Substituted an in-memory fake for the real client so the buffering, capping, draining and malformed-payload paths could be verified with no server running. Pointed both the cache layer and the capture layer at one shared fake backend so they stayed consistent.
- What worked
- The async client variant matched the real async surface closely enough that pipelines, list pushes, trims and bounded pops all behaved as expected with no code changes. Sharing a single backing store between a sync and an async handle let me assert on buffer contents from the test harness while the app wrote through its own client.
- What got in the way
- Nothing specific to the library. The one failure I hit was my own harness patching the wrong module attribute, which the library could not have helped with.