Fix unbounded worker task leak from the chain-contention registry — Closes #261

[conradbzura]

Summary

Change the task factory's module-global chain-contention registry, _task_contexts, from a strong dict[int, asyncio.Future] to a weakref.WeakValueDictionary, so a registered task's entry drops as soon as the task is garbage-collected — even when the add_done_callback(_release) cleanup never runs because the per-dispatch worker loop (ttl=0) was torn down before the callback fired. This bounds worker-side task and registry bookkeeping to in-flight dispatches instead of lifetime dispatch count, removing the dominant driver of the 0.9.3 → 0.10.0 dispatch-latency regression (an ever-growing global task set made the per-dispatch asyncio.all_tasks() teardown scan O(N-growing)). The _PENDING reservation sentinel and the chain-contention guard semantics are unchanged.

Closes #261

Proposed changes

Hold registered tasks weakly (runtime/context/factory.py)

Replace the strong _task_contexts dict with a weakref.WeakValueDictionary. Entries now drop when their task is collected regardless of whether _release fires, so a stranded done-callback (worker loop destroyed first) no longer pins the task forever. _PENDING remains a strongly-referenced module singleton, so reserved slots never evaporate; the id-reuse safety invariant still holds (a live weak entry implies its task, and thus its context, is still alive); _release remains an eager best-effort cleanup and the factory-displacement backstop.

Test cases

#	Test Suite	Given	When	Then	Coverage Target
1	test_factory	An armed task registered on a worker loop whose _release callback is stranded by loop teardown	The last strong reference is dropped and gc.collect() runs	Its weakref clears and its registry key is gone — the leak is bounded (fails on the old strong dict)	Weak-registry eviction
2	test_factory	A _PENDING reservation held during a re-entrant create_task	gc.collect() runs inside the reservation window	The re-entrant call still raises ChainContention and the original task completes	_PENDING survives GC
3	test_factory	A live task running under an explicit armed context	The context is re-passed after a forced gc.collect()	ChainContention still fires — a live task pins its own weak registry value	Guard intact after GC