Reuse the element scratch buffer across rows in json_get_array

[samuelcolvin]

Follow-up to #139.

The json_get_array fast path buffered each row's element slices in a SmallVec<[&str; 8]>, which allocates once per row for arrays longer than eight elements.
This replaces it with a Vec<Range<usize>> created once per batch and cleared per row, so its capacity is reused across rows. The buffer grows only when a row has more elements than any earlier row in the batch, so the allocation count per batch depends on the widest row, not on the number of rows.
Storing byte ranges rather than borrowed &str is what lets one buffer outlive each per-row closure call.

Elements are now sliced from the input &str by range, an O(1) char-boundary check, instead of a std::str::from_utf8 pass over each element.
jiter stops on ASCII structural bytes, so every range boundary is a char boundary.

invoke_array_scalars_direct takes FnMut so the closure can capture the buffer.
The smallvec dependency is removed.

Null-on-error behaviour is unchanged: ranges are only sliced after the whole array parses, and the existing malformed_late_element_does_not_append_partial_values test still covers that.
The equivalence test runs rows of 5, 9, 2, 0 and 1 elements in sequence through the shared buffer.

Benchmarks

The existing json_get_array_array bench uses eight elements, which fit the old inline buffer, so this adds json_get_array_array_wide with 32 elements.
Both are 1,024-row Utf8View batches, cargo bench --bench main release builds on an Apple M3 Max, criterion --save-baseline on main's source and --baseline on this branch, same bench file for both.
Two before/after pairs, reported as criterion's mean estimate:

bench	elements	before	after	change
json_get_array_array	8	382 µs	331 µs	-12.8%
json_get_array_array	8	373 µs	322 µs	-13.0%
json_get_array_array_wide	32	1.490 ms	1.200 ms	-18.0%
json_get_array_array_wide	32	1.485 ms	1.179 ms	-20.0%

The eight-element gain comes from dropping the per-element from_utf8 pass; the extra gain at 32 elements is the per-row heap allocation the old buffer spilled into.
A first baseline run measured 487 µs and 1.876 ms, well outside the other two, and was discarded as noise.

cargo test and cargo clippy --all-targets -- -D warnings pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_01L2MYrU7T15KLX6ASKAMu7Y

[codecov-commenter]

Codecov Report

❌ Patch coverage is 95.83333% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 93.09%. Comparing base (2538b69) to head (c607e24).

Files with missing lines	Patch %	Lines
src/json_get_array.rs	95.45%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #142      +/-   ##
==========================================
+ Coverage   93.05%   93.09%   +0.03%     
==========================================
  Files          18       18              
  Lines        1815     1824       +9     
  Branches     1815     1824       +9     
==========================================
+ Hits         1689     1698       +9     
  Misses         70       70              
  Partials       56       56              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Copilot review overview

🟢 Approval recommended

The implementation preserves behavior and is adequately tested, with only a minor allocation-claim discrepancy.

Review effort: Balanced
Findings: 1

Open (1)

• Vec may reallocate multiple times despite single-allocation guarantee · New

What changed in this PR

Reuses a batch-scoped scratch buffer in json_get_array to reduce parsing allocations.

Changes:

• Stores element byte ranges in a reusable Vec.
• Allows mutable direct-path callbacks.
• Removes smallvec and adds a wide-array benchmark.

File	Description
src/​json_get_array.rs	Implements reusable range buffering.
src/​common.rs	Accepts FnMut callbacks.
Cargo.toml	Removes smallvec.
benches/​main.rs	Adds a 32-element benchmark.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.