Share the failure check in project creation tests

[Deicyde]

Follow-up to #152 and #162. Test-only: tests/test_project_create.py goes from 2176 to 2039 lines (+85/-222).

• _fails. The failure tests that call create_project each spelled out pytest.raises, the call with the default package and release, and the code check. _fails(target, code, **options) does that and returns the error for message checks; _refused calls it and keeps its unchanged-parent check. 32 tests now call _fails and 8 more call _refused. The five pytest.raises blocks left call internal helpers, not create_project.
• Checks that got stronger. Four of the tests moved to _refused check more than before. test_rejects_nonsticky_shared_parent and test_version_conflict_and_floor_messages_name_the_problem now check that the parent is unchanged, and the second also checks its code, project-version-invalid. test_package_name_reserves_the_longest_lake_artifact_filename and test_macos_tmp_alias_is_rejected_but_private_tmp_is_supported compare the whole parent listing instead of checking one name. The other four keep equivalent checks.
• Interrupt after publish. test_interrupt_after_publish_reports_that_the_target_exists repeated test_fsync_failure_after_publish_reports_that_the_target_exists with KeyboardInterrupt in place of OSError, minus the message checks. The fsync test is now parametrized over both, so the interrupt case checks the message too. create_project calls fsync once after the rename, so raising on the first call after publish, as the fsync test does, covers what the old test's raise-on-every-call did.

No open PR edits this file, and the branch merges cleanly with all 12.

Validation at exact head cd4fee28: ruff check autoform_cli servers tests is clean, and tests/test_project_create.py passes (257).

[Deicyde]

Review at cd4fee2

Verdict: merge-ready. The refactor is test-only and no check got weaker. CI is green at this head.

How it was checked:

• Same cases. pytest --collect-only at main 89dff279 and at this head both give 257 cases. The only id change is test_interrupt_after_publish_reports_that_the_target_exists, which becomes test_fsync_failure_after_publish_reports_that_the_target_exists[interrupt]. CI selects nothing in this file by node id.
• Same calls. _fails defaults to package="Project", release_id=_RELEASE, which is what every replaced call passed explicitly. In test_workspace_substitution_fails_before_publication, versions always carries release_id. The surrogate test now passes a Path where it used to pass os.fspath(...). _validate_target calls os.fspath first (autoform_cli/project/create.py:452), so it sees the same string.
• Same checks.
  • Every replaced block keeps its exception type (ProjectCreateError), its code check and its message checks. One test used to compare (code, message) as a tuple; it now checks the code in _fails and the message after.
  • The 8 tests moved to _refused add a parent-listing comparison. That comparison covers each per-file check it replaces. In the two symlink tests, real starts empty, so listing through the link is the old not list(real.iterdir()).
  • Of the 8, test_never_overwrites_existing_target is also stronger than the description says: the old file-only snapshot missed a leftover empty stage directory, and the listing catches it.
• Interrupt merge.
  • After the rename, create_project calls os.fsync only once (create.py:219). The except BaseException and finally paths call only _publication_state and os.close. So raising on every fsync after publish (old) and raising on the first one (new) exercise the same path.
  • The interrupt case now also checks both message strings. They follow from _commit_uncertain("target", False, None) (create.py:281).
• Remaining pytest.raises blocks. The five left call internal helpers (_load_creation_release_descriptor, _parse_release_bundle, _reopen_bound_parent, _rename_noreplace, _open_parent), not create_project, as the description says.

Landing

• Onto main 89dff279: merges cleanly.
• Drop duplicate code from project inspection and the Lake metadata reader #174 and Share fixture and verdict helpers in the project inspection tests #180: all three merge cleanly onto main together, and tests/test_project_create.py there is byte-identical to this head. ruff check autoform_cli servers tests is clean. In the combined run, every tests/test_project_create.py case passed. The only failures were 10 cases in tests/test_project_inspect.py, and the same 10 fail on unchanged main in this checkout, which has no autoform_cli install (see the Drop duplicate code from project inspection and the Lake metadata reader #174 and Share fixture and verdict helpers in the project inspection tests #180 reviews). CI passes them.
• Publish blueprint sites transactionally #165: the description says no open PR edits this file, but Publish blueprint sites transactionally #165 (draft, head cef4a8ab) now does. It adds ".autoform-publication-*/" to the expected .gitignore lines, near line 240. A three-way merge of that change onto this head is clean. Publish blueprint sites transactionally #165 already conflicts with main in other files.
• Order: no constraint.

Posted by PR swarm: Review #173 #174 #180