codegen: two sibling native-instance registration forms are still name-keyed and module-wide (follow-up to #9847)

[proggeramlug]

Follow-up to #9847 / #9857, which fixed one of three registration forms in
lower_assign that key a native-instance tag on the identifier text and
scope it to the whole module. The other two, and one fallback, were left alone
deliberately — #9847's A/B does not exercise them, and widening the change would
have put behaviour outside the evidence. This issue records that decision so the
remainder is documented rather than a loose end.

What #9857 fixed, and what it left

#9857 changed the bare-assignment form X = <native module>.<method>(...) to
key on the LocalId the target resolves to. Whole-bundle class-bearing
NativeMethodCall nodes in cli_2.1.112.js went 410 → 98
(child_process::Instance 194 → 0, fs::Instance 40 → 0, path::Instance
92 → 15), with the 70 named-class nodes unchanged as a registered control.

28 catch-all ::Instance sites survive. They split cleanly:

surviving site	count	assessment
stream::on	4	plausibly a genuine stream handle
https::on	3	plausibly genuine
https::destroy	3	plausibly genuine
stream::pipe, stream::once	2	plausibly genuine
https::write, https::end	2	plausibly genuine
subtotal — look correct	14
path::startsWith	7	wrong
path::slice	3	wrong
path::replace	2	wrong
path::toLowerCase, path::split, path::match	3	wrong
subtotal — still mis-typed	14

path has no instances at all, and every one of those 14 is a
String.prototype method. They are the same defect #9847 described, reaching
the binding through a route #9857 did not change.

The three remaining routes

All in crates/perry-hir/src/lower/expr_assign.rs, in the same
lower_assign prologue:

1. x = new NativeClass(...) — registers via both
   register_native_instance (scoped, name-keyed) and
   push_module_native_instance (module-wide, name-keyed). The scoped half
   limits the damage, but the module-wide half has the identical homonym
   collision.
2. x = y propagation — when y is a known native instance, the tag is
   copied to x with push_module_native_instance, module-wide by name. This
   one can also launder a tag past a scope boundary.
3. The unresolvable-name fallback in fix(hir): scope a bare-assignment native-instance tag to the binding, not the name (#9847) #9857 itself — when the assignment
   target resolves to no local (a bare global, or a module-level binding
   referenced from a body lowered before that binding was pre-registered), the
   tag is still registered and resolved by spelling. This is the documented hole
   fix(hir): scope a bare-assignment native-instance tag to the binding, not the name (#9847) #9857 kept, mirroring the one miscompile: an uncalled function containing new Proxy(arr, {}) makes an unrelated function's indexed read loop run zero iterations #7775 kept for proxies; it is strictly no worse
   than the previous behaviour, but it is not nothing.

I have not determined which of the three produces the 14 path sites.

The obvious shape, untested

The same LocalId treatment #9857 applied — resolve the assignment target with
ctx.lookup_local and register through register_local_id_native_instance,
keeping the name-keyed path only for a target that resolves to no binding. Form
1 would additionally want its register_native_instance half examined rather
than assumed. This is a hypothesis, not a plan: unlike #9847 it has no A/B
behind it yet, and the path/stream/https sites should be traced to their
producing route before anything is changed.

How to measure it

perry compile --no-auto-optimize --trace hir <bundle>, then count
NativeMethodCall { module: "M", class_name: Some("C"), … } occurrences. The
HIR dump is emitted before codegen, so the compile can be stopped as soon as
codegen: <file> appears on stderr — roughly 3 minutes rather than a full
build. Keep the named-class subtotal (70 on this bundle) as a control: a fix to
these routes must not move it.

Note on the catch-all

lower_assign's class-name match ends _ => Some("Instance"), so any method
on any recognised native module tags its target. Narrowing that is a separate
question from scoping — it changes which calls dispatch natively — and is not
proposed here, but it is why a single mis-scoped registration reaches so far.

[proggeramlug]

Fixed by #9904, landed on main via merge train #9922. The train merged as its own branch, so the close-keyword never fired — closing manually, verified on main.