Hanzi Jiang 8412179b95 Isolate the metrics sweep failure and pin the test sweep period ...

- Wrap the registry sweep and metric emission in try/catch: an
  allocation failure under memory pressure now costs only the two UDF
  metrics for that cycle instead of discarding the whole asynchronous
  metrics snapshot. Memory pressure is exactly when these metrics are
  read, so blanking every metric for the cycle was the worst failure
  mode.
- Pin `asynchronous_metrics_update_period_s` to 1 in the test config:
  the suite's polling windows are calibrated to a 1 second sweep and
  previously relied on the server default.
- Fix the `sample` comment: on non-Linux the metrics are absent, not 0
  (the call site gates emission).

2026-06-12 17:47:56 -04:00

..

udf.xml

…