README: switch download links to .json.gz (auto-download UX)

README.md

@@ -28,7 +28,7 @@ FumingY/MergeFree/
 ├── README.md            ← this file
 └── bundles/
     ├── README.md        ← bundle schema + usage details
-    └── <root_id>_hybrid.json    ← one per neuron
+    └── <root_id>_hybrid.json.gz    ← one per neuron
 ```
 
 ## Direct download URL pattern
@@ -37,31 +37,40 @@ Each JSON has a stable `resolve/main/...` URL; suitable for
 `=HYPERLINK()` in Google Sheets:
 
 ```
-https://huggingface.co/datasets/FumingY/MergeFree/resolve/main/bundles/<ROOT_ID>_hybrid.json
+https://huggingface.co/datasets/FumingY/MergeFree/resolve/main/bundles/<ROOT_ID>_hybrid.json.gz
 ```
 
 Spreadsheet column formula (column A = root id, column B = link):
 
 ```
 =HYPERLINK(
-  "https://huggingface.co/datasets/FumingY/MergeFree/resolve/main/bundles/" & A2 & "_hybrid.json",
+  "https://huggingface.co/datasets/FumingY/MergeFree/resolve/main/bundles/" & A2 & "_hybrid.json.gz",
   "download"
 )
 ```
 
 ## How to review
 
-1. Open https://fumingyang-felix.github.io/MERGE_FREE_DEMO/ (or clone +
-   serve `MERGE_FREE_DEMO/index.html` locally).
-2. Click **↑ Import bundle**, point to one of these JSONs.
-3. Step through each suspect window (`→` / `Enter`); for each one
+1. Click the download link → browser saves `<root>_hybrid.json.gz`
+   to your Downloads folder.
+2. Open https://fumingyang-felix.github.io/MERGE_FREE_DEMO/ (or
+   clone + serve `MERGE_FREE_DEMO/index.html` locally).
+3. Click **↑ Import bundle** and select the downloaded `.json.gz`
+   file — the static site decompresses it in-browser.
+4. Step through each suspect window (`→` / `Enter`); for each one
    answer:
    - **Real merge?** — yes / no / skip / unsure
    - **SPLIT WHICH?** — multi-select the colored token clusters that
      should actually be split off
-4. Click **↓ Export decisions** when done — that JSON is the
+5. Click **↓ Export decisions** when done — that JSON is the
    reviewer's output.
 
+**Why `.json.gz` and not `.json`?** Hugging Face serves small JSON
+files with `Content-Disposition: inline`, so plain `.json` links
+render in-browser instead of triggering a download.  Browsers
+always download `application/gzip`, and the static site handles the
+decompression for you.
+
 ## Generation
 
 Bundles are produced in batch by
@@ -75,9 +84,10 @@ repo) running the same pipeline as the streamlit UI:
 5. cuts via the **hybrid** partitioner (wsbm if its DL ≤ 100, else
    spectral+)
 
-The `_hybrid` filename suffix tags the partition method used for the
-cuts.  Future spectral_plus / wsbm-only variants would land alongside
-as `<root>_spectral.json` / `<root>_wsbm.json` in the same folder.
+The `_hybrid` filename suffix tags the partition method used for
+the cuts.  Future spectral_plus / wsbm-only variants would land
+alongside as `<root>_spectral.json.gz` / `<root>_wsbm.json.gz` in
+the same folder.
 
 ## Schema