Print each PyHealth log line once in Jupyter and Colab

[solarsys]

Why

pyhealth/__init__.py gives the pyhealth logger its own stdout handler, and the logger also propagates to the root logger. Jupyter, Colab and logging.basicConfig() all give the root logger a handler, so in those settings every PyHealth message prints twice: dataset loading, task processing, the Trainer's epoch summaries. It makes notebooks noticeably harder to read (it showed up throughout the Colab runs for #1281).

What

• Set propagate = False on the pyhealth logger, next to where PyHealth adds its handler, with a comment showing how to opt back in.
• Why this fix and not "only add the handler when root has none": that check would run at import time, so a script that imports PyHealth and then calls logging.basicConfig would still print twice. Turning off propagation is deterministic and matches how most libraries that print by default behave.
• Trade-off: an application that configures the root logger (say, to a file or a log collector) no longer receives PyHealth's messages there automatically. It can get them back with two lines, now documented:

  logging.getLogger("pyhealth").removeHandler(pyhealth.handler)
  logging.getLogger("pyhealth").propagate = True

• Unchanged: the INFO level, the stdout handler, child loggers (pyhealth.trainer, pyhealth.datasets, ...) still reach PyHealth's handler, and Trainer's log.txt file logging is unaffected.
• Docs: a "Controlling PyHealth's log output" entry in docs/faq.rst. Example: examples/logging_setup.py (no data needed) shows the default, a quieter level, and routing through root.

Once this is merged, the logging.getLogger("pyhealth").propagate = False line in the tutorials' install cell (#1281) can be dropped.

Tests

tests/core/test_logging_setup.py (synthetic, no data):

• with a root handler attached, a message from a pyhealth.* logger is printed once by PyHealth's handler and not by the root handler (fails on master, passes here);
• INFO is still printed and DEBUG is not;
• removing the handler and re-enabling propagation routes messages through root;
• Trainer's set_logger still writes log.txt, and the message prints once.

Also ran the modules that use assertLogs on PyHealth loggers and the Trainer-based tests (test_early_stopping, test_model_mode, test_xgboost_model, test_nested_processor_truncation, test_default_cache_dir): 45 tests OK. tools/check_pr_rules.py passes.

🤖 Generated with Claude Code

[jhnwu3]

lgtm