Closes #56.

Compiling Data.Array aborted with Unexpected bound reference: Ref Nothing (Local (Name "b")) 1, which blocked the purescript-lua-arrays port entirely. The single declaration Data.Array.foldRecM was enough to trigger it.

Root cause

The IR addresses locals with per-name De Bruijn indices: Ref (Local n) k means the k-th enclosing binder named n, counting from the innermost. When an optimizer pass removes a binder, every reference that pointed past that binder (index ≥ 1, because it was skipping over it) has to drop by one. Two passes removed a binder and skipped that repair:

• betaReduce drops the applied lambda. In foldRecM f b array = tailRecM2 go b 0, the outer parameter b collides with tailRecM2's own third parameter b. Inlining tailRecM2 (a single-use dictionary accessor) beta-reduces under that inner b; capture avoidance shifts the outer reference up to index 1, and removing the inner binder left it stranded there instead of lowering it back to 0.
• dead-code elimination blanks an unused named binder to ParamUnused. If an inner lambda's parameter is unused but its body still references an outer binder of the same name (at index 1, skipping the inner one), blanking the binder must lower that reference to 0. This is the same class of bug; the well-scopedness property below found it.

Both sites now repair the survivors with a new unshift, the inverse of shift 1. To keep the two from drifting apart, shift is refactored to share its scope-tracking traversal (overFreeIndex) with unshift; the strict minIndex < index guard leaves genuine inner references alone and keeps the Natural index from underflowing.

Tests

There was no unit coverage for the De Bruijn primitives at all, so the original bug could only surface as an integration crash. This adds a layer that exercises them directly:

• Types spec: round-trip unshift . shift 1 ≡ id and a substitute-identity law as properties, plus the classic capture-avoiding substitution examples (capture under a same-named binder, pass-through of an unrelated binder, a shadowing binder stopping the substitution, a reference reaching past one).
• Optimizer spec: a well-scopedness invariant (optimizing a well-scoped expression must never produce an unbound local) run over a generator of well-scoped terms through the full pipeline. This is what surfaced the DCE sibling bug. Deterministic regressions pin both passes: the foldRecM shape for beta reduction, a shadowing unused binder for DCE.
• Golden: Golden.TailRecM2Shadow, an eval golden over a tailRecM2-based foldRecM paraphrase whose accumulator parameter is named b to reproduce the original collision.

Verified: cabal test is green, and pslua --entry Data.Array now compiles purescript-lua-arrays to luacheck-clean Lua.