Run gc.collect() at session teardown to promptly release session-scoped resources #2483

Summary

When a Shiny session ends, most session-scoped objects are released promptly via refcounting, but a fraction become entangled in reference cycles involving the session machinery (AppSession, reactive domain, callback registries, the ExitStack in _run_impl). Those objects survive until the cyclic GC runs — and because CPython's cyclic collector is triggered by allocation counts rather than wall-clock time, a long-running but idle process may not collect them for hours. For objects holding external resources (HTTP connections, file handles), this means resources outlive their sessions by an unbounded, unpredictable amount of time.

Proposing that Shiny call gc.collect() at the end of session teardown (_run_session_ended_tasks), since session end is a natural, infrequent collection point.

Reproduction

I ran a real Shiny 1.7.0 app whose server function creates an httpx.Client holding a keep-alive connection (the same pattern as creating an LLM client like chatlas.ChatOpenAI() at session start), then drove 120 real websocket sessions through it with the proper init handshake, counting the app process's open TCP connections to the endpoint with lsof:

# app.py
import httpx
from shiny import App, ui

app_ui = ui.page_fluid("fd test")

def server(input, output, session):
    client = httpx.Client()
    client.get("http://127.0.0.1:9876/")  # local HTTP/1.1 keep-alive server

app = App(app_ui, server)

Results:

Two things stand out:

1. ~10% of sessions leave a lingering connection until the cyclic GC runs. All lingering sockets were ESTABLISHED.
2. An idle process never collects. The count was flat across 60 idle seconds; it only dropped once new sessions created enough allocations to trigger the collector.

Why this matters

This affects any app that creates resource-holding objects per session — HTTP clients (httpx/openai/anthropic SDK clients), database connections, temp files. In a long-running process with bursty traffic (e.g., a Shiny app that's quiet overnight), dead sessions' connections can linger for many hours, holding file descriptors and server-side keep-alive slots. It's not an unbounded leak — collection does eventually happen — but cleanup timing is unpredictable and decoupled from session lifetime.

Explicit close() at session end is the right user-level fix where available, but Shiny can't rely on every object in every server function being explicitly closed.

Proposal

Call gc.collect() at the end of _run_session_ended_tasks() in shiny/session/_session.py. Session ends are infrequent relative to allocations, so the cost should be negligible, and it makes cleanup prompt for all session-scoped cyclic garbage, not just the objects that happen to get explicit teardown.

A lighter-weight alternative would be gc.collect(0) (youngest generation only), though that may not catch cycles involving longer-lived session objects.

Happy to put together a PR if this sounds reasonable.