AI junk

[ghost]

In wsgi_app, the call to response(environ, start_response) sits outside the inner try/except that catches dispatch errors. When this call raises (for example, a streaming response body fails or the client disconnects mid-delivery), the error variable remains None. The finally block then calls ctx.pop(error) with error=None, so all teardown_request functions receive exc=None instead of the actual exception. This silently discards the error, making it impossible for teardown logic to make informed cleanup decisions (such as rolling back a database transaction versus committing).

The fix wraps response(environ, start_response) in a try/except BaseException that captures the error into the error variable before re-raising. The finally block then passes the correct exception to ctx.pop(error). Using BaseException ensures that KeyboardInterrupt, SystemExit, and other non-Exception throwables during WSGI response delivery are also captured. The exception is always re-raised after capture — it is never swallowed or handled, preserving the existing contract that wsgi_app propagates unexpected errors to the WSGI server.

A regression test is added that creates a Response subclass whose __call__ raises RuntimeError, registers a teardown_request handler, and asserts that the handler receives the RuntimeError instance. This test fails on the unpatched code (teardown receives None) and passes with the fix. The full Flask test suite (492 tests) passes with no regressions.