Improve benchmark reliability

[RowDaBoat]

Finding

The previous comparison labeled any median change above a fixed 5% as significant, while its CSV retained only one median per benchmark. A same-source CI comparison reproduced false 5.1-5.5% regressions, and a local same-binary run produced false 13-18% improvements in churn rows.

The first revision preserved inner samples, but an optimization trial exposed a second issue: samples from one process are correlated and cannot stand in for independent runs. Two process medians moved by several percent while the pooled samples still produced directional verdicts.

Changes

• Batch fast read and iteration workloads to roughly 1 ms per timing sample while reporting normalized operation time.
• Preserve raw timing and memory samples for diagnostics and preserve one median per independent process run.
• Compare 15 matched process pairs, alternating candidate/baseline execution order to balance drift.
• Base verdicts on conservative 99.9% paired-run median intervals plus a 1% minimum directional effect.
• Report uncertain results as inconclusive and retain compatibility with the existing four-file comparison command.
• Add focused regression tests and a standalone comparison command.

Validation

• Complete tests suite passes, including the benchmark regression tests.
• tests/manycomponents.nim passes with -d:ArchetypeWords=2.
• A 15-pair, 30-process identical-binary comparison classified every benchmark as unchanged.
• The previously suspicious two-process data is now correctly inconclusive.
• Public vecs source and API are unchanged.

This branch is the required baseline for subsequent optimization work.