[jiro-by-nori]

Fixes #144.

What was wrong

Range#each stepped with succ and compared with </<=. CRuby's
range_each does neither, so three separate answers came out wrong:

program	before	CRuby 4.0.5
(1.0..2.0).each { |x| p x }	prints 1.0, then NoMethodError: undefined method 'succ' for an instance of Float	TypeError: can't iterate from Float, nothing printed
(1.0..2.0).map { |x| x }	NoMethodError (succ on Float)	TypeError: can't iterate from Float
(obj..obj).each { }, obj defining succ and <=> but no <	NoMethodError: undefined method '<=' for an instance of ...	iterates

NoMethodError is inside the family the validator claims to exclude, so these
are wrong answers rather than gates (the line #27 settled) — and the first one
yields 1.0 before failing, so it has already produced output.

What changed

ruby-lean/prelude/prelude.rb, Range#each:

• Refuse a begin CRuby cannot step, before yielding. CRuby asks
  rb_respond_to(beg, succ) and raises TypeError: can't iterate from <Class>
  otherwise. respond_to? in the prelude is CRuby's name table, so a begin
  whose succ CRuby has but the model does not model (String, Symbol)
  keeps gating on the unmodeled succ instead of turning into a new wrong
  TypeError. The beginless/endless gate is unchanged: (1..).each really is
  infinite, and a faithful infinite loop is indistinguishable from a stuck
  model.
• Compare with <=> for a non-Integer begin, tested before each step, with
  the end value yielded once after the loop. That is the shape CRuby's
  range_each has: an object needs succ and <=> only, and a succ that
  answers the receiver itself is yielded once and not stepped again.
• An Integer begin keeps the native </<= loop — same answers, a step
  rather than three.

ruby-lean/RubyCore/Generated/{Prelude,PreludeJson}.lean are regenerated
(make gen, checked by make gen-check).

Test plan

New regression case difftest/corpus/regressions/range-each-non-succ-begin.rb
(sidecar status: fixed), which pins:

• the TypeError for a Float begin on the block, map and to_a forms, and
  that nothing is yielded first;
• that an Integer begin still iterates, including against a Float end;
• that a class with succ and <=> only iterates (inclusive one visit,
  exclusive none, then three visits), matching CRuby byte-for-byte;
• that a succ returning self with <=> 0 yields one element and stops.

How I validated it:

• Reproduction first: with the sidecar at status: open and the old
  prelude, cd difftest && uv run difftest run --tier regressions --sut lean
  reported still_open: 1 with the diff
  stdout: '[TypeError, can't iterate from Float]...' vs '1.0\n[NoMethodError, ...]'
  and exception: None vs ('NoMethodError', "undefined method '<=' ...").
• After the fix: the same run reports 223 ran, 222 agree, 1 sut_unsupported
  and held for the new case, with no other outcome changed (the one
  sut_unsupported is the pre-existing sorbet-hash-gate.rb, gated).
• make check: ALL CHECKS PASSED — build, every proof book, the axiom
  audits, the checker corpus, difftest over bootstraptest, the regressions and
  tier-3 corpora, the unit tests and their coverage, the docs. The bootstraptest
  watermark is unchanged (ran 1287, agree 1087, sut_unsupported 194, not_compared 6, disagree 0), so no baseline moved.
• make check-not-weakened: CHECKS NOT WEAKENED.
• Also checked by hand against CRuby 4.0.5 that ("a".."c") and (:a..:b)
  still gate rather than raising the new TypeError, and that
  (0..10).step(2)/each_slice gates are untouched.

Note

Authored by Jiro, an autonomous agent (Nori sessions), on behalf of
@sam-xif — including the code and this description.