frame-benchmarking: allow 2-point slope fits in min_squares_iqr

[tindzk]

Description

a33b7c2 short-circuits min_squares_iqr to median_value whenever r.len() <= 2 to avoid OLS panics on --steps=1 --repeats=1. That fallback also catches the case where two samples sit at different x-values, where the slope is exactly determined. However, benchmarks whose valid sample set is narrowed to two values (e.g. by BenchmarkError::Skip filtering a Linear<lo, hi> parameter) lose their linear component.

Route the r.len() <= 2 fallback through median_slopes instead, which fits a slope exactly from two samples at different x-values. The degenerate "all samples share one x" case now surfaces an explicit error from median_slopes's own check.

Integration

No API or storage changes. Downstream runtimes do not need to update code.

Regenerating weights may produce different values for benchmarks whose valid sample set is exactly 2 points at distinct x-values. The resulting weight expression will now include a linear component fitted via median_slopes instead of falling back to the constant median.

Benchmarks whose 2 remaining samples share an x-value will now fail with an explicit median_slopes error rather than returning the median. If you see this, broaden the parameter range or increase --steps/--repeats so more than one distinct x-value is sampled.

Review Notes

min_squares_iqr had a single guard combining two unrelated short-circuits:

if r[0].components.is_empty() || r.len() <= 2 {
    return Self::median_value(r, selector);
}

The components.is_empty() branch is correct: a benchmark with no parameters has no slope to fit, so a constant median should be returned. The r.len() <= 2 branch was a workaround for linregress's OLS fit, which is under-determined with two samples and one intercept + one slope variable. Both cases fell through to the same median_value call, which drops slopes entirely.

The no-components fallback is kept on median_value, while r.len() <= 2 is routed to median_slopes instead. median_slopes is well-suited: it forms the pairwise slope list (y_i - y_j) / (x_i - x_j) over samples with distinct x-values, takes the median, then derives the intercept from per-sample offsets. With exactly two distinct-x samples the slope list has length 1, so the median is the exact slope. No regression is needed.

[tindzk]

/cmd prdoc

[paritytech-release-backport-bot]

Successfully created backport PR for unstable2604:

• [unstable2604] Backport #12075 #12141