test(fixtures): multi-bit Hack ALU fixtures (parametric + single-file)

[jeffersonmourak]

Summary

Two multi-bit Hack ALU test fixtures, each demonstrating a different language usage pattern:

Parametric project (tests/fixtures/projects/alu_parametric/):

• A parametric Hack-style ALU sub-circuit (alu.circ) generic over <W>, plus four per-width root files (alu_2bit.circ, alu_4bit.circ, alu_6bit.circ, alu_8bit.circ) that instantiate it.
• Each wrapper imports the parametric core and supplies a width-specific ripple-carry adder — the language can't express the carry chain parametrically (no iteration over bit positions).
• Truth-table goldens included for the 2-bit (1024 rows) and 4-bit (16384 rows) wrappers; 6-bit (262k rows) and 8-bit (4M rows) are over the default --truth-table cap and are exercised via the .wasm runtime instead.

Single-file 4-bit (tests/fixtures/circuits/alu_4bit_multibit.circ):

• A flat single-file companion to the existing scalar alu_4bit.circ — same Hack ALU logic, width-4 buses, no imports.
• Truth-table golden matches the parametric 4-bit golden bit-for-bit on all 16,384 rows.

Why this is useful even with the workarounds

Both fixtures exercise three multi-bit-language papercuts surfaced while building them; each is worked around inline and documented in the fixture comments:

1. Scalar→width-W broadcast can't go through a {c,c,...} concat directly into a parametric sub-circuit instance port. The wrappers route every scalar control through a wire[W] first.
2. Sliced macro outputs feed scalar macros only via intermediate scalar wires — slice-into-scalar-macro trips the topology builder otherwise.
3. xor[W] macro produces incorrect simulation values (or breaks topology build entirely) when its output flows through a wire-buffered slice chain into per-bit scalar macros. Both fixtures hand-expand the bit-parallel XOR propagate from and/not primitives. The variant that trips InternalError rather than producing wrong values appears specifically in the feedback chain from a sub-circuit instance's output back into its own input.

When these are fixed in the validator / topology builder / simulation engine, the workarounds collapse to one-line xor[W] calls and the goldens stay bit-identical — a clean diff later.

Test plan

• zig-out/bin/circ-compile tests/fixtures/projects/alu_parametric/alu_2bit.circ --truth-table produces a 1024-row table matching the 2-bit golden.
• zig-out/bin/circ-compile tests/fixtures/projects/alu_parametric/alu_4bit.circ --truth-table produces a 16384-row table matching both the parametric 4-bit golden and the existing scalar alu_4bit.truth.golden on every canonical Hack ALU operation.
• zig-out/bin/circ-compile tests/fixtures/circuits/alu_4bit_multibit.circ --truth-table produces a 16384-row table identical to the parametric 4-bit golden (`diff` exit 0).
• zig-out/bin/circ-compile tests/fixtures/projects/alu_parametric/alu_6bit.circ --truth-table --truth-table-cap=18 produces a 262k-row table (no golden checked in at this size).
• zig-out/bin/circ-compile tests/fixtures/projects/alu_parametric/alu_8bit.circ -o /tmp/alu_8bit.wasm builds a self-contained ~955 KB .wasm.
• Fixture-only — no source-code changes.