js_class_method_bind still binds a movable heap string's interior in four places #7747 missed (timer handles, TextDecoder/TextEncoder)

[proggeramlug]

#7747 fixed two callers that bound a movable GC heap string's interior as a bound-method name. Four more sites in the same neighbourhood still do it.

The contract

js_class_method_bind(instance, method_name_ptr, method_name_len) stores the name pointer in the bound closure (capture 1) and dispatch_bound_method (closure/dispatch/bound.rs:12) re-reads it at call time. Its doc requires the pointer to stay valid for the closure's lifetime, which codegen satisfies with rodata.

#7747's commit message states the failure mode exactly:

get_field_by_name_tail passed key + size_of::<StringHeader>() — the interior of a movable GC heap string that is unreachable once the read returns — so a copying minor could relocate or reclaim the bytes the closure names. […] Whether the stale bytes still spell the method is an allocator property, not a program property, which is why this passed locally and took a SIGSEGV on conformance-smoke shards 7 and 8.

It fixed the two Buffer arms (object/field_get_set/buffer_own_prop.rs and the computed-key arm in object/polymorphic_index.rs), both of which now bind a 'static literal via buffer_dispatch::buffer_method_name_static.

Still unfixed, same file, same pattern

Each of these computes let key_ptr = (key as *const u8).add(size_of::<crate::StringHeader>()) and hands that pointer to js_class_method_bind:

• crates/perry-runtime/src/object/field_get_set/get_field_by_name_tail.rs:48 — timer-handle method key on a small-handle receiver (is_timer_handle_method_key + timer::is_known_timer_id).
• crates/perry-runtime/src/object/field_get_set/get_field_by_name_tail.rs:121 — the same arm for an already-stripped handle-band receiver.
• crates/perry-runtime/src/text.rs:649 and :670 (text_handle_property) — TextDecoder.prototype.decode and TextEncoder.prototype.encode/encodeInto read as VALUES. That function's own docstring says value reads are the reason it exists (K.decode.bind(K) — "the shape a minified SDK's cached decodeText helper takes"), so this is the intended hot path, not an edge. Reached from get_field_by_name_tail.rs:53 and :125, get_field_by_name.rs:875, and ic_miss.rs:540, all of which pass the same heap-string interior.

Not affected

crates/perry-runtime/src/symbol/get.rs:262 binds let mname = b"values" — a 'static byte-string literal. Fine as written.

Suggested shape

The same one #7747 used: a macro-generated *_method_name_static(name) -> Option<&'static str> for each surface (timer handles, TextDecoder, TextEncoder), so the bound closure names a literal rather than a borrow of its argument.

Testing note from #7747, which applies here too

The inequality against the key string could pass with the bug present […] Comparing the BYTES only fails on a host where the freed memory has already been reused, which is the lucky-allocator problem these tests exist to avoid. Identity with the literal cannot be lucky.

So the test should assert pointer identity with the static literal (see gc/tests/buffer_bound_method_name.rs), not that the bytes still read correctly.

Provenance

Found while bisecting #8117's two pass -> crash gap regressions. Those two bisect to #7314 and are a different defect — these sites are not on their path (their receivers are small handles, not Buffers). Filing separately so the finding is not lost inside that thread.