[wufangyong973]

Closes #39.

Fossel has a strong measure retrieval changes before merging culture, but bench/README.md stopped at describing what the benchmark does and never explained the contribution side — how to add a new query/expected-result pair to the eval set and interpret the result.

This adds a single Adding a benchmark case section, placed between the existing Datasets and Tuning sweeps sections, that covers:

• the shape of a dataset entry — memories[] ({key, note, type, tags}) and queries[] ({id, query, relevant, category, rationale}) in bench/datasets/repo-memory.json, cross-referenced by key
• the four-step add-a-case flow: add a memory if needed, add a query, re-run npm run bench to refresh the committed snapshot under bench/results/, then verify with npm run bench:check
• how to read the output (the search vs context surfaces, byCategory, hit@k / MRR, and the per-mode miss list as the design tool)
• the merge rule: keep a retrieval change only when the numbers support it

Everything is grounded in files that already exist — the committed dataset, bench/run.ts (the --check / --no-write snapshot logic) and the current README.md. Pure documentation, no harness change. A contributor should now be able to add a case and interpret the result without reading the harness source.

[vercel]

@wufangyong973 is attempting to deploy a commit to the vignesh's projects Team on Vercel.

A member of the Team first needs to authorize it.