fix: VarBinArray.DictMode reads dict offsets in a uniform loop, unblocking vectorization (#243)

[dfa1]

Summary

VarBinArray.DictMode.forEachByteLength called dictReadOff(long) twice per row, and each call recomputed the invariant offset width, ran an invariant bounds check, and dispatched a variable-target switch(dictValOffPType). Per the CLAUDE.md hot-loop rule, a variable-target switch per element blocks C2 auto-vectorization (regression introduced by #215's true-width offset read).

The fix branch-splits the hot loop, hoisting both ptype dispatches out of the per-row body:

• I32 fast path (forEachI32OffsetByteLength): the common FSST/dict shape where value offsets are 32-bit. The dictCodesPType switch is hoisted once, and each specialized loop reads codes at a single fixed stride and computes lengths via i32OffsetLength at a constant 4-byte offset stride — leaving a uniform, vectorizable body with zero per-row switches.
• Wide-offset general path: 64-bit and narrow offset ptypes keep the existing per-row dictReadCode/dictReadOff path (documented slow path).

Scalar accessors (getBytes/getByteLength/getString) intentionally keep the existing dictReadOff helper: its per-call VortexException bounds guard (ADR 0003) is pinned by getString_widthOvershootsBuffer_throwsVortexException, which constructs the array in // Given and expects the throw at access time — so the check cannot move to the constructor.

Tests

• Existing VarBinArrayTest stays green.
• Added forEachByteLength_i32Offsets_resolvesViaDict (fast path) and forEachByteLength_i64Offsets_resolvesViaDict (general path).
• ./mvnw test -pl reader -q passes.

Closes #243

🤖 Generated with Claude Code