[lesnik512]

Same change as modern-python/fastapi-sqlalchemy-template#105. Reverts the backpressure half of #77.

Granian's backpressure limits open connections per worker, not requests in flight. Each worker takes a semaphore permit before accept() and releases it when the TCP connection closes (granian v2.8.3, src/workers.rs lines 704 and 1074). An idle keep-alive connection holds its permit, so with backpressure set to the pool size (5 by default) a reverse proxy that keeps upstream connections open can take every slot by itself. New connections then wait in the backlog until Granian drops an idle one at --http1-header-read-timeout, 30 s by default, even when the app is doing nothing.

Reproduced locally with backpressure=2, two idle HTTP/1.1 keep-alive clients, and then a third request:

Granian settings	third request served after
--backpressure 2	30.0 s
--backpressure 2 --http1-header-read-timeout 5000	5.0 s
default backpressure	0.0 s

Changes

• app/__main__.py no longer passes backpressure, so Granian uses its default (backlog / workers).
• test_main asserts that backpressure is not passed.
• AGENTS.md: the pool sizing paragraph says why backpressure stays at its default.

Overload is still bounded. A request that finds the pool empty waits for a connection, and after DB_POOL_TIMEOUT it gets a 503 from database_unavailable_handler. The 30 s default for that timeout is unchanged here. Lowering it would shed load sooner, and that is a separate decision.

Testing

just test: 87 passed, 100% coverage. ruff format --check and ruff check are clean. ty check runs in CI, because the dev environment does not install on macOS (pyroscope-io fails to build).