Prevent reload realm barriers from stalling behind retired async work #850

Problem

The exact full-V8 navigation realm gate stalled on the first top-level replacement while a retired iframe/top-level fetch remained blocked in the host loader. The replacement was accepted, but the fresh realm could not service a trivial binary invocation.

Investigation localized the stall to cancel_detached_frame_context_tasks: erasing a pending fetch destroyed its std::async future while the isolate was held. The future destructor waits for the blocked loader, so cancellation synchronously blocked navigation. Full-realm teardown also built a weak wrapper graph that was immediately discarded later in the same retirement path.

Fix scope

• cancel retired-context fetches and clear their V8 context/resolver handles without destroying unfinished futures;
• let the ordinary ready-task drain reap cancelled host loads after completion;
• during full-realm retirement, run disconnect reactions and register detached roots for native reclamation without building redundant weak wrapper graphs;
• include the replacement index in barrier diagnostics.

Acceptance

• all 31 replacement realms complete with the host fetch deliberately blocked;
• retired timer, fetch, IndexedDB, MessagePort, Worker, and WebSocket callbacks remain absent;
• storage/cookies survive or isolate according to origin;
• stale globals/listeners/styles are absent;
• the first replacement remains below 250 ms and total replacements below 10 seconds;
• cancelled loads introduce no polling or idle wake loop.

Implementation

Implemented on fix/reload-realm-retirement-850 in commit 9f3fa40f. The focused full-V8 gate completed 31 replacements in 75 ms with unchanged heap and listener counts.