fix(o11y): record output request + HEC ACK-poll latency in ms (PIPE-1404)

[Dylan-M]

Proposed Change

blitz.output.request_latency and blitz.output.hec.ack_poll_latency were declared in seconds and recorded via .Seconds(). Both use OTel's default histogram buckets (0, 5, 10, 25, ... 10000), where the first boundary is 5. Output latencies almost always run well under 5 seconds, so every sample landed in the first bucket. The histograms carried no distribution signal.

Both now record in milliseconds, where the default buckets fit. A new output.DurationMillis helper converts a duration to fractional milliseconds. It stays float so a sub-millisecond sample is not truncated to zero. Truncating would also understate the histogram sum. The request-latency feeders (tcp, otlp_grpc, hec) and the HEC ACK-poll site record through it. Both metric.yaml units change from s to ms.

Stacked on #291.

How the reviewer can validate

Note: the manual validation procedure below is an AI-generated script. Review it before running.

Manual:

1. make build
2. Start a throwaway TCP sink: nc -lk 127.0.0.1 5140 >/dev/null &
3. Run blitz to it with the self-metrics endpoint on:

   ./blitz --generator-type apache-common --generator-apache-common-rate 50ms \
     --output-type tcp --output-tcp-host 127.0.0.1 --output-tcp-port 5140 \
     --metrics-port 9100
   

4. curl -s localhost:9100/metrics | grep request_latency
5. Confirm the histogram renders in milliseconds: blitz_blitz_output_request_latency_milliseconds_* (previously ..._seconds_*). On a localhost sink, sub-5ms sends correctly sit in the first bucket. Any latency above 5ms now lands in a higher bucket, where in seconds it would have collapsed into the first.

The HEC ACK-poll histogram (blitz_blitz_output_hec_ack_poll_latency_milliseconds_*) validates the same way against a HEC endpoint with ack enabled.

Automated: go test ./output/.... TestDurationMillis covers the conversion including sub-millisecond preservation. Manual-reader tests assert the ms unit on both blitz.output.request_latency and blitz.output.hec.ack_poll_latency.

Checklist

• Changes are tested
• CI has passed

[Dylan-M]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• fix(o11y): record output request + HEC ACK-poll latency in ms (PIPE-1404) #292 👈 (View in Graphite)
• feat(o11y): lifecycle startup-latency metrics via new runtime weaver registry (PIPE-1066) #291
• feat(o11y): bring stdout/nop outputs fully into the self-telemetry spine (PIPE-1066) #290
• feat(o11y): per-batch output send spans + HEC ACK-poll span (PIPE-1066) #289
• feat(o11y): wire fix generator into the self-telemetry spine (PIPE-1066) #288
• feat(o11y): self-logs via shared entry-point zap->OTel bridge + OTLP log export (PIPE-1066) #287
• feat(o11y): self-tracing via runtime spans, adapter emit spans, and OTLP trace export (PIPE-1066) #286
• feat(o11y): per-instance metrics via injectable MeterProvider, drop package globals (PIPE-1066) #285
• feat(embed): add TelemetrySettings provider bundle + NopTelemetry (PIPE-1066) #284
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.