Implements the worker/wrapper design from #24: curried functions stop paying a Lua call plus a closure allocation per argument at their saturated call sites.

What changed

• The IR gains the definition-side counterpart of feat: represent application as an n-ary AppN call node (#179) #199's AppN: Abs becomes the singleton pattern synonym of an n-ary AbsN, so every unary rewrite rule compiles unchanged. The Lua backend emits one n-ary function for a multi-parameter AbsN, dropping the trailing run of unused parameters and eliding the trailing run of Prim.undefined arguments on the call side (a non-trailing one lowers to nil).
• A new pass (Language.PureScript.Backend.IR.Uncurry) splits every binding with manifest arity n ≥ 2 and at least one saturated call site into a worker f$w = AbsN [p₁…pₙ] body plus a curried wrapper f = λx₁.…λxₙ. f$w(x₁, …, xₙ), and rewrites the saturated sites to direct worker calls: add(x)(y) becomes add$w(x, y). Partial applications and higher-order uses keep going through the wrapper, so a mixed-use function wins at its saturated sites instead of being disqualified. Unreferenced wrappers are removed by the new post-uncurry optimize+dce fixpoint. Top-level candidates are collected module-wide; local (Let-bound) candidates per top-level site, since GUC is per-site. Recursive functions split too: mutual recursion becomes direct worker-to-worker calls, and a local loop like go acc n typically loses its wrapper entirely, which is the shape Lower self-recursive tail calls of uncurried workers to while loops (loopification) #181 needs.
• One deliberate deviation from the issue design: the wrapper is not marked inline-always. The call-site inliner that plan relies on is Budgeted call-site inlining of dictionary methods to collapse non-Effect/ST monadic chains #180, which does not exist yet; today's inliner substitutes at every occurrence, which would paste lambda literals into argument positions (a closure allocation per evaluation at map f xs sites) and drop the shared top-level wrapper. The pass rewriting its own saturated sites gets the same result on exactly the sites that benefit. A second addition to the plan is the saturated-site precondition: without it a split with no sites would leave a permanent wrapper→worker indirection, because the use-once revert path does not reach recursive-group members or local sibling references.
• betaReduce handles the exact-arity n-ary redex (subsuming betaReduceUnusedParams), which is what collapses a single-use worker pasted into its one call site. It declines on repeated parameter names, where left-to-right substitution would steal occurrences from the parameter Lua actually binds.
• FlattenDeepBinds recognises the n-ary shapes: a chain step through w(action, \x -> rest) and spine depth through any argument of an n-ary call. Without this, Golden.LongCallbackChain and Golden.LongApplyChain would fail the Lua 5.1 nesting check the moment their sites get uncurried; the step rebuild preserves the call shape, since w(a, k) and w(a)(k) denote different programs.
• The WellApplied invariant generalizes to exact-arity application of literal lambdas plus the trailing-ParamUnused condition, and now rides along with GUC at every pass boundary, so the checked runner verifies it on every golden. DCE blanks only a dead suffix of an AbsN parameter list to keep the trailing-run condition.
• Minted names ($w, $p<i>, $u<i>) are deterministic and bypass the shared supply, so downstream $kont/$tmp numbering stays put.

Out of scope

Lifting runFnN/mkFnN stays with #198 (it needs the #178 lifter; AbsN unblocks the mkFnN half). Effect-shaped functions mostly stay curried for now: thunk-forcing sites only appear after magicDo, which runs later, so the pass never sees them saturated. The same applies to the $kont helpers minted by the flattening pass. Both gaps are tracked in #200 (a late, post-flattening run of the same pass). Constructors, which the codegen still emits as curried closure chains, are untouched and tracked in #201.

Verification

• cabal test all: 599 examples, 0 failures. The pass has 11 unit tests (split, partial, over-application, recursive groups, local Lets, per-site scoping, @inline never, unused parameters, annotations) and 3 property tests over generated modules (the output lints clean for GUC and WellApplied, a second run is a no-op, and WasRewritten is signalled exactly when the module changed). Four more tests run the whole checked pipeline to fix the split's interaction with inlining and DCE: the wrapper is dropped once every site calls the worker, a used-once function leaves no residue, an exported function whose only saturated consumer is dead comes out unsplit, and an @inline always binding is claimed by the inliner before the split. Alongside are the n-ary beta and flattening cases and the mirrored AbsN tests for alphaEq, countFreeRefs, the linter, and codegen. The new tests were run red first: 9/11 pass unit tests fail against an identity pass, 5 of the beta/flattening tests fail against the pre-change implementations, and the property and pipeline tests fail against a pass mutated to mint duplicate wrapper binders and to over-report Rewritten.
• The new Golden.Uncurry golden runs saturated, partially applied, higher-order, mutually recursive, locally recursive, over-applied and unused-parameter shapes through lua against a hand-written eval oracle. Every pre-existing eval/golden.txt is byte-identical: the semantic net for the whole transformation held.
• Bench counters: Bench.BindChain drops one per-call FNEW, one trace-abort site and one blacklist entry with an unchanged workload result. A new Bench.CurriedStep macro pins the canonical target, a hot saturated two-argument loop. Compiling it with the pre-split pslua (main at d9d2d58) and with this branch, at n=100000 with an identical result=5000050000, the split removes the per-iteration closure the recursive go allocated: function-body FNEW drops from 6 to 5, distinct trace-abort sites from 4 to 3, and blacklisted bytecodes from 5 to 4. Median wall-clock falls from 22.0ms to 18.1ms on PUC Lua 5.1, from 22.6ms to 17.3ms on LuaJIT, and from 22.6ms to 17.9ms on LuaJIT with the JIT off. The three trace aborts that remain are the curried FFI closures (eqInt, intAdd, intSub), which is Lift the *.Uncurried wrappers to direct AppN calls #198's half.
• Structural golden churn is large but mechanical: every golden.ir moves with the AbsN Show shape, and 17 golden.lua files change where splits fire.

Closes #24.