fix(docs): the chunk sweep resolved nothing, and the write-up said otherwise (#110)
Two review findings on #109, both valid.
The chunk sweep reported the best of three interleaved rounds per size and
fitted a two-term cost model to those five points. Re-run with five rounds and
every sample retained, the interquartile range within a single chunk size
reaches 2.55s while the entire spread between sizes is 1.43s - the noise on one
arm is larger than the difference between arms, so the sweep resolves nothing,
and which size wins depends on the estimator (median and minimum pick 1000, the
mean picks 100). The per-entry coefficient survives at ~1.2-1.3 ms because it
dominates; the per-chunk term lands anywhere from ~1 to ~10 ms, which is 0.2s to
2.3s of scaffolding on a 22,000-entry import rather than the 0.23s published.
The conclusion is unchanged and never rested on the ranking - 100 stays - but it
now carries its uncertainty, in the code comment and both pages that repeated
it. The benchmarking page uses the episode to show why a single number without a
spread hides whether a comparison means anything.
Separately, a figure measured on RouterOS 7.24.1 with 22,857 entries had been
written into a table headed RouterOS 7.22.1 with ~28,700 IPs. The historical
table is restored intact and the current measurement gets its own, with its own
conditions.
Both locales; parity gate green.