Code Review: feat: add NIP, TUCK, PICK, ROLL stack operations

Overview

Adds four new stack manipulation words across all three layers: ODS table definitions, Forth-to-MLIR translation, and MemRef conversion patterns. PICK and ROLL are dynamic (runtime-indexed), with ROLL using an scf.for shift loop. Clean 282-line addition.

Correctness

All four conversion patterns are correct. Verified each against standard Forth semantics:

• NIP (a b -- b): Loads top, decrements SP, stores at new SP. ✓
• TUCK (a b -- b a b): Loads both, rearranges, increments SP. ✓
• PICK (xn...x0 n -- xn...x0 xn): Pops n, computes SP' - n, loads and pushes. ✓ Edge case 0 PICK = DUP works.
• ROLL (xn...x0 n -- xn-1...x0 xn): Pops n, saves target, shifts with scf.for, stores saved at top. ✓ Edge cases: 0 ROLL = no-op, 1 ROLL = SWAP, 2 ROLL = ROT — all verified.

Stack pointer convention (SP points to top element) is consistent with existing patterns.

Code Quality & Style

• Follows existing project patterns exactly (constructor, OneToNOpAdaptor alias, boilerplate structure). Good consistency.
• Stack effect comments on each struct match Forth standard.
• ODS definitions are well-structured with proper summary/description fields.
• Pattern registration list reformatting is a clean diff.

Potential Issues

1. No bounds checking on PICK/ROLL n values — A negative or out-of-range n would read/write out of bounds on the 256-element memref. Consistent with the project's design (untyped stack, programmer ensures safety), so not a blocker, but GPU kernels using these words need care.

2. ROLL with n=0 — When n=0, rolledAddr == spAfterPop, so the loop range is empty (zero iterations). The saved value is stored back at the same position. Correct but results in a redundant load/store. Minor, not worth optimizing.

3. Pure trait on PICK/ROLL — Fine at the Forth dialect level (pure SSA transformations on !forth.stack). Consistent with SwapOp, RotOp, etc.

Test Coverage

• Translation test (stack-ops.forth): Verifies all four words parse to correct forth.* ops. ✓
• Conversion test (stack-manipulation.mlir): Tests MemRef lowering with CHECK patterns covering all load/store/loop sequences. Pushes literal 2 before PICK and ROLL to exercise dynamic indexing. ✓
• Missing: no end-to-end pipeline test in test/Pipeline/. Not a blocker since conversion tests cover the patterns, but would add confidence.

Suggestions

1. Update CLAUDE.md — The Supported Words section lists DUP DROP SWAP OVER ROT but does not include NIP, TUCK, PICK, ROLL. Should be updated to keep docs in sync.
2. Consider a pipeline test (optional, low priority).

LGTM with the documentation nit. 👍