codegen: a constructor bail-out lowers the body and then discards its globals and declares (method.rs:928)

[proggeramlug]

crates/perry-codegen/src/codegen/method.rs:917-931 — the "parent class has no
callable constructor symbol" bail-out inside constructor lowering — lowers the
method body into the module and then discards every global and declaration that
body produced:

} else {
    // No callable ctor symbol — bail.
    stmt::lower_stmts(&mut ctx, method_body).with_context(|| {
        format!("lowering body of method '{}::{}'", class.name, method.name)
    })?;
    // Fall through to the default ret at end.
    if !ctx.block().is_terminated() { … ctx.block().ret(DOUBLE, &undef); }
    let _ = std::mem::take(&mut ctx.ic_globals);
    let _ = std::mem::take(&mut ctx.typed_parse_rodata);
    let _ = std::mem::take(&mut ctx.pending_declares);
    return Ok(());
}

The three std::mem::takes discard the values; the normal exit at
method.rs:1306-1343 does the opposite with the same three — it drains them
into llmod as inline_cache_global_definitions, raw globals and
declare_functions.

The lowered body is not discarded with them. ctx writes its blocks into
the function llmod already holds, so on this path the module keeps IR that
references:

• every @perry_ic_N inline-cache slot the body's property accesses allocated
  (expr::inline_cache_global_name / inline_cache_global_definition),
• every raw global pushed to typed_parse_rodata — today that is
  expr/strings.rs, expr/array_literal.rs, expr/object_literal.rs,
  expr/typed_feedback.rs, expr/instance_misc1.rs, expr/v8_interop.rs and
  concat_site_cache.rs,
• every runtime symbol queued in pending_declares,

none of which is ever defined or declared.

How it fails

Loudly, at the in-process LLVM parse, with use of undefined value '@…' — the
same shape as #9859, where five js_segments_view_* calls were emitted with no
declare and twelve passing HIR-level unit tests could not see it. It is not a
miscompile and cannot produce a wrong answer at runtime.

There is a second, quieter consequence: this path returns without running
llmod.ic_counter = ic_end, so the site ids this body consumed are handed back
out to the next function in the module and two sites can be issued the same
@perry_ic_N / @perry_concat_site_N name → an LLVM redefinition error.
Also loud.

Why it has not fired

Presumably every constructor that reaches this arm today has a body trivial
enough to allocate no IC site, no rodata global and no pending declare. That is
a property of the corpus, not of the code, and nothing enforces it.

How I found it

Not from a failure — from reading the drain sites before adding a new one. The
regex literal-site key (perf/regex-literal-site-key) emits one
@perry_regexp_site_<prefix>__<n> per regex literal through
typed_parse_rodata, so a regex literal inside such a constructor would be the
next victim. The new lowering carries a comment pointing here so the next person
to see use of undefined value knows where to look.

Suggested fix

Drain instead of discard: hoist the ic_globals / typed_parse_rodata /
pending_declares emission (and llmod.ic_counter = ic_end) into a single exit
helper that every return from this function goes through. A regression test
wants a class whose parent has no callable constructor symbol and whose own
constructor body contains something that allocates a global — an array literal
or a property read is enough — and asserts the emitted module defines every
global its IR names.

Filed by the regex-subsystem lane of the cc-perf campaign; not blocking that
work, since the failure mode is a compile error rather than a wrong answer.

[proggeramlug]

Fixed by #9896, landed on main via merge train #9901. The train merged as its own branch, so the close-keyword never fired — closing manually, verified on main.