[wufangyong973]

Closes #39.

bench/README.md explained what the benchmark does but not how to add a
case, even though the project's standing rule is to measure retrieval
changes before merging and to keep a change only if the numbers support
it.

Adds an "Adding a benchmark case" section covering:

• the shape of the committed eval set (bench/datasets/repo-memory.json),
  with concrete examples of a memory entry and a query entry, plus the
  recency-order and no-write-path-inference constraints that keep the
  numbers honest;
• how to run npm run bench (with --verbose to see every miss) and which
  of the two reported tables is the retrieval-quality number;
• how npm run bench:check works (fails on drift, does not write) and
  when to update the committed snapshot in bench/results/.

Docs-only, 92 lines added to one file.

[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.

[7vignesh]

Superseded by your revised version in #55 - the more concise rewrite is the one we are taking. Closing as superseded, not rejected. Thanks for tightening it up. (#55 closes #39.)