feat(ir): uncurry the shapes that only appear after magicDo and flattening - #264

Closes #200.

The worker/wrapper uncurrying pass (#24) now runs a second time at the end of the IR pipeline, after flattenDeepBinds, followed by a single dce pass. The late run catches the saturated call sites that do not exist at the early run's position: the effect-run spines magicDo completes, and the $kont helpers minted by flattening. The helpers are saturated by construction, so their curried kontN(live)(x) calls become kontN$w(live, x) and the wrappers die in the closing dce pass.

One correction to the issue's premise surfaced while updating goldens. An effect function with two or more real arguments is already split by the early run, because its real-argument spine is saturated before magicDo runs: bindE (f a b) k contains the full f(a)(b). For those functions the late run adds nothing; the thunk stays inside the early worker, and absorbing it there is a possible follow-up. The family the late run actually wins is the unary effect action. f :: A -> Effect B has manifest arity 1, below the split threshold, until magicDo appends the thunk parameter, and the late split absorbs that parameter into the worker: a fully applied effect statement compiles to the single call f$w(a) instead of f(a)() with a thunk allocated per call. The new Golden/UncurryEffect test covers this end to end, with an eval oracle for semantics.

Three pieces of groundwork make the rerun sound:

• isEffectRun now recognises an effect run by its trailing EffectRunArg instead of only the singleton application. Without this, rewriting f(a)(b)(run) to f$w(a, b, run) would strip the statement's protections and dead code elimination would silently drop the effect. New DCE and inliner tests cover the protections; the Lua codegen already erased trailing markers from n-ary argument lists.
• The pass is now rerun-safe. A wrapper built by the earlier run (recognised by its delegate-call body) is never split again, since that would mint a duplicate $w binding. It re-registers its arity instead, so saturated sites that appeared between the runs still reach the existing worker. A non-wrapper candidate whose $w name is taken is left alone entirely.
• The cleanup after the late run is a lone dce pass, not the optimize+dce fixpoint the issue sketched. Optimize's use-once inlining would paste the $kont workers (each called exactly once) back into their call sites one fixpoint round at a time, re-nesting exactly what flattenDeepBinds had just flattened and tripping Lua's parser nesting cap again.

Golden churn is confined to the four Long*Bind modules (the $kont split) plus the new test, and every eval oracle passes unchanged. The full suite is green (899 examples) and the randomized specs hold across five extra seeds. The new bench/macro/effect_step bench records the improved shape and commits baseline counters. Wall-clock on that micro is flat: the step body's ST-ref FFI calls are still curried and dominate the saved call and thunk allocation, which is itself an argument for the follow-ups below.

Two follow-up candidates came out of this work and stayed out of scope: absorbing the thunk parameter into workers the early run already split (the logShow$w(x, true)() shape visible in the Primops golden), and distributing an effect-run application into IfThenElse branches, which an if-terminated effect loop still needs before loopification (#181) can turn it into a while loop.