refactor: convert h5py writer to with self.parent.indented():

[devdanzin]

Completes the generator indentation-API migration (#145, #146; tracked under #106) for the last and largest holdout: the h5py code writer (fusil/python/h5py/write_h5py_code.py).

Every manual addLevel(1)/restoreLevel(...) pair — ~150 call sites across all 9 emitting methods — is replaced with the with self.parent.indented(): context manager, so each block's nesting mirrors the generated code's nesting and a level can't leak. The file now has zero manual addLevel/restoreLevel calls. Net −997/+814 lines (the try/finally + level-variable boilerplate goes away).

Methods converted

_dispatch_fuzz_on_h5py_instance (preserving its entry-level return contract that the parent dispatcher relies on), fuzz_one_h5py_class, _write_h5py_file, _write_h5py_dataset_creation_call, _fuzz_one_attributemanager_instance, _fuzz_one_file_instance, _fuzz_one_group_instance, and the large _fuzz_one_dataset_instance (~800 lines).

Verification — all behaviour-preserving

The h5py writer is not exercised by the golden-output test (it needs numpy+h5py, which the golden harness disables). So this was verified with a dedicated before/after diff harness that drives all 9 methods and snapshots ~1.36 MB of generated code:

• byte-identical before vs. after, under a pinned PYTHONHASHSEED and a seeded uuid4 (both otherwise per-process random and would mask the comparison).
• Full suite 328 OK; ruff check + ruff format --check clean.

With this merged, the entire Python-fuzzer generator uses the context manager — the only remaining manual idiom anywhere is the single genuine callee-opens-a-level contract between _dispatch_fuzz_on_instance and the h5py dispatcher.

🤖 Generated with Claude Code