fix(text): handle code points > U+FFFF in levenshteinDistance

[lionel-rowe]

Fixes #6013

[lionel-rowe]

Benchmarks:

import { levenshteinDistance as current } from "jsr:@std/text@1.0.6/levenshtein-distance";
import { levenshteinDistance as next } from "https://raw.githubusercontent.com/lionel-rowe/std/levenshtein-unicode/text/levenshtein_distance.ts";

for (const [name, fn] of Object.entries({ current, next })) {
  Deno.bench(`${name} (ASCII)`, () => {
    fn("a".repeat(10), "b".repeat(10));
    fn("a".repeat(100), "b".repeat(100));

    fn("a".repeat(10), "");
    fn("a".repeat(100), "");
    fn("", "b".repeat(10));
    fn("", "b".repeat(100));

    fn(
      "a".repeat(100) + "b".repeat(100) + "a".repeat(100),
      "b".repeat(100) + "a".repeat(100) + "b".repeat(100),
    );
  });
}

for (const [name, fn] of Object.entries({ current, next })) {
  // will give wrong result with `current`, but just testing perf here
  Deno.bench(`${name} (with emoji)`, () => {
    fn(
      "a".repeat(100) + "💩".repeat(100) + "a".repeat(100),
      "💩".repeat(100) + "a".repeat(100) + "💩".repeat(100),
    );
  });
}

for (const len of [1e0, 1e1, 1e2, 1e3, 1e4]) {
  for (const [name, fn] of Object.entries({ current, next })) {
    Deno.bench(`${name} (string length ${len.toLocaleString("en-US")})`, () => {
      fn("a".repeat(len), "b".repeat(len));
    });
  }
}

Performance is almost identical on my machine, typical run:

benchmark                        time/iter (avg)        iter/s      (min … max)           p75      p99     p995
-------------------------------- ----------------------------- --------------------- --------------------------
current (ASCII)                          34.4 µs        29,030 ( 24.3 µs …   1.4 ms)  31.1 µs 128.5 µs 184.8 µs
next (ASCII)                             35.8 µs        27,930 ( 24.4 µs … 893.6 µs)  31.8 µs 134.5 µs 197.2 µs
current (with emoji)                     33.0 µs        30,280 ( 22.5 µs … 682.6 µs)  29.4 µs 129.3 µs 183.6 µs
next (with emoji)                        36.8 µs        27,200 ( 24.2 µs … 656.6 µs)  39.8 µs 110.0 µs 183.1 µs
current (string length 1)                67.9 ns    14,720,000 ( 49.7 ns … 194.9 ns)  75.8 ns 142.3 ns 155.8 ns
next (string length 1)                   69.1 ns    14,470,000 ( 52.6 ns … 206.7 ns)  73.9 ns 139.5 ns 161.2 ns
current (string length 10)              391.7 ns     2,553,000 (310.8 ns … 587.3 ns) 421.5 ns 580.7 ns 587.3 ns
next (string length 10)                 386.4 ns     2,588,000 (318.8 ns … 641.8 ns) 404.6 ns 610.2 ns 641.8 ns
current (string length 100)               6.0 µs       167,000 (  5.7 µs …   6.5 µs)   6.1 µs   6.5 µs   6.5 µs
next (string length 100)                  6.1 µs       163,500 (  5.6 µs …   7.2 µs)   6.3 µs   7.2 µs   7.2 µs
current (string length 1,000)           222.4 µs         4,496 (185.2 µs …   1.4 ms) 222.7 µs 397.6 µs 569.9 µs
next (string length 1,000)              212.3 µs         4,711 (174.1 µs …   1.1 ms) 214.1 µs 413.4 µs 525.8 µs
current (string length 10,000)           18.1 ms          55.3 ( 16.8 ms …  19.9 ms)  18.6 ms  19.9 ms  19.9 ms
next (string length 10,000)              17.4 ms          57.5 ( 16.1 ms …  19.0 ms)  17.7 ms  19.0 ms  19.0 ms

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Please upload report for BASE (main@a550998). Learn more about missing BASE report.
Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #6014   +/-   ##
=======================================
  Coverage        ?   96.29%           
=======================================
  Files           ?      494           
  Lines           ?    39541           
  Branches        ?     5837           
=======================================
  Hits            ?    38076           
  Misses          ?     1423           
  Partials        ?       42           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[lionel-rowe]

Previous fix was still buggy as it's tough to keep the various increments and length measurements in sync, given that sometimes you want the code point length and sometimes the code unit length. As a result I just switched to using [...str] char arrays, which surprisingly still has no measurable impact on perf (updated benchmarks above).

[kt3k]

LGTM. Thanks!