test: bind moto server to an ephemeral port to fix xdist deadlock

[d-v-b]

This is a claude-authored PR that changes our S3 test fixture to use an ephemeral port managed at runtime by moto itself. I discovered this while commissioning claude to find new pytest features that we can use to improve our test suite.

🤖 AI text below 🤖

Problem

Running the full suite locally with pytest -n auto deadlocks near the end (~98%), with every xdist worker parked at 0% CPU. It has been this way since the shared moto fixture landed (#4018, 2026-06-02).

Root cause: the session-scoped moto_server fixture in tests/conftest.py bound a fixed port (MOTO_SERVER_PORT = 5555). Under pytest-xdist, session scope means once per worker, so when two or more workers reach an S3 test concurrently they race to bind the same port. The first wins; every loser blocks forever in ThreadedMotoServer.start() — its werkzeug thread dies on "Address already in use" before ever setting the ready Event that start() waits on (no timeout).

Minimal repro: construct two ThreadedMotoServer(port=5555) and call start() on both — the second never returns.

CI never sees this because run-coverage runs the suite single-process, and the one -nauto CI target (run-hypothesis) doesn't touch S3.

Fix

• moto_server binds port=0 (ephemeral) and yields the actual endpoint from server.get_host_and_port(). This also makes concurrent pytest sessions on one machine safe, not just concurrent workers.
• Since the endpoint is now only known at runtime, test_fsspec.py's module-level endpoint_url constant becomes an endpoint_url fixture (absorbing the old s3_base alias), and its consumers take it as a parameter. tests/test_docs.py already consumed the fixture value, so it needs no changes.

Verification

• pytest tests/test_store/test_fsspec.py -n 4: 102 passed in 8.8s (previously a guaranteed deadlock shape — 4 workers, 4 servers).
• Full suite pytest --ignore tests/benchmarks -n auto: 6196 passed, 277 skipped, 4 xfailed in 55.54s on a 10-core M-series machine, where it previously hung indefinitely.
• Diagnosis receipts: sampling the hung workers showed all threads in __psynch_cvwait; kill -ABRT produced faulthandler dumps pointing at Event.wait ← ThreadedMotoServer.start ← moto_server fixture setup.

🤖 Generated with Claude Code

[d-v-b]

this is a small test-time-scoped quality-of-life improvement to our test suite that does not affect public code. I will self-merge.