[mvzink]

peeking_read_while appends to a caller's buffer and leaves the first non-matching character unread. Use it for word, numeric-prefix and dollar-value scans; peeking_take_while shares the same loop. This avoids temporary suffix strings and copies, complementing apache#2226.

Compared with the word-only change cc080fdd, the measured helper took 1.9% more tokenizer time with 8-character identifiers and 2.4% less with 128-character identifiers. The word-only change took 10.5%/9.2% less time than the original baseline; the helper was not measured directly against that baseline in the same batch.

These are medians of paired mean-time ratios from ten alternating pairs of the committed MySQL identifier benchmarks, with spans and output destruction included. Both comparisons used Linux on a Ryzen 5 3600, Rust 1.98.1, a fixed CPU and the performance governor. Within each comparison, both arms used identical locks and default features. Adapter throughput was not measured.

Original baseline: 1ea2488a632762e85f2cdb1c8bbe08f80800338e; word-only: cc080fdd3cccc379622370c13073b3fe98fdcfaf; measured helper: 3a150f27b5650585ca1b99a252870f29111b310f. Copy the added benchmark into the original baseline.

Related: apache#2036, apache#1557. AI-assisted.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}