PerfStar Timing Inconsistencies vs Manual Runs & Unexpected MT Evaluation Behavior (+ OrchardCore on Windows)

While using PerfStar for performance evaluation of MSBuild scenarios, we are observing inconsistencies in reported timings that reduce confidence in the measurements.
Specifically:

1. Mismatch with manual testing
Timings reported by PerfStar differ from manually executed runs of the same scenarios.
The discrepancy is consistent across multiple runs and scenarios.

2. Unexpected MT (multi-threaded) evaluation performance
Evaluation time in MT mode is higher than in non-MT mode.
This is counterintuitive, as MT is expected to improve or at least not regress performance.
The issue is particularly visible in evaluation phase metrics (as opposed to build phase only).

## Expected Behavior

PerfStar timings should reasonably align with manual measurements (within expected variance).
MT mode should not be consistently slower than non-MT for evaluation.

## Observed Behavior

PerfStar produces timings that diverge from manual runs.
MT evaluation is consistently slower than non-MT in some scenarios.

## Questions

Are PerfStar metrics collected differently than in manual runs?
Are MT scenarios configured identically between PerfStar and manual testing?