feat(bench): add quality and throughput benchmark workspace

[jan-kubica]

Summary

Adds packages/bench, a private workspace that benchmarks the deterministic pipeline (NER off) over the contract fixture corpus, as groundwork for publishing reproducible performance and quality numbers and for comparing other anonymization tools on the same corpus.

Stacked on #193: the bench imports the built dist like a production consumer, which is how the non-Western corpus bundling regression fixed there was found; the first commit here is that fix.

What's included

• Scorer (src/scorer.ts): span-level, per-label precision/recall/F1 with one-to-one matching in two modes (exact bounds; label + largest overlap), plus unit tests.
• Quality runner (src/run-quality.ts): scores predictions against the reviewed .snapshot.json reference annotations, per label, per language, micro-averaged. --predictions file.json scores an external tool's output through the same scorer; the interchange format and label-mapping expectations are documented in the README.
• Throughput runner (src/run-throughput.ts): one-time costs (dictionary load, search preparation) separated from steady-state latency; warmup + measured passes with per-document medians and corpus chars/s.
• Renderer producing results/RESULTS.md from the JSON reports; results from a developer machine are committed alongside the methodology README.

Methodology note

The reference annotations derive from reviewed pipeline output, so the pipeline's own score against them is ~100% by construction; the README states this explicitly and frames the number as a regression/drift signal. The meaningful outputs are throughput and, next, cross-tool comparisons via the predictions interchange format.

Verification

bun run lint, bun run typecheck (6/6 tasks), bun run format:check, and bun test in packages/bench pass. check:version and packlist tooling are unaffected (explicit package lists; bench is private).

[gemini-code-assist]

Code Review

This pull request introduces a new benchmarking package, @stll/anonymize-bench, to measure the quality and throughput of the anonymization pipeline. It also refactors non-Western name imports in the main package to use literal import specifiers, ensuring correct bundler resolution. The reviewer identified three key areas for improvement: filtering readdirSync in fixtures.ts to avoid crashing on non-directory entries, deduplicating city names in dictionaries.ts using a Set for better memory efficiency, and adding a guard in run-throughput.ts to prevent a potential division-by-zero error when calculating characters per second.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Using readdirSync directly on CONTRACTS_DIR will return all directory entries, including hidden files (such as .DS_Store) or any non-directory files. Attempting to call readdirSync on these files in the next line will throw an ENOTDIR error and crash the benchmark. Filtering for directories only using withFileTypes: true prevents this issue.

  const languages = readdirSync(CONTRACTS_DIR, { withFileTypes: true })
    .filter((dirent) => dirent.isDirectory())
    .map((dirent) => dirent.name)
    .toSorted();
  for (const language of languages) {
    const languageDir = join(CONTRACTS_DIR, language);

[gemini-code-assist]

The mergedCities array is populated by pushing every city from every country, which can result in duplicate city names (e.g., "London" in both GB and US/CA). Deduplicating the list using a Set avoids redundant entries in the search automaton and improves memory efficiency.

Suggested change

	const mergedCities: string[] = [];
	for (const { country, entries } of cityResults) {
	citiesByCountry[country] = entries;
	for (const entry of entries) {
	mergedCities.push(entry);
	}
	}
	const mergedCitiesSet = new Set<string>();
	for (const { country, entries } of cityResults) {
	citiesByCountry[country] = entries;
	for (const entry of entries) {
	mergedCitiesSet.add(entry);
	}
	}
	const mergedCities = Array.from(mergedCitiesSet);

[gemini-code-assist]

If a document is processed extremely quickly or due to timer resolution limits in virtualized environments, medianMs can be 0. This would result in a division-by-zero error, making charsPerSecond evaluate to Infinity. Adding a guard prevents this.

    charsPerSecond: medianMs === 0 ? 0 : Math.round(doc.text.length / (medianMs / 1_000)),