Contract eta-expanded type constructors

[rochala]

First of the output-alignment fixes from the Scala 2 / Scala 3 stdlib diff. Independent of #146.

The bug

TASTy stores a type constructor passed where F[_] is expected as the eta-expansion [A] =>> C[A], and the printer rendered it literally. So the Scala 3 output was worse than the Scala 2 output for the very same class — and inconsistent within a single line, with List plain and the other constructor expanded:

// Scala 3 (TASTy), before:
class NonEmptyList[+A] extends NonEmptyCollection[A, List, [A] =>> NonEmptyList[A]] with Product with Serializable
// Scala 2 (pickles), same class:
class NonEmptyList[+A] extends NonEmptyCollection[A, List, NonEmptyList] with Product with Serializable
// Scala 3, after — now identical to the Scala 2 rendering:
class NonEmptyList[+A] extends NonEmptyCollection[A, List, NonEmptyList] with Product with Serializable

This was 51 of the 132 differing lines in the stdlib comparison — the largest category, and every one favoured the Scala 2 rendering.

Scope of the contraction

Only a genuine eta-expansion: every parameter passed straight through, in order, with no bounds of its own. Two things deliberately keep their current rendering, and two pre-existing tests pin them:

• [A <: AnyRef] =>> F[A] — the bound says something F does not.
• [A] =>> F[G[A]] — a composition, not a plain constructor.

Verified against a real artifact: cats.Applicative still shows all 11 of its [α] =>> F[G[α]] composition lambdas.

Verification

• End-to-end per the CLAUDE.md rule, on cats-core_3:2.10.0: NonEmptyList now renders identically to its Scala 2 counterpart, =>> count 0; Applicative unchanged at 11.
• lib.test 372/372.
• The new fixture (CellarSelfBox) reproduces the case, and the test was confirmed to fail without the fix: eta-expansion not contracted: class CellarSelfBox[A] extends CellarBox[[A] =>> CellarSelfBox[A]].

Remaining alignment items (verified, not in this PR)

• scala.jdk.CollectionConverters shows 20 class *HasAs* wrappers but 0 of the defs that attach them — PublicApiFilter.isSyntheticSym drops them.
• scala.Char emits val MaxValue: with a raw U+FFFF and val MinValue: with a raw NUL into the Markdown.

Two categories from the diff turned out not to be cellar bugs: the $default$N accessors don't reproduce outside the experimental stdlib build (0 occurrences on cats-core_3, cats-effect_3, pureconfig), and toString() vs toString reflects genuinely different declarations in the two encodings.

🤖 Generated with Claude Code