fix: emit per-client metrics in benchmark summary

[cbermudez97]

Problem

When a benchmark runs multiple clients, the end-of-run per-client metrics were not truly per-client on the default (non-Prometheus) path — the per-client/per-method entries came out empty or reflected data that wasn't actually separated by client.

Each client maps to a k6 scenario (scenario name == client name). The default metrics collector reads k6's summary.json (--summary-export) and expects per-client-per-method submetrics named like http_req_duration{scenario:C,req_name:M}. But k6 only writes a tag-filtered submetric into summary.json when a threshold is defined on it, and the generator only registered thresholds keyed by req_name alone (and only when a call defined thresholds). So no per-scenario (client) submetric was ever emitted, and the parser had nothing per-client to read. The base http_req_duration metric in the summary is aggregated across all clients.

The Prometheus remote-write path was unaffected — every exported series keeps its scenario label.

Fix

Register always-true per-client × per-method submetric thresholds in the generated k6 config so k6 emits the per-scenario breakdown into summary.json:

• http_req_duration{scenario:C,req_name:M} → max>=0
• http_reqs{scenario:C,req_name:M} → count>=0
• http_req_failed{scenario:C,req_name:M} → rate>=0

The conditions can never fail, so the k6 exit code is unaffected. Keys are unquoted and use {scenario:C,req_name:M} ordering to match the name lookup in summary_fallback.go — k6 stores submetric names verbatim (it does not strip quotes or re-sort tags). Existing user-configured thresholds are left untouched (still global across clients). No parser or downstream changes were needed.

Verification

Ran a real k6 benchmark against two mock RPC servers with 20 ms / 60 ms injected latency:

• go build ./... passes.
• summary.json now contains all per-client submetric keys in the exact format the parser matches.
• P99 coverage went from 0% → 100% (4/4 methods).
• Final exports are genuinely distinct per client: geth ~21 ms avg, nethermind ~61 ms avg — matching the injected delays — with per-method counts, latencies, and P99 all populated.
• k6 exit code stayed 0 (no false threshold failures); Prometheus path unaffected.