feat(array): add scalar probes with lazy child contexts

[joseph-isaacs]

Summary

Add a borrowed scalar probe API so an encoding can retain preparation across a short series of random reads. This is the first PR in a two-PR stack; all encodings, including primitive, PCO, RLE, and RunEnd, use unit state and the default execute_scalar path here.

The encoding implementations and paired benchmarks are in #9844, based on this PR’s branch. Merge this API layer first.

Changes

• Add ArrayRef::probe(Once | Repeated) and bounds-checked scalar access including nullness.
• Pass Option<&mut ProbeCtx<'a, ProbeState<'a>>> through the existing array vtable. The context combines encoding-local state with lazy, reusable child probes indexed by source slot.
• Keep construction allocation-free and initialize contexts once on first in-bounds repeated access. Store small contexts inline, with a heap fallback and typed destruction for larger contexts.
• Test default behavior, bounds/nulls, lazy child creation, slot/source identity, independent caches, split mutable access, and storage initialization, moves, alignment, and destruction.

API Changes

Custom OperationsVTable implementations must declare type ProbeState<'a> = (); unless they need local state. Overriding probe_scalar is optional; its default preserves existing scalar execution.

Validation

• cargo nextest run -p vortex-array -p vortex-pco -p vortex-fastlanes -p vortex-runend: 3,886 passed, one skipped.
• cargo test --doc -p vortex-array: passed.
• Clippy with all targets/features for those four crates: passed.
• Four isolated Miri tests of the current erased storage: passed.
• Workspace Clippy was attempted and blocked by the nvCOMP build script downloading the NVIDIA archive in this environment.
• cargo +nightly fmt --all, Python crate formatting check, and git diff --check: passed.

Cargo checks use RUSTC_WRAPPER= after the sandbox's sccache permission failure. Encoding optimizations, their tests, and paired benchmarks belong in the stacked follow-up.

[codspeed-hq]

Merging this PR will regress 1 benchmark

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 13 improved benchmarks
❌ 1 regressed benchmark
✅ 2183 untouched benchmarks
⏩ 218 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	dict_canonicalize_gt_u8_neon[1000000]	487.1 µs	541.9 µs	-10.11%
⚡	WallTime	arrow_checked_add_u32_neon[16384]	20.4 µs	13.5 µs	+51.69%
⚡	WallTime	filtered_sink_i64_avx2[OneNullInEight]	31.2 µs	21.9 µs	+42.2%
⚡	WallTime	mul_u64_nonnull_neon	21.3 µs	15.1 µs	+40.77%
⚡	WallTime	filtered_sink_i64_avx512[OneNullInEight]	31.2 µs	22.3 µs	+40.12%
⚡	WallTime	filtered_sink_i64_avx2[NineNullsInTen]	16.4 µs	13.1 µs	+25.3%
⚡	WallTime	filtered_sink_i64_avx512[NineNullsInTen]	16.5 µs	13.5 µs	+22.41%
⚡	WallTime	dict_canonicalize_gt_u8_avx2[16000000]	8 ms	6.8 ms	+17.62%
⚡	WallTime	multiply_shapes_neon[(16384, PerRowPerRow)]	20.3 µs	17.3 µs	+17.56%
⚡	WallTime	mul_i64_nonnull_neon	20.2 µs	17.2 µs	+17.51%
⚡	WallTime	filtered_owned_i64_avx2[OneNullInEight]	25.8 µs	22.3 µs	+15.9%
⚡	Simulation	allocate_drop_arrow[0]	456.9 ns	402.7 ns	+13.45%
⚡	Simulation	compact_sliced[(1024, 90)]	1.5 µs	1.4 µs	+11.92%
⚡	Simulation	compact_sliced[(2048, 90)]	1.6 µs	1.4 µs	+11.42%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ji/array-probe-api (8146a99) with develop (d1fe2dc)}

Footnotes

1. 218 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩