[ci] run sampler_test in-process (drop from SUBPROC_TEST_PATTERN)

[tiankongdeguiji]

What

Move sampler_test to the in-process test group (drop .sampler_test. from SUBPROC_TEST_PATTERN). All other entries stay.

Why

SUBPROC_TEST_PATTERN runs each matching test in a fresh python -m subprocess. For the graphlearn-sampler tests this was added (#76) to dodge graphlearn's module-global SERVER_LAUNCHED guard, which raises duplicate server launch detected when launch_server runs twice in one process.

Since #554 tzrec owns the server launch (inlined, without touching graphlearn's SERVER_LAUNCHED), so that guard no longer fires. But removal has to be decided per entry — what matters is where the server is launched, since that's where the new liveness watchdog's os._exit runs:

entry	isolation reason	still needed?
sampler_test	dup-launch guard — but launches the server only in forked _sampler_worker children	No → in-process
dataset_test	launches the server in the main process → watchdog could os._exit the whole runner	Yes
odps_dataset_test	main-process launch + distributed RANK/WORLD_SIZE + init_process_group; skipped w/o ODPS creds	Yes
parquet_dataset_test	distributed RANK/WORLD_SIZE/MASTER_PORT + forked ranks	Yes
tdm.gen_tree	mutates USE_HASH_NODE_ID + forks mp.Process clustering workers	Yes
convert_easyrec_config_to_tzrec_config_test	needs PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python set before protobuf import (_run_env)	Yes

Only sampler_test is safe to de-isolate: its watchdog runs in the worker child (can't kill the runner) and its USE_HASH_NODE_ID mutation is reset in setUp.

Verification (local, GPU box with graphlearn)

• _gather_test_cases now puts sampler_test's 22 cases in the in-process group and leaves dataset/odps/parquet/gen_tree/convert in the subprocess group.
• sampler_test's 22 tests pass in-process — including the watchdog test, whose os._exit(1) kills only its worker child (runner survives), and a hash+int run together.
• A/B confirming the removed root cause: graphlearn's launch_server twice in one process raises duplicate server launch detected; tzrec's inlined launch does not.
• Even with the watchdog poll forced to 0.3 s, the main-process dataset_test runner survives the recurring graphlearn server-child crash (the sampler is GC'd → stop.set() before the crash) — confirming the watchdog hazard is real only for main-process launchers and is confined by their subprocess isolation.

🤖 Generated with Claude Code

https://claude.ai/code/session_012uPyFcxcLAdNpgFFCYSZ2f