A blind scan of the pre-fix token_scoping.py flagged the #1885 connection leak

[Tomosu-AI]

We used the #1885 leak as a test case for our reliability scanner (disclosure, we build Tomosu). We checked out cf5adcd from before the #1886 fix and gave the scanner only mcpgateway/middleware/token_scoping.py. We didn’t give it the issue, the load test or the patch. Before running it, we wrote down the pass condition: “Identify unsafe database-session cleanup where an operation before close() can fail.”

It flagged “Unsafe database-session lifecycle” at High, noting that an operation before close can fail and leave the connection unreleased. That’s the same failure the 3,000-user load test turned up.

One place we’d like your view. #1886 nests the commit in its own try so close() always runs. Our scan suggested dropping the commit from that cleanup completely, since the path looked read-only to us. We didn’t load test that, so it may well be wrong. Does that path ever need the commit?

Write-up with the method: tomosu.ai/case-studies/contextforge-connection-leak.html