Patched row-encode kernel

[joseph-isaacs]

Part 22 of 25 in the stacked PR series adding vortex-row.

This PR contains exactly one commit; review just that diff in isolation.

What this commit does

Replaces the stub RowSizeKernel / RowEncodeKernel impls for Patched with real implementations.

Size pass: per-row size matches the inner array exactly because patches share the inner dtype. Just delegates to dispatch_size on the inner array.

Encode pass: delegates to dispatch_encode on the inner array, then walks the patch indices and overwrites each patched row's value bytes in place. Patched arrays in our hot paths are Primitive-typed (BitPacked with patches, etc.), so the kernel checks DType::Primitive upfront and declines for anything else.

Pre-cursor snapshot is captured before the inner encoder advances cursors, so the overlay knows each row's slot start position.

Adds patched_i32_* bench triplet. Patched-specific tests live next to the kernel in kernels/patched.rs::tests.

Stack

#	PR	Title	Branch
1	#7986	vortex-row: crate scaffolding	claude/row-c01-crate-scaffolding
2	#7987	vortex-row: add SortField and RowEncodeOptions	claude/row-c02-sortfield-options
3	#7988	vortex-row: codec for fixed-width canonical types	claude/row-c03-codec-fixed-width
4	#7989	vortex-row: codec for varlen canonical types	claude/row-c04-codec-varlen
5	#7990	vortex-row: codec for nested canonical types	claude/row-c05-codec-nested
6	#7991	vortex-row: compute_sizes helper and RowSize ScalarFn	claude/row-c06-rowsize-scalarfn
7	#7992	vortex-row: RowEncode ScalarFn	claude/row-c07-rowencode-scalarfn
8	#7993	vortex-row: convert_columns + tests + bench scaffolding	claude/row-c08-convert-columns-tests-bench
9	#7994	Skip ListView validation in row encoder output	claude/row-c09-skip-listview-validation
10	#7995	Add validity fast-path helper for the four pattern-matching encoders	claude/row-c10-validity-fast-path
11	#7996	Skip zero-init of output buffer	claude/row-c11-skip-zero-init
12	#7997	Auto-vectorize pure-fixed offsets construction	claude/row-c12-vectorize-pure-fixed-offsets
13	#7998	Auto-vectorize mixed-path offsets construction	claude/row-c13-vectorize-mixed-offsets
14	#7999	Rewrite varlen 32-byte block encoder with copy_nonoverlapping	claude/row-c14-varlen-block-copy-nonoverlapping
15	#8000	Walk VarBinView rows directly in row encoder hot loop	claude/row-c15-walk-varbinview-directly
16	#8001	Add arithmetic-write fast path for fixed-before-varlen columns	claude/row-c16-arith-write-fast-path
17	#8002	Specialize Constant for the arithmetic-write fast path	claude/row-c17-specialize-constant-arith
18	#8003	RowSizeKernel and RowEncodeKernel dispatch helpers	claude/row-c18-kernel-dispatch-helpers
19	#8004	Inventory-based registry for downstream encoding kernels	claude/row-c19-inventory-registry
20	#8005	Constant row-encode kernel	claude/row-c20-constant-kernel
21	#8006	Dict row-encode kernel	claude/row-c21-dict-kernel
22	#8007	Patched row-encode kernel	claude/row-c22-patched-kernel
23	#8008	RunEnd row-encode kernel (vortex-runend)	claude/row-c23-runend-kernel
24	#8009	BitPacked row-encode kernel (vortex-fastlanes)	claude/row-c24-bitpacked-kernel
25	#7985	FoR and Delta row-encode kernels (vortex-fastlanes)	claude/row-pr3-kernels

Base of this PR: #8006 (claude/row-c21-dict-kernel)
Next in stack: #8008 (claude/row-c23-runend-kernel)

Combined context

For the full design + rationale, see PR #7985 (top of stack).

[codspeed-hq]

Merging this PR will not alter performance

⚠️ 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.

⚡ 2 improved benchmarks
❌ 1 regressed benchmark
✅ 1218 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	chunked_varbinview_into_canonical[(100, 100)]	358.4 µs	325 µs	+10.26%
❌	Simulation	chunked_varbinview_canonical_into[(100, 100)]	273.3 µs	309 µs	-11.56%
⚡	Simulation	chunked_varbinview_canonical_into[(1000, 10)]	197.9 µs	161.9 µs	+22.25%

Tip

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

---

_{Comparing claude/row-c22-patched-kernel (b86974e) with develop (faf7e42)¹}

Footnotes

1. No successful run was found on claude/row-c21-dict-kernel (62e41b4) during the generation of this report, so develop (faf7e42) was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩