Summary

• rewrite decode_mapping in src/decode.rs from split(';')/split(',') iteration to a cursor-based single-pass byte scanner
• remove per-segment heap allocation (Vec<i64>) and decode into a fixed [i64; 5] stack buffer
• preserve existing decode semantics and errors while reducing parsing overhead
• add malformed mapping coverage for bad segment size and VLQ leftovers

Key Changes

• delimiter-driven scanner over mapping.as_bytes()
• line and column state updates done inline while scanning
• new parse_vlq_segment_into(mapping, &mut cursor, &mut [i64; 5]) -> Result<usize> helper

Tests

• cargo fmt
• cargo check --all-targets --all-features
• cargo test
• cargo bench --bench simple -- parse/real_large --warm-up-time 0.1 --measurement-time 0.2 --sample-size 10

Benchmark Note

Quick parse benchmark for parse/real_large improved from ~531 MiB/s to ~644 MiB/s in this environment (~21% throughput increase).