leaderboard: docs cleanup — stale CLAUDE.md path, LFS contributor note

Two small fixes surfaced while auditing the leaderboard workflow docs:

- leaderboard/CLAUDE.md referenced data/results.json, which has not
  existed since PR #35 split data files (the curated entries live in
  data/leaderboard.json now).

- leaderboard/CONTRIBUTING.md had no mention of git-lfs, even though
  data/{leaderboard,extractions,scan_results}.json are LFS-tracked
  per data/.gitattributes. A first-time contributor without `git lfs
  install` reads pointer files and trips validate.py / build scripts.
  CI uses lfs: true on checkout (added in #46) but local setup was
  undocumented. Added a "Local Setup" section with the one-time
  install/pull commands.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: MilkClouds

leaderboard/CLAUDE.md

@@ -4,7 +4,7 @@ This directory is a **standalone static leaderboard site** — it is NOT vla-eva
 
 ## Key distinction
 
-- `leaderboard/data/results.json` contains scores **reported in published papers**. These are manually curated, not produced by running `vla-eval`.
+- `leaderboard/data/leaderboard.json` contains scores **reported in published papers**. These are manually curated or produced by the `extract.py` / `refine.py` pipeline — not produced by running `vla-eval`.
 - `vla-eval` is the evaluation harness (the parent repo). Its runtime outputs go to `results/` or user-specified paths, never here.
 
 Do not conflate leaderboard data with vla-eval reproduced results. They are independent.

leaderboard/CONTRIBUTING.md

@@ -2,6 +2,17 @@
 
 > **Note on evaluation protocols:** Benchmark evaluation protocols are not fully standardized across the VLA community. Different papers may use the same benchmark name but differ in training regimes, task subsets, or evaluation conditions — making scores not always directly comparable. This leaderboard records all available results transparently and documents known protocol differences, but gaps remain. We actively welcome contributions: score corrections, missing results, protocol clarifications, and proposals for standardization.
 
+## Local Setup
+
+`leaderboard/data/{leaderboard,extractions,scan_results}.json` are stored in **Git LFS** (see `leaderboard/data/.gitattributes`). Without LFS smudging, scripts will read pointer files and fail. Once per machine:
+
+```
+git lfs install                 # install the LFS hooks
+git lfs pull                    # smudge any pointer files in the current checkout
+```
+
+CI workflows (`pages.yml`, `update-data.yml`, `leaderboard-validate.yml`) already pass `lfs: true` to `actions/checkout`.
+
 ## Data Structure
 
 Data is split into focused files under `leaderboard/data/`: