feat(benchmarks): auto-generate Interpretation/Recommendation and add -report-only flag

generate_report.go:
- buildInterpretation: derives narrative from p99/p50 tail-latency ratio,
  per-action complexity trend (% increase vs discover baseline), concurrency
  scaling efficiency (GOMAXPROCS=1 vs 16), and cache warm/cold delta
- buildRecommendation: identifies the best throughput/cost GOMAXPROCS level
  from scaling efficiency and adds production sizing guidance

run_benchmarks.sh:
- Add -report-only <dir> flag: re-runs parse_results.go + generate_report.go
  against an existing results directory without rerunning benchmarks

REPORT_TEMPLATE.md:
- Replace manual placeholders with __INTERPRETATION__ and __RECOMMENDATION__
  markers filled by the generator

benchmarks/reports/REPORT_TEMPLATE.md

@@ -36,11 +36,11 @@ adapter-internal latency from network variables.
 
 ### Interpretation
 
-_Review the numbers above and add interpretation here._
+__INTERPRETATION__
 
 ### Recommendation
 
-_Add sizing and tuning recommendations here._
+__RECOMMENDATION__
 
 ---