[reginaldalfret]

Summary

In PercentileTimer, negative durations were unconditionally passed to bucket resolution and incremented the bucket count for min (bucket 0 by default), while the underlying Timer dropped them according to the Timer contract. This caused the sum of bucket counts to exceed timer.count() and artificially skewed percentiles toward zero.

Additionally, PercentileTimer did not override Timer.record(long[] amounts, int n, TimeUnit unit), falling back to the default interface loop.

This change brings PercentileTimer in line with PercentileDistributionSummary (see #1303).

Root cause

PercentileTimer.record(long, TimeUnit) called:
java final long nanos = restrict(unit.toNanos(amount)); timer.record(amount, unit); counterFor(PercentileBuckets.indexOf(nanos)).increment();
When amount < 0, timer.record(amount, unit) dropped the value, but restrict(negative) clamped the nanos to min (0ns by default) and incremented bucket 0.

Changes

1. spectator-api/.../PercentileTimer.java:
   • Guarded bucket counter increment with if (amount >= 0).
   • Implemented record(long[] amounts, int n, TimeUnit unit) for batch updates, matching PercentileDistributionSummary.record(long[], int).
2. spectator-api/.../PercentileTimerTest.java:
   • Added negativeAmountIgnored test asserting that negative durations do not increment count() or any bucket counters.
   • Added batchRecordMatchesSingleRecord and batchRecordMatchesSingleRecordDefaultRange tests verifying that batched updates match single updates across zero, positive, and negative amounts.