js_class_method_bind: perry-stdlib's handle-property dispatch has the same interior-pointer bug in eight sub-dispatchers

[proggeramlug]

Summary

#8133 named four js_class_method_bind sites binding a movable heap string's interior. Fixing it (#8177) surfaced six in perry-runtime — and the same defect systemically in perry-stdlib's handle-property dispatch, where eight sub-dispatchers capture the forwarded name pointer directly.

The contract being violated

js_class_method_bind(instance, name_ptr, name_len) stores the name pointer in the bound closure's capture, and dispatch_bound_method (closure/dispatch/bound.rs:12) re-reads it at call time. So the pointer must outlive the bind — it cannot point into a movable GC heap string, whose contents are unreachable once a collection moves it.

Where it still happens

Eight sub-dispatchers in perry-stdlib forward the caller's pointer straight into the bind:

• sqlite — four sites
• tls — two sites
• emitter_als — two sites

Reachable from ordinary code that stores a method rather than calling it:

const f = db.run;        // sqlite
const g = emitter.on;    // emitter_als
const h = als.getStore;  // emitter_als

Also found on perry-runtime's primitive-receiver arm ((5).toString), which #8177 did not cover.

Why this will not show up by accident

#8177 established the reproduction conditions, and they are narrow:

• A literal dec.decode lowers to rodata and never reaches these arms at all. Only a computed key does — const k = "dec" + "ode"; obj[k].
• Pre-fix, the symptom was a silent wrong answer: decode=undefined followed by a throw, where node prints decode=hi. No instrument fired.
• Under PERRY_GC_ZEAL=1 PERRY_GC_PROTECT_FROMSPACE=1 it became a SIGBUS reported as RETIRED FROM-SPACE … retired_by_minor=#0.

So a test that binds a literal name and calls it immediately will pass on broken code. Any fixture here needs a computed key and enough allocation to actually move the string; #8177's used ~400k allocations.

Suggested shape

#8177's fix is worth copying rather than re-deriving: it replaced the is_timer_handle_method_key predicate with a 'static-returning lookup and removed key_ptr/key_len from text_handle_property entirely. With no predicate and no pointer parameter, the bug is unwritable rather than merely fixed. The stdlib dispatchers can take the same treatment.

Related

#8133 (closed by #8177), #7747 (the first two sites), docs/src/internals/gc-rooting-invariant.md.