GC/threads: process-global address-keyed caches outlive thread arenas (3rd instance: #11319, #11462, #11470) — audit + gate #11471

One bug class, three instances in one day. Process-global state keyed by (or holding) a heap address outlives the thread whose arena owned that address. After the thread exits, Arena::drop frees its blocks, and another thread's arena can reuse the same addresses. The stale entry then describes a different object.

1. perry-stdlib lib tests: GC young-log rule-1 assertion fires in readline listeners_provider test on main #11319 / fix: perry-stdlib lib tests pass on clean main (#11417, #11319) #11422: closure side tables (process-global) kept a dead thread's entries, so a new closure at a reused address inherited them (young-log rule-1 abort). Fixed by release_closure_side_table_owners_in_ranges in Arena::drop.
2. fix(runtime): cache node:tls helper arrays per thread (fetch::lifecycle test flake) #11462: tls.rs cached tls.rootCertificates and sigalgs arrays in process-global AtomicU64 GC roots, which pinned whatever object later occupied the address (fetch::lifecycle flake; a dangling pointer in real perry/thread programs). Fixed by making them per-thread.
3. fix(runtime): a thread exit forgets its typed arrays in the global kind cache (#11463) #11470: PERRY_TA_KIND_CACHE / INLINE_OWNING_U32_CACHE (process-global, keyed by address) kept "this address is a typed array" after the owning thread exited, so a Promise at the reused address was misread as a Uint8Array (streams::pipe flake). Fixed by invalidate_kind_caches_in_ranges in Arena::drop.

Gap. scripts/gc_runtime_root_holders.py enumerates statics that can hold heap pointers as GC roots. It does not cover process-global lookup caches keyed by heap addresses, which aren't roots but are just as stale after arena reuse. Each instance so far was found by chasing a flake.

Ask.

• Audit every process-global cache or table in perry-runtime and perry-stdlib keyed by or storing a heap address, and make each one per-thread or invalidated in Arena::drop.
• Add a gate, like the root-holder inventory, that requires every such static to declare its thread-exit story, so the next one fails CI instead of flaking.