[stevevanhooser]

Syncs DID-python to VH-Lab/DID-matlab#227 (bb39399). That merge makes "Bridge coverage vs DID-matlab" report drift on sqlitedb.m and database.m for every DID-python PR, and this clears it: the check reports 0 problems again.

Most of #227 (one transaction per write operation, replay after reopen, and the pre-commit replaced-file check) mirrors what is already here from #93 and #94. Two parts weren't, and are ported here.

Changes

• delete_original sources are deleted only after the write operation commits. _delete_after_commit queues them and _journal_committed deletes them, instead of deleting during ingest, so a rollback no longer leaves the only copy of a file unreferenced in FileDir. Reclaimed cached copies already waited for the commit. A source that two files share is deleted once, and one that's already gone isn't a warning, matching MATLAB's suppression of MATLAB:DELETE:FileNotFound.
• _recover_transaction doesn't reopen a database file that no longer exists. Reopening created a new, empty database at the path. It now raises the original error with FileNotFoundError as its cause, matching MATLAB's DID:SQLITEDB:DatabaseFileGone.
• Bridge: sqlitedb and database are bumped to bb39399, the last commit touching each file, with notes added.
  • database.m's change (journal_mode=MEMORY instead of OFF for Validate=false) has nothing to port, since Python never changes journal_mode.
  • The remaining design differences are recorded: BEGIN IMMEDIATE versus a deferred BEGIN, changes() versus rowcount, and where the fields cache is restored. Each has the same effect.

Test changes

• New: a rolled-back add keeps its original, and deletes it once the add commits; nothing is left queued after an operation.
• Updated: a deleted database file is an error, and no new database is created at the path.
• test_open_doc_locations' MATLAB-shaped fixture relied on the old order. Its two files share one original, and the first file's delete_original removed it before the second was copied. Both are now ingested, so test_a_file_with_no_ingested_copy_and_no_original_is_an_error removes filename2's ingested copy itself.

Locally: black --check, ruff check (0.16.5), pytest (802 passed) and check_bridge_coverage.py against DID-matlab main (f4a4905), which reports 0 problems.

---

Generated by Claude Code