete the ns.Def site.
2. Delete the Go func it referenced, once nothing else calls it.
3. Drop the name from knownShadowedHandRegistrations.

The ratchet keeps the list honest in both directions: a new duplicate fails the test, and a listed name that is no longer double-registered fails it too, so the list cannot go stale.

Done when the accepted set is empty and the test asserts emptiness directly.

Gate

ns-publics diff empty, pkg/rt green under go test -count=2. The failure mode to watch for is deleting a var's only surviving registration, which the ns-publics diff catches.

[mparrett]

One of the twelve, some, now has a second in-flight change: the perf/vm-prepared-call branch (#720, stacked on #719) rewrites the hand-written rt.Some in native_prims.go to invoke bytecode predicates through PreparedCall, which is what recovers the #700 regression on that shape. Whichever body survives this retirement needs to carry that adoption — retiring the hand registration in favor of a generated body that calls ec.Invoke per element would silently reintroduce the per-element entry cost.

Same criterion for the broader #697 drain: a callback-taking native needs PreparedCall (or an in-bytecode loop) or it ships a per-element boundary tax; the non-callback names among the twelve are unaffected.