Regression workflow

Baseline and candidate

1. Freeze scenario, device profile, build settings, camera/input script, warm/cold state, and duration.
2. Run multiple baseline samples; retain raw traces and summary metrics.
3. Apply one change.
4. Run the same number of candidate samples in comparable thermal/background conditions.
5. Compare milliseconds, memory, bandwidth, hitch count, and percentile relevant to the budget.
6. Inspect a representative trace when the aggregate moved unexpectedly.

Minimum report:

Metric	Baseline	Candidate	Delta	Budget	Result
median frame ms
p95/p99 or worst-frame ms
bound thread/pass ms
hitch count over threshold
memory/load/net metric

Do not call a change a win when it shifts cost outside the sampled interval, trades a rare severe hitch for a better average, changes image quality/gameplay, or exceeds another budget. Keep a safety margin; meeting 16.67 ms only in an ideal trace is not a robust 60 fps target.