[fallintoplace]

Summary

• keep init_device_contexts from clearing the thread-local device map while checking for duplicate initialization
• add a regression test showing duplicate init continues to fail after the first duplicate error

Why

ctx.devices.replace(None).is_none() used the duplicate-init predicate itself to replace the existing map with None. That made the first duplicate call return Context already initialized, but it also dropped the initialized state, so a later call could initialize again.

Validation

• cargo fmt --check
• git diff --check
• cargo test -p cuda-async duplicate_init_preserves_existing_device_contexts (blocked locally: CUDA toolkit headers are not installed; cuda-bindings could not find cuda.h under /usr/local/cuda)

[nihalpasham]

LGTM 👍🏾

• The old check wiped the live device map even on the error path; the fix restores it.

Reviewer note, verified: the added duplicate-init test passes and genuinely fails on the old code.