perf(array): reuse probe state in RLE, RunEnd and PCO

[joseph-isaacs]

Uses the scalar probe API from #9843 to retain preparation across repeated lookups, including recursive child access. Stacked on #9843; the diff is one commit.

let mut probe = array.repeated_probe();
probe.execute_scalar(10, &mut ctx)?;
probe.execute_scalar(11, &mut ctx)?; // Reuses preparation at each visited slot.

Encodings

• FastLanes RLE: RleProbeState keeps the slice's base value offset. state.split() holds it while the indices, offsets and values children are read through ProbeChildren::slot.
• RunEnd: the ends child probe is held across the binary search, so a repeated read keeps the child's preparation within and between searches; the values child is probed for the selected run.
• PCO: PcoProbeState keeps the validity mask, prefix ranks for non-null positions, page boundaries and every page decoded so far, so each page is decoded at most once per probe. A one-off read decompresses the single row as before.

Each encoding implements probe_scalar(state, index, ctx) once and reads children with state.slot(i); the same body serves one-off and repeated reads. scalar_at delegates through ProbeState::once.

Tests and benchmarks

One-off and repeated access over sliced and nullable inputs; RunEnd and RLE state reuse; PCO physical types, all-null rows and probes that outlive their source. The PCO scalar bench from #9896 now reads through one repeated probe, so CodSpeed tracks the clustered cases against that baseline.

Local medians, 1024 clustered lookups, execute_scalar per row (develop) vs a repeated probe:

array	per-row	repeated probe
PCO	5.54 ms	24 µs
PCO, nullable	4.17 ms	52 µs
RunEnd(PCO, PCO)	24.5 ms	4.1 ms
RLE, nullable	176 µs	133 µs

Scattered PCO lookups over 16 pages go from 5.4 ms to 122 µs, one decode per page; single lookups are equal between the two paths.

Validation: tests for vortex-fastlanes, vortex-runend and vortex-pco, workspace clippy, formatting.

🤖 Generated with Claude Code

[codspeed]

Merging this PR will regress 5 benchmarks

⚠️ 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

⚡ 6 improved benchmarks
❌ 5 regressed benchmarks
✅ 2197 untouched benchmarks
⏩ 293 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	decode_primitives[f32, (1000, 512)]	42 µs	64.2 µs	-34.61%
❌	WallTime	dbp_assemble_kernel_avx512[(I128, 1024)]	467 ns	559 ns	-16.46%
❌	WallTime	filtered_owned_i64_avx2[OneNullInEight]	22.4 µs	25.8 µs	-13.15%
❌	Simulation	take_fsl_u32_random[128, 100]	208.2 µs	237.2 µs	-12.25%
❌	Simulation	take_fsl_f16_random[256, 100]	204.8 µs	230.6 µs	-11.19%
⚡	Simulation	scalar_access[(1024, false, false)]	135,445.9 µs	791.7 µs	×170
⚡	Simulation	scalar_access[(1024, true, false)]	97 ms	1.3 ms	×78
⚡	Simulation	scalar_access[(1024, false, true)]	135.5 ms	2.8 ms	×48
⚡	Simulation	take_fsl_u32_random[256, 10]	159.4 µs	124.4 µs	+28.22%
⚡	WallTime	dbp_assemble_kernel_avx2[(I128, 1024)]	531 ns	464 ns	+14.44%
⚡	Simulation	take_fsl_random[128, 10]	139.4 µs	126.6 µs	+10.1%

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-encodings (96bb450) with develop (24d67e8)}

Footnotes

1. 293 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. ↩

[myrrc]

LGTM