Closes #174.

What

Every top-level binding currently lives in the module-scope table M, so every inter-binding reference is a hash lookup. This adds Language.PureScript.Backend.Lua.Promote, a Lua-level pass that promotes the bindings actually worth it to real chunk locals: M.foo = e becomes local foo = e, and reads become local/upvalue accesses.

• Selection: a binding qualifies when its field is written by exactly one top-level M.x = e statement, is read at least once, and its name isn't used as a variable anywhere else in the chunk (a name collision, which only hand-written FFI produces, declines rather than renames). Qualifying bindings are promoted in descending static read count while the chunk's locals budget allows.
• Upvalue accounting: the pass computes every function proto's upvalue demand bottom-up (demand(f) = ownOuterRefs(f) ∪ children's pass-through demand), lexically resolved. When a proto would exceed the upvalue ceiling, promoted-field references inside it demote back to M.x, cheapest reads first, and the binding is mirrored into M (M.x = x right after local x = e) so the demoted reads still find it.
• Recursion: a binding read before its own initializer (self-recursion, or an earlier member of a mutually recursive group) is pre-declared and initialized by plain assignment instead.
• Budgets are pure degradation: every overflow, whether locals, upvalues, or a name collision, falls back to exactly today's M-only form, so the target's per-function limits (Function at line XXX has more than NNN upvalues #19) stay handled by construction rather than by assuming programs stay small.

Promotion runs before the existing stage-1 per-function field caching (Language.PureScript.Backend.Lua.Localize) in optimizeChunk, so that pass now only ever caches whatever M traffic promotion left behind. ADR 0001 is updated from "stage 2 planned" to record the implemented design.

Red → green

The spec commit lands first against the existing identity-placeholder skeleton: 13 examples, 8 failures (promotion, pre-declaration, both budgets, the collision decline, and the stability precondition all red; the two limits-driven end-to-end eval cases pass trivially since M always survives a no-op). The feat commit turns all 13 green.

Corpus

50 of the corpus's 62 golden modules move; 35 of those drop the M table entirely once every read in the chunk is a local/upvalue access. Net -39 lines. Where M survives, it's either the budget-pressure cases (FieldCaching, UncurriedLift, Uncurry) exercising stage 1's own caching on the residual traffic, or a genuine zero-read field, which is a golden-harness artifact rather than a production one: compileCorefn always IR-links a module AsModule, rooting DCE at every export, even when that module is Lua-linked AsApplication for its eval check. DerivedFunctor.Test's functorEither dictionary survives IR-level DCE as an export even after the one call site that used it gets constant-folded away, so it reaches Lua codegen with no readers left; promotion correctly leaves a zero-read field in M rather than spending a local slot on a dead store. A production AsApplication build roots IR-level DCE at the entry point instead and would prune these before Lua codegen ever sees them. All 31 eval oracles are byte-identical (never auto-accepted), and luacheck is clean under the same flags the test suite uses.