[michaelthyregod]

Compare-Builds.ps1 misread the BenchmarkDotNet CSV reports, so summary.md could be wrong.

Problem

Write-Summary stripped everything but digits and dots from the Mean and Allocated cells, then read them as ns and B. BenchmarkDotNet writes the unit into each cell and picks it per row (952.654 μs, 1,280.41 KB). So:

• such rows were off by 1000x or 1024x;
• a ratio was wrong whenever two builds of the same case landed in different units.

Change

• Each cell is parsed as a value and a unit, then converted. Times: ns, μs/µs/us, ms, s. Sizes: B, KB, MB, GB.
  • NA (a failed benchmark) drops the row.
  • - counts as no allocation.
  • An unknown unit throws instead of producing a wrong number.
• Reports are read only from this run's round-* folders. The baseline source trees live in the output directory too, and a report left in one of them would otherwise count as a build.

Verification

No benchmark run was needed: the script's functions were tested directly.

• Cells: 12 samples convert correctly, across every unit, with thousands separators and -; an unknown unit is rejected.
• Real data: on the reports of an earlier three-way run, the regenerated summary.md is identical to the original, even with a stray report in a source tree.
• Mixed units: with one build's cells rewritten into μs/ms and KB/MB, the summary matches up to the rounding of the rewritten values. The previous parsing got 20 of 24 rows wrong; for example, 5.53 ms showed as "5.5 ns" with ratio 0.000.

🤖 Generated with Claude Code