perf: fast-path long identical partial ratio inputs

[arindamsikder]

Summary

• Return the full-length, 100-point alignment immediately for identical inputs longer than 64 characters.
• Apply the shortcut in the shared implementation used by direct, alignment and cached scoring.
• Cover empty/short inputs, the 64/65 boundary, 50,000-character separate buffers, iterator overloads, wide strings and cutoff handling.

Problem

The example in rapidfuzz/RapidFuzz#430 compares "hello" * 10000 with itself. The current C++ implementation evaluates shorter prefixes before reaching the full-length match. On this Linux/GCC 15.2 machine, the original direct, cached and alignment calls each exceeded a 15-second probe limit.

Solution

Check exact equality only when the lengths match and exceed 64. Keep the existing prefix/suffix search and tie ordering for nonidentical inputs; the public empty-input and above-100 cutoff guards remain in effect.

Performance tradeoff: this is not a universal speedup. Five alternating base/patch benchmark runs of 200,000 comparisons of 32-character strings differing at the first or last character showed medians of 1.07232 → 1.12651 seconds (+5.05%) and 1.09900 → 1.16506 seconds (+6.01%). The 50,000-character example completed in 0.00154 / 0.00166 / 0.00153 seconds for direct/cached/alignment, including scorer construction. These are single-machine measurements, not cross-platform guarantees.

Testing

All execution used a network-disabled Bubblewrap sandbox with no account credentials or home directories mounted. C++11, GCC 15.2.0, CMake 3.31.6, Catch2 2.13.10:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release   -DCMAKE_CXX_STANDARD=11 -DCMAKE_CXX_STANDARD_REQUIRED=ON   -DCMAKE_CXX_EXTENSIONS=OFF -DRAPIDFUZZ_BUILD_TESTING=ON   -DFETCHCONTENT_SOURCE_DIR_CATCH2=/catch2 -DFETCHCONTENT_FULLY_DISCONNECTED=ON
cmake --build build -j 2
ctest --test-dir build -C Release --output-on-failure -j 2
build/test/test_fuzz PartialRatioIdenticalInputs --reporter compact
build/test/test_fuzz --reporter compact

• Full suite: 10/10 passed. Focused regression: 329 assertions; affected fuzz suite: 624 assertions.
• New regression compiled against old implementation: timed out at 8 seconds; final implementation passed within the same cap. Repeated after final test expansion.
• AddressSanitizer + UndefinedBehaviorSanitizer Debug affected suite: 624 assertions passed (-fsanitize=address,undefined -fno-omit-frame-pointer).
• RAPIDFUZZ_ENABLE_LINTERS=ON with pinned project_options v0.26.2: affected test_fuzz builds and passes. The all-target strict build fails on five -Werror=null-dereference diagnostics at unchanged Levenshtein_impl.hpp:1124; the same target on pristine base produces the identical diagnostics. No warning flags were weakened.
• External deterministic comparison: 9,246 output rows byte-identical on base and patch, including scores, alignment indices, cached results, cutoff boundaries, short binary-string corpus, 63/64/65/128 lengths, early/late differences and mixed wide/integer values.
• git diff --check, added-line security scan and independent AI diff review passed. Standalone clang-format and other compiler/platform matrices were not available locally. Catch2 emits its existing CMake deprecation warning.

Related Issue

Related to rapidfuzz/RapidFuzz#430. This addresses the identical-input C++ case only, not general worst-case complexity, Python interrupt handling, or a downstream Python dependency update/release.

Implemented and validated with Hermes AI automation; the independent local review was also AI-based, not a human maintainer approval.