[ondrejmirtes]

A benchmark compared by its growth had nothing for tally accept to write: a slower case meant editing the attribute by hand. Accept now fits the measured instructions of a method's data sets against the presets and writes the lowest one they grow within, as tally run compares them.

slower  App\PairsBench::benchPairs: O(n) to O(n^2)

Accepted the complexity class of 1 benchmark.

• Up when the instructions grew more than the class allows, such as #[Constant] to #[Logarithmic], down when a lower preset bounds them too, such as #[Linear] to #[Logarithmic] (listed as faster).
• Order: Constant, Logarithmic, Linear, Linearithmic, Quadratic, Polynomial 3, 4, …, Exponential. The size argument stays as written; the import of the replaced class goes when nothing else uses it.
• The replacement is a set of CodeEdits in the file's AttributeIndex, like the expectations since 5ed6578, so a file with both is rewritten in one pass. The new import takes the place of the removed one when it sorts there.
• Classes are compared at the measured sizes only, so #[Exponential] comes before a preset that allows more growth at the largest data set. Otherwise any growth would fit some high-degree polynomial.
• A class that grows exactly like the current one keeps it (#[Polynomial(degree: 1)] stays). #[Complexity] with its own formula is left alone.
• A method no preset bounds is reported: Cannot accept App\FooBench::benchFoo: no preset complexity class bounds how its instructions grow, write a #[Complexity] that does. The other changes are still written, and accept exits with 2.
• The report now suggests accepting a slower growth case instead of loosening the class, and shows the accept command for it.

PresetFitTest covers each neighbouring pair of presets, up and down. CompareAndAcceptCommandTest runs accept end to end on a fixture.

Checks: lint, cs, PHPStan, tests, 100 % coverage, and Infection on the changed files at 100 % MSI pass locally. make bench was not run; ExpectationChanges::plan() changed, so the Benchmarks job may report new numbers.

Known limitation: accept with --filter fits the class to the selected data sets only.