What changed

• Generate a 256-byte operand-size table at comptime from the authoritative Op.spec() schema.
• Inline Op.operandSize() as one indexed load instead of re-running instruction metadata and operand-layout classification at runtime.
• Document that the table is derived cache data, not a second hand-maintained schema.

Why

ReleaseFast profiles attributed roughly 31% of Richards, 19% of DeltaBlue, and 22% of RayTrace samples to runtime operand-size metadata. Merged opcode-family handlers call this path on every execution even though the bytecode schema is immutable.

Impact

The serialized remote 12-pair A/B improved all six interpreter macros: Richards 0.582x, DeltaBlue 0.725x, Crypto 0.681x, RayTrace 0.748x, Navier-Stokes 0.765x, and Splay 0.849x, for a 0.720x geometric mean. The default production-tier geometric mean was 0.706x. Interpreter prop_access measured 0.576x, and the full micro suite reported no regression past the paired-run threshold. Dispatch counts and bytecode encoding are unchanged.

Validation

• zig build test-fast
• exhaustive Op: schema/table equality tests
• full non-cached test262: 48,653 pass / 1,324 fail, 97.35% (exact main pass set)
• independent code review: no findings
• serialized remote 12-pair A/B, both tiers: no regressions