test(stress): debug build for the soak server + per-interval throughput

[cryo2010]

What

Two changes to the per-workload stress soaks (nimble stress*), aimed at making the soaks better at their actual job: forcing and diagnosing intermittent crashes.

1. Build the soak server -d:release, not -d:danger

The soaks exist to force intermittent crashes and instability, but -d:danger disables all runtime checks (bounds/overflow/nil) plus assertions. That works against the goal: without checks, a defect doesn't fail at its site, it silently corrupts memory and crashes somewhere unrelated (the hardest kind to trace). It also makes --panics:on a no-op, since the Defects panics govern are never raised in the first place.

-d:release keeps optimizations on (realistic timing, so races/lifecycle bugs still reproduce) but keeps the checks on, so a defect fails fast at its source. Added --stackTrace:on --stackTraceMsgs:on --lineTrace:on --panics:on --debugger:native for a native traceback on failure.

Tradeoff: with checks on, the per-cell RSS/throughput are a debug-build baseline, not the peak ceiling. That's the correct tradeoff for a debugging build; the Dockerfile comment is updated to say so.

2. Per-interval throughput on every report line

Non-streaming workloads (requests, methods, ws, sse) previously reported only cumulative 2xx tallies, so a throughput dip or declining trend was invisible unless you diffed successive lines by eye. Added a per-interval 2xx completion rate (requests/s / messages/s / events/s per the workload), computed the same way fmt_xfer already does MB/s for the streaming workloads. 0 in the segment means nothing completed in the interval, i.e. a stall.

[requests h2 sync] 200x1481767 | 33018 requests/s | RSS 29MB | heap 7MB | fds 42 | t=45s
[requests h2 sync] final 200x1493782 | 802 requests/s | RSS 29MB | heap 6MB | fds 41 | t=60s

Notes

• The rate's resolution is VORTEX_REPORT_SECONDS (default 60s). For dip-hunting, run with a smaller value (e.g. VORTEX_REPORT_SECONDS=5) for a finer time series.
• The rate is interval throughput (delta/dt), not a running average, so a dip shows as a low number on that one line rather than being smoothed away.

Test

• python3 -c "import ast; ast.parse(...)" on the client passes.
• The changes are local-only stress tooling (Docker soaks), not a CI gate.