[bjester]

Summary

• Adds characterization tests that pin what a full network push and pull send and store, for the sync layer refactors in M2–M4. They cover both the async flow (ASYNC_OPERATIONS) and the legacy flow used with servers before 0.6.0
• A spy on SessionWrapper.request records each request and passes the call through to the real method. Per request, it records the method, path, query params, body encoding (json/gzip), body keys, the values of the stable body fields (filter, push, records_total, records_transferred, transfer_stage), the record count of each pushed chunk, and is_retryable. Generated IDs are normalized to <id>
• Asserts the final local transfer session fields, the Buffer, Store and app rows on both sides, and the byte counts the server stores
• Fails a sync that sends more than 100 requests or runs longer than 120s, so a stalled sync fails instead of hanging the test run
• For the legacy tests, the live server runs with MORANGO_TEST_LEGACY_SERVER, which turns off ASYNC_OPERATIONS and FSIC_V2_FORMAT in server2_settings.py. LiveServer and BasePushPullClientTestCase take extra server environment variables for that
• Extracts the live server harness from test_syncsession.py into integration/helpers.py as BasePushPullClientTestCase, so the new tests reuse it without re-running PushPullClientTestCase's tests

Reviewer guidance

• The recording goes beyond the plan's (method, path, body keys, transfer_stage) tuple. Mutation checks showed the narrower tuple let these through, and the expanded recording catches each: dropping is_retryable, sending buffer chunks as JSON instead of gzip, un-flipping the byte counts in the pull PATCH, and adding a query param to the buffer GET
• The legacy tests emulate an older server, and the client keeps its full capabilities, as a current Kolibri does when it syncs with an older one. Tests that pin today's legacy network sync behavior #364 proposed reducing the client capabilities instead, but that fails today on the server for both push and pull, which is now tracked in Clients that disallow async operations cannot sync with servers that allow them #416
• The legacy pull moves 11 records, not 10. The client creates the transfer session before it serializes, so it sends an empty client_fsic, and the legacy server queues the user from the initial pull again. This is today's behavior, pinned as-is
• URL paths are literal strings on purpose, not reverse() or the api_urls constants. The client sends them to peers on other morango versions, so a renamed route must fail these tests
• chunk_size is set in the test class rather than inherited, because the expected request lists depend on it and the harness is not frozen
• Once this merges, test_network_characterization.py must not change until M4 is complete

Issues addressed

Closes #363
Closes #364

Related: #416

AI usage

Claude Code implemented the tests from the issues and the implementation plan. I had a separate Claude subagent review the async tests against the issue and spec. Its mutation checks showed regressions that the plan's narrower request tuple missed, so I had the recording strengthened, and directed extracting the shared harness into integration/helpers.py. While implementing the legacy tests, Claude found that the approach in #364 fails against the current code, so we switched to emulating a legacy server and filed #416. Claude ran the integration suite and the mutation checks to verify the changes.

🤖 Generated with Claude Code