test(io): cover interrupted parser cache writes

[hwendler]

Summary

Retain @hwendler's interrupted-cache-write regression for AtomWorks 3.0.

The release parser already implements atomic cache publication: it serializes to a unique temporary file, explicitly selects compression, replaces the destination only after serialization succeeds, and cleans up on interruption. The original implementation change and unrelated documentation formatting are therefore removed from this PR's diff.

The remaining change is one test in tests/io/components/test_cache_atomic.py. It writes partial bytes and raises KeyboardInterrupt during serialization through the public parse(..., config=ParseConfig(...)) API, then checks that neither a cache entry nor any temporary file remains. It uses the checked-in 2HHB fixture and requires no external structure download. This complements the existing tests for failure during final file replacement.

Validation

• Five focused cases passed: the new interruption regression, two existing buffer/cache round trips, cache-key disambiguation, and cache-directory defaults.
• Ruff check and format check passed for the new test file.
• Release-relative diff contains only the new test file.

Contributor commits are preserved; no force push or history rewrite. Target: release/atomworks-3-0.

[nscorley]

Thanks for this fix! Just kicked-off the tests - once those pass likely good to merge

[nscorley]

Thank you very much for catching this problem and putting up this PR! A few relatively minor comments, then would be excited to merge!

[m1c0l]

The interrupted-cache-write problem reported and fixed here by @hwendler is also addressed independently in the AtomWorks 3.0 parser on release/atomworks-3-0 (replacement writer).

I validated that replacement implementation against this PR's failure scenario:

• Interrupting serialization after writing partial bytes publishes no cache entry and removes the temporary file.
• An interrupted replacement preserves the previous complete, readable entry.
• Actual SIGKILL during serialization never publishes a partial target, both with and without an existing entry. A subsequent write succeeds.
• A real 2HHB parse/cache round trip succeeds, including coordinates, annotations and bonds; the cache hit is verified without invoking the CIF loader.

All three focused regression tests passed (the SIGKILL test covers both destination states). As expected, SIGKILL can leave an orphan temporary file, but the published cache remains valid or absent.

Credit to @hwendler for identifying the production failure and providing the interrupted-write regression. This validates the 3.0 replacement; it does not imply the fix is already on production or that this PR was merged.