fix: drop stale Ollama tags, and make add_models.py the catalog's only source

.githooks/pre-commit

@@ -1,6 +1,6 @@
 #!/bin/sh
 #
-# Runs the same three checks as .github/workflows/ci.yml before a commit is
+# Runs the same four checks as .github/workflows/ci.yml before a commit is
 # created, so a red X on GitHub is caught here instead of arriving as mail
 # minutes later. This is the whole reason the hook exists: every failed CI run
 # in this repository's history was a commit that would not have passed
@@ -72,4 +72,22 @@ run "Clippy" "yukaridaki uyarilari duzelt" \
     cargo clippy --all-targets --all-features -- -D warnings
 run "Test" "cargo test --all-features" cargo test --all-features
 
+# data/models.json is generated from scripts/add_models.py. A hand edit to the
+# JSON is overwritten by the next run of the script, so it is stopped here.
+# `python3` on Windows can be the Store stub that exits with an error, so each
+# candidate is tried rather than just looked up.
+py=""
+for candidate in python3 python; do
+    if "$candidate" -c "" >/dev/null 2>&1; then
+        py=$candidate
+        break
+    fi
+done
+if [ -n "$py" ]; then
+    run "Catalog" "degisikligi scripts/add_models.py'ye tasi ve calistir" \
+        "$py" scripts/add_models.py --check
+else
+    echo "pre-commit: python bulunamadi, katalog kontrolu atlaniyor" >&2
+fi
+
 echo "pre-commit: hepsi gecti"

.github/workflows/ci.yml

@@ -51,6 +51,21 @@ jobs:
       - name: Clippy
         run: cargo clippy --all-targets --all-features -- -D warnings
 
+  catalog:
+    name: Catalog
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      # data/models.json is generated from scripts/add_models.py; a hand edit
+      # to the JSON would be lost on the script's next run.
+      - name: Check the catalog matches its generator
+        run: python scripts/add_models.py --check
+
   fmt:
     name: Format check
     runs-on: ubuntu-latest

README.md

@@ -587,8 +587,9 @@ Databricks · Tencent · Baidu · ByteDance · LG AI · Allen AI · TII ·
 Hugging Face · OpenBMB · Liquid AI · Arcee · Nous Research · Perplexity ·
 Shanghai AI Lab · BAAI · Nomic · Jina · and more.
 
-The catalog is generated by `scripts/add_models.py`, which merges entries by
-`id` so re-running it corrects a record rather than duplicating it.
+The catalog is generated by `scripts/add_models.py`, the only place a record is
+written. CI runs it with `--check`, so a hand edit to `data/models.json` fails
+the build instead of being lost on the script's next run.
 
 ---
 

docs/ROADMAP.md

@@ -109,6 +109,11 @@ image would each need release automation and signing.
    runtime tags, partial geometry and implausible values
 4. Rebuild; the catalog is embedded at compile time
 
+Correcting or removing a record works the same way: edit or delete its entry in
+the script and run it. Never edit `data/models.json` by hand. The script
+rewrites every record from its own entries, and `--check` (run by CI and the
+pre-commit hook) fails while the two disagree.
+
 Geometry (`layers`, `hidden_size`, `kv_heads`, `head_dim`) is optional but
 worth having: with it the KV cache is exact, without it llmspec falls back to
 a heuristic derived from one model's proportions. Supply all four fields or