perf: optimize strategy selection and Teddy 2-byte fingerprint

[kolkov]

Summary

• Implement Teddy 2-byte fingerprint (reduces false positives by ~90%)
• Reorder strategy selection to prioritize DigitPrefilter for digit-lead patterns
• Add isDigitLeadPattern() helper for pattern classification

Performance Impact

Pattern	Before	After	Change
literal_alt	31ms	8ms	+4x faster
version	8.2ms	2ms	+4x faster
IP	3.9ms	5.5ms	-43% (trade-off)

Trade-off Analysis

The IP pattern regression is an acceptable trade-off:

• We remain 2.2x faster than Rust regex on IP patterns (Rust: ~12ms)
• The gains on literal_alt and version patterns significantly outweigh the IP regression
• See perf: Research IP pattern optimization strategies #62 for future IP optimization research

Technical Details

Teddy 2-byte Fingerprint

Changed default from 1-byte to 2-byte fingerprint in prefilter/teddy.go:

• 1-byte: ~25% false positive rate on typical text
• 2-byte: <0.5% false positive rate

Strategy Reorder

Moved DigitPrefilter check before tiny NFA fallback in meta/strategy.go:

• Ensures digit-lead patterns use specialized prefilter
• Prevents single-byte inner literals (like .) for digit patterns

Test Plan

• All existing tests pass
• Pre-release check passes
• Benchmarks validated locally
• Trade-off documented in perf: Research IP pattern optimization strategies #62

[github-actions]

Benchmark Comparison

Comparing main → PR #61

Summary: geomean 173.0n 172.7n -0.17%

⚠️ Potential regressions detected:

geomean                               ³                +0.00%               ³
geomean                               ³                +0.00%               ³
geomean                         ³                +0.00%               ³
geomean                         ³                +0.00%               ³
AhoCorasickVsStdlib/coregex_IsMatch-4                   1.249µ ± ∞ ¹    1.252µ ± ∞ ¹   +0.24% (p=0.032 n=5)
IPRegex_Find/coregex_1KB_sparse-4                       2.307µ ± ∞ ¹    4.298µ ± ∞ ¹  +86.30% (p=0.008 n=5)
IPRegex_Find/stdlib_1MB_sparse-4                        1.208m ± ∞ ¹    2.155m ± ∞ ¹  +78.45% (p=0.008 n=5)
IPRegex_Find/stdlib_1MB_dense-4                         8.084µ ± ∞ ¹   15.050µ ± ∞ ¹  +86.17% (p=0.008 n=5)
Find/hello-4                                            531.2n ± ∞ ¹    535.6n ± ∞ ¹   +0.83% (p=0.032 n=5)
IsMatch/literal-4                                       57.00n ± ∞ ¹    58.54n ± ∞ ¹   +2.70% (p=0.008 n=5)

Full results available in workflow artifacts. CI runners have ~10-20% variance.
For accurate benchmarks, run locally: ./scripts/bench.sh --compare