Exercise native allocation failures through the public API

Decision: Valid hardening gap.

Evidence: The Python layer has 100% branch coverage, but native allocations use `std.heap.c_allocator` directly across callback queues, timers, DNS, transports, and process setup. Sanitizers and the lifecycle soak detect corruption and growth but do not inject allocation failures.

Impact: Partial-construction cleanup, fatal-error latching, and ownership rollback can regress without any deterministic test reaching those branches.

Next action: Introduce test-build allocator indirection or another deterministic failure boundary, then drive failures through public asyncio APIs rather than native private helpers.

Acceptance criteria: Public-API tests fail each allocation point in scheduling, timer, DNS, transport write, datagram send, and process construction paths and prove there are no leaked references, descriptors, handles, or resumable half-failed states.

Start by tracing the public asyncio APIs for scheduling, timers, DNS, transport writes, datagram sends, and process construction, then locate the native c_allocator uses behind those paths. Add deterministic test-build allocation failures and make public-API tests cover each allocation point, with no leaked references, descriptors, handles, or resumable half-failed states.

Written by the indexing model from the issue text.