[jiro-by-nori]

Summary

Enumerable#minmax accepted a comparison block and never called it: it answered
[min, max] under the default <=>. This was a silent wrong answer for a valid
program, not a gate:

[1, 2, 3].minmax { |a, b| b <=> a }                    # before: [1, 3]  CRuby: [3, 1]
%w[xxx x xxxx xx].minmax { |a, b| a.size <=> b.size }  # before: ["x", "xx"]  CRuby: ["x", "xxxx"]
{foo: 0, bar: 1, baz: 2}.minmax { |a, b| a[1] <=> b[1] } # before: [[:bar, 1], [:foo, 0]]  CRuby: [[:foo, 0], [:baz, 2]]

The block is CRuby's comparison over consecutive elements, not a key
function. enum_minmax compares two consecutive elements once, then the smaller
with the running minimum and the larger with the running maximum, and compares a
trailing odd element last, so a program that logs its comparisons observes that
order. The block form now runs that loop; without a block the pair is the same
under any total order and Array#minmax answers [min, max], so that path is
unchanged. minmax_by still gates as unmodeled.

Fixes #128.

Test plan

The new regression program difftest/corpus/regressions/enumerable-minmax-block.rb
(hand-filed, status: fixed) is replayed by make conformance on every build. It
pins both the answer and the exact sequence of comparisons for:

• descending comparators, comparisons by String#size, and a Hash's pairs;
• even and odd element counts, the empty and one-element collections, equal
  pairs, and a trailing unpaired element;
• an Enumerator receiver, and a comparison whose result depends on how often it
  has run (only a loop with CRuby's call order answers that the same way);
• the no-block form, unchanged.

Before the change the same program disagreed on the first line (the block's log
was empty); after it bin/ruby-lean --compare reports
AGREE the model and CRuby print the same and end the same way.

Validation

All run locally on CRuby 4.0.5 with the pinned toolchain:

• make conformance — 1287 bootstraptest programs: 1087 agree, 194 unsupported,
  0 disagree, baseline held; the regressions tier and tier3 agree.
• make books — 1399 jobs, success.
• make soundness — SOUNDNESS CHECK PASSED.
• make metatheory — axioms are propext, Classical.choice, Quot.sound only.
• make difftest-test — 189 passed, coverage baseline held.
• make lint — ruff, RuboCop, shellcheck, actionlint clean.
• make gen-check — all generated sources fresh.
• make check-not-weakened — CHECKS NOT WEAKENED.

Authored by Jiro, an autonomous agent. The analysis, the code, and this
description are the agent's.