[jiro-by-nori]

Fixes #114.

Authored by Jiro, an autonomous agent (Nori sessions) on behalf of @sam-xif.

What was wrong

Range#first was a native builtin with arity 0, registered next to Range#begin
and sitting in zeroArgBids. Two of CRuby's three shapes were therefore wrong,
in opposite directions:

expression	model before	CRuby 4.0.5
(1..5).first(2)	ArgumentError: wrong number of arguments (given 1, expected 0)	[1, 2]
(1...5).first(9)	ArgumentError (given 1, expected 0)	[1, 2, 3, 4]
(1..).first(2)	ArgumentError (given 1, expected 0)	[1, 2]
(1..5).first(1, 2)	ArgumentError (given 2, expected 0)	ArgumentError (given 2, expected 1)
(nil..5).first	nil	RangeError: cannot get the first element of beginless range
(nil..5).first(2)	ArgumentError (given 1, expected 0)	same RangeError

The counted forms are a wrong ArgumentError — inside the family the validator
claims to exclude — and the bare form on a beginless range is a wrong answer,
not a gate. Range#last had the same defect (#110); Array#first(n) and
Array#last(n) already follow CRuby.

The fix

Range#first is now a prelude method in ruby-lean/prelude/prelude.rb, the same
treatment #size/#each/#cover? get and the same shape the #last fix takes:

• the bare form answers #begin, after the beginless check;
• the counted form is Enumerable#first(n): it stops at the end of the range,
  so an over-large n is the whole range and an endless range simply yields n
  elements;
• the beginless check comes first, before the arity and count checks, because
  that is the order in rb_range_first — so a beginless range raises the same
  RangeError for every argument shape;
• the count follows NUM2LONG and rb_ary_new2: a negative count is
  ArgumentError: negative array size (or size too big), and a count beyond a
  machine word either way is RangeError: bignum too big to convert into 'long';
• a non-Integer count, or a range whose endpoints are not Integers, gates
  with a named Unsupported rather than guessing CRuby's per-class each
TypeError (can't iterate from Float, no implicit conversion of String into Integer, …). Those are the only shapes this PR declines.

Range#begin keeps its own arity: (1..5).begin(0) is still
ArgumentError (given 1, expected 0), and the regression case pins it.

Test plan

Ran on Debian with Lean v4.32.2, CRuby 4.0.5 and the pinned Sorbet gems.

• make check — make check: ALL CHECKS PASSED: make lint, gen-check,
  lean, books (every proof book, the checker included), metatheory (axiom
  audit), soundness (SOUNDNESS CHECK PASSED), comparator
  (Your solution is okay!), desugar-unit, emitter-unit, desugar-test,
  desugar-coverage, difftest-test, conformance, feature-loading,
  book-checks (72/72), docs.
• make check-not-weakened — CHECKS NOT WEAKENED.
• make conformance — 1287 ran, 1087 agree / 0 disagree / 194 unsupported,
  exactly the recorded baseline (difftest/coverage-baseline.json is unmoved).
  The regressions and tier-3 corpora agree.
• The new difftest/corpus/regressions/range-first-counted.rb (status fixed,
  28 labelled cases) is reported held by
  uv run difftest run --tier regressions --sut lean, which went from 220 to
  221 held cases with 0 failures.
• bin/ruby-lean --compare difftest/corpus/regressions/range-first-counted.rb
  — AGREE after the change. On origin/main's prelude the same command
  reports DIFFER, and the model's stdout is the wrong-arity
  ArgumentError on every counted case plus beginless: nil.
• Probed beyond the corpus case: 28 further expressions (inclusive/exclusive,
  endless, descending, empty, single-element and negative ranges; n of 0, 5, 6
  and 9; bignum counts) all print byte-identical output to CRuby, and the seven
  gated shapes print the named Unsupported reason (exit 3) instead of an
  answer.

Not in this PR

Range#last's counted form is #110 (its own PR). Two defects found while
probing Range#first are filed separately rather than scope-crept in:
#115 (Range#count is aliased to #size, so a block is silently ignored and the
item argument is rejected) and #116 (Enumerable#index is invented, so
Range#index answers where CRuby raises NoMethodError).