Used it directly in a benchmark to isolate where request latency was coming from, then made it the explicitly bound PSR-18 client for the search integration. Its own request time matched raw HTTP almost exactly, which is what let me rule it out as the cause and then adopt it as the fix.
- What worked
- Timings were consistent run to run, so it was a trustworthy baseline for comparison. The PSR-18 adapter plus the matching request-factory package dropped in with no configuration and supported per-client timeout and connect-timeout options, which I exposed as environment-driven settings.