Summary

• Thread full-width and i16 Smi BitAnd successors through Lantern's existing Int32 fast path.
• Preserve the ordinary bytecode and both JIT tiers; non-Int32 operands still use the canonical coercion/BigInt/throw path.
• Deliberately exclude LdaSmi8 to avoid charging Splay's 2.08 million nonmatching compact loads.
• Add direct-transfer, exclusion, Double, coercion, and caught-throw coverage plus benchmark and conformance history.

Why

Crypto executes 3,158,688 LdaSmi16 -> BitAnd pairs and 1,488,595 full-width LdaSmi -> BitAnd pairs. The retained implementation bypasses 4,646,882 indirect successor dispatches while keeping all logical bytecode telemetry unchanged.

Broader opcode and split-handler prototypes were rejected because they regressed zero-hit workloads or caused a whole-function LLVM layout phase change. The retained in-arm hybrid adds only 96 native bytes and uses an unlikely side edge so the dominant ordinary decode path remains favored.

Performance

Pinned x86_64, CPU 3, 40+40 paired physical roles:

• six-macro neutral geometric mean: 0.9795x (2.05% faster)
• targeted Crypto 60+60: 0.9710x
• zero-hit arith_loop 60+60: 0.9895x
• worst macro: Splay 1.0055x
• 60+60 confirmations: DeltaBlue 0.9881x, Splay 1.0095x

No workload crossed the 2% regression ceiling.

Validation

• Focused bytecode-stats/coercion tests: 4/4 pass
• Full ReleaseSafe units: 3,184 pass, 464 intentional skips
• Exact-main/candidate bitwise-and test262 pass lists: identical at 29 pass / 1 known fail
• ReleaseSafe gc-threshold=1 bitwise-and sweep: identical 29/1 result
• Full non-cached test262: 48,653 pass / 1,324 fail, no change
• ShadowRealm sweep: 63 pass / 1 fail, no change
• Complete Crypto stats reports are byte-identical after normalizing only direct-transfers