Zero-copy SlicedString search (indexOf/startsWith/endsWith/includes)

[lahma]

Summary

SlicedString is a zero-copy view over a backing string (produced by slice/substring/substr for large results). It already overrides Equals/GetHashCode/AsSpan for zero-copy, but the search methods IndexOf/StartsWith/EndsWith/Contains were inherited from the base JsString, which routes through ToString() — materializing (allocating) the whole substring on every search. A fresh large slice that is searched and then discarded therefore allocated its full content on each indexOf/startsWith/endsWith/includes call.

This makes the four base search methods virtual and overrides them in SlicedString to search directly over AsSpan(). Ordinal char comparison is binary/sequence equality, so the parameterless span overloads produce results identical to the base StringComparison.Ordinal paths (verified across all target frameworks).

Benchmark

Jint.Benchmark/StringSliceBenchmarks.cs — SearchOnSlice creates a fresh large slice and searches it each iteration (the shape that previously materialized on every call). Default BDN job, .NET 10, Ryzen 9 5950X, same-runtime worktree A/B:

Benchmark	Baseline	This PR	Δ
SearchOnSlice — time	378.2 ms	27.5 ms	13.7× faster
SearchOnSlice — alloc	1,163,559 KB	395 KB	−99.97%
SearchOnFlat (guard) — time	54.9 ms	50.6 ms	within noise
SearchOnFlat (guard) — alloc	161 KB	161 KB	unchanged
SliceLargeDiscard / SliceThenRead	no regression

SearchOnFlat is a dispatch-bound guard searching a plain (non-view) JsString; making the methods virtual is below the noise floor (dynamic PGO devirtualizes the monomorphic call sites).

Correctness

• New differential test StringTests.SlicedStringSearchMatchesFlatString compares sliced-view search against an identical flat string across many needles and positions, both before and after the view materializes.
• dotnet test Jint.Tests: 3138 passed, 0 failed.
• Test262 (strict + sloppy): 99260 passed, 0 failed.

🤖 Generated with Claude Code