Fix flaky DatadogMetricsTest (broken wait helper, origin detection, UDP races)

[adamw]

Problem

DatadogMetricsTest was flaky on CI (e.g. the failure on #5369, where the 4xx counter assertion ran before the datagram arrived) and failed 5/5 when run on a dev machine. Four root causes:

1. The waitReceiveMessage helper was broken. Its m.startsWith(expected) check could never match — the default sentinel was "method" and others were "4xx"/"0.3", while every datagram starts with tapir. or datadog.. Its polarity was also inverted (it kept looping while the expected message was present, rather than until it appeared), and its first clause was unbounded (a potential infinite hang if no message ever arrived). The effective behaviour degraded to "return as soon as ANY message arrives", so the assertions that followed raced in-flight UDP datagrams. The "0.3" sentinel in the duration test was additionally wrong — the last histogram sent is 3.3.
2. Origin detection breaks exact-match assertions on dev machines. java-dogstatsd-client enables origin detection by default; on machines where a container id resolves, it appends |c:<container-id> to every datagram, so all exact-string assertions fail (hence 5/5 local failures). Now disabled via .originDetectionEnabled(false) on every client.
3. Client-side aggregation delays and reshapes counters. The client aggregates metrics client-side and flushes every ~2 seconds, so datagrams arrive late, and counter increments can be split or merged across flush windows.
4. Unsynchronized cross-thread access. MockStatsDServer.receivedMessages was a plain var written by a daemon receiver thread and read by the test thread, with no @volatile.

Fix

• Replaced all waitReceiveMessage calls with eventually-wrapped assertions, using the same idiom as ServerMetricsTest (class-level Eventually.PatienceConfig, 15s timeout / 150ms interval for slow CI); deleted the broken helper.
• Added .originDetectionEnabled(false) to all five client builders.
• Tagged "collect requests active" and "collect requests total" as Retryable: aggregation can in rare cases split or merge counter increments across 2-second flush windows, which no amount of waiting can heal — the suite's existing retry mechanism handles this.
• Made receivedMessages @volatile.

Verification

On a machine where origin detection resolves a container id:

• Baseline (before fix): datadogMetrics/testOnly ...DatadogMetricsTest — 0/5 passed (all 5 fail on the appended |c:<container-id>).
• After fix: 3 consecutive runs — 5/5 passed, 3/3 green.
• datadogMetrics2_12/Test/compile and datadogMetrics3/Test/compile — success.
• datadogMetrics/Test/scalafmtCheck — success.

🤖 Generated with Claude Code