Replace ts-toolbelt with local type utilities

[connorshea]

Fixes #148.

AI Disclosure: Generated with Claude Code, Opus 5. Reviewed and tested by me, added a bunch of tests where there weren't any to ensure no regressions. I'm happy to split this into smaller pieces if needed (e.g. add the tests in one PR, then add port the local type utils, or split the type util ports into smaller sets rather than one big PR).

---

ts-toolbelt's entry point is a namespace barrel: out/index.d.ts eagerly imports 13 _api.d.ts
namespaces, so all 241 of its files are pulled into the program of every project that consumes
these types. skipLibCheck hides the type-checking cost but not the parse and bind cost.

Most of that cost is one type. Iteration/IterationMap is a 202-entry lookup table (-100..100)
whose elements are typed keyof IterationMap — a 202-member literal union — because ts-toolbelt
predates TypeScript 4.5 and had to count without variadic tuples.

Measured on a project importing a single ramda function:

	files	memory	type instantiations
before	253	92 MB	7,309
after	13	73 MB	3,441

(The 19-function benchmark shows the same shape; instantiations roughly halve once the Iteration
table is gone.) With skipLibCheck: false the check time for the barrel alone was 0.35s and
57,135 types.

What

The twelve types this package actually needed from ts-toolbelt are reimplemented in
types/util/_internal.d.ts, recursing over variadic tuples rather than counting — so there is no
arithmetic and no ±100 element cap.

package.json now has no dependencies block at all; the package has zero runtime
dependencies.

A util file whose name starts with _ is internal: npm run build imports from it and ships it
alongside the other declarations, but does not re-export it, so none of this widens the public API.
That convention is documented in CONTRIBUTING.md and implemented in scripts/buildScripts.mjs.

The generated es/index.d.ts differs from the previous build only in its import header and in
three renames (x → PlaceholderBrand, F.Parameters → the global Parameters, F.Return →
ReturnType). Every exported declaration keeps its shape.

How it was checked

• A differential battery type-checks the new implementations against ts-toolbelt's originals for
  invariant type identity — (<T>() => T extends A ? 1 : 2) extends (<T>() => T extends B ? 1 : 2)
  — plus separate probes for inference behaviour, which identity does not capture. That mattered:
  writing Narrow as A extends [] ? A : NarrowRaw<A> produces an identical type but defers the
  conditional, and a deferred conditional is not an inference site, so ['a', 'b'] widened to
  string[]. Routing through the Try indirection is load-bearing and is commented as such.
• A mutation sweep over types/util/_internal.d.ts — break one behaviour, run
  build + typecheck + tsd, restore — initially left 13 of 17 mutations alive, because
  curry, curryN, binary, nAry, flatten, unnest, mergeDeepLeft and mergeDeepRight
  had no test files at all. That sweep found a genuine bug: Flatten recursed structurally into
  each element instead of un-nesting to a fixpoint, so flatten(x: number[][][]) returned
  number[][] and flatten(x: [number[][], string]) silently dropped the array. Fixed in the last
  commit, along with tests that bring the survivors down to two.
• The two remaining survivors are unobservable through the public API and are kept for fidelity
  with the original: Primitive's bigint (the only branch it gates produces a never
  intersection either way) and MergeProp's optional-key branch (an optional key's type always
  includes undefined, so the OK extends undefined fallback covers it).
• ts4/es/index.d.ts compiles clean under a real TypeScript 4.9.

npm run build, npm run typecheck, npm run lint and npx tsd (73 files) are all green.

🤖 Generated with Claude Code