Tier 2: Fixed-size arrays and optimizations

[kavanaghpatrick]

Summary

• Add 4 new array opcodes (ALLOC, ALOAD, ASTORE, ALEN) wired through full pipeline: opcodes → VM → bytecode → Metal kernel → GPU glue → C compiler
• Transpiler support for arr = [0.0] * N, arr[i] read/write, arr[i] += val augmented assignment
• sum(arr) and len(arr) builtins for array variables
• Constant folding: 4.0 * 3.14159 → single PUSH at compile time
• Type inference: skip unnecessary float coercion when dividend is known-float
• GPU arrays: thread-local storage (8 arrays x 256 floats per thread)
• 831 tests passing (48+ new tests added)

Closes #28

Test plan

• Array allocation, read, write, augmented assignment in VM
• Array bounds checking and error messages
• Transpiler handles [0.0] * N, subscript access, sum(), len()
• Constant folding reduces instruction count
• Type inference skips float coercion for known-float variables
• Bytecode encoding for all 4 array ops
• Metal kernel compiles with array storage and opcodes
• C compiler emits array declarations and access code
• Full regression: 831 tests, 0 failures

🤖 Generated with Claude Code