fix(codegen): alias UDT array-element loop-locals so field mutation writes back

[luisleo526]

Summary

Pine array elements of a user-defined type are references. A global-scope, non-var, loop-nested UDT VarDecl bound from a UDT-array-element lvalue (z = arr.get(i) / .first / .last) that is later field-mutated must alias the element so the mutation writes back into the array.

Previously this mis-lowered to a value copy:

• a global while loop hoisted the local to a class member whose in-loop init became a value-copy assignment, and
• a global for loop kept it a true local but the function-local alias path no-ops at global scope (_current_func_body is None).

Both silently dropped the field mutation, causing stale-array over-fire.

Fix

_register_udt_array_get_ref_locals detects exactly this shape (UDT-array-element RHS AND global field-mutation AND loop-nested), suppresses the hoisted class member, and emits a fresh per-iteration UDT& z = arr[i]; reference — the same alias form the non-hoisted function-local path already produces.

Strictly gated: read-only get-locals keep value-copy semantics and function-local get-locals are excluded, so all other output is byte-identical.

Verification (independent R7 regression gate)

• pytest: 1449 passed / 1 skipped / 0 failed (PYTHONHASHSEED=0).
• Byte-identity proof: transpiled 257 corpus/validation + 77 data/standard + 2 finnp17 beneficiaries under PYTHONHASHSEED=0, fix vs HEAD. Only the 2 finnp17 dirs differ; every corpus and standard strategy is byte-identical. The finnp17 diff is precisely Zone z = Zone{} / z = active_zones[i] → Zone& z = active_zones[i].
• Corpus gate: excellent=239, strong=12, anomaly=1 — no demotion.
• Sweep: finnp17-atm (+ twin) weak → strong; engine/TV trade-count delta collapses from a ~1300-trade over-fire to +5; 4 tripwires + 3 3commas DCA/grid sentinels unchanged; 0 downward tier moves.

🤖 Generated with Claude Code