fix(ci): bump model cache version to pick up lora_config.json

The LoRA intent classifier model was missing lora_config.json due to
timing issue during initial download. Bumping cache version from v1 to
v2 forces re-download of all models with the corrected configuration.

This fixes the 40% accuracy drop (empty predictions) by ensuring the
LoRA auto-detection logic finds lora_config.json and properly identifies
the model type.

Signed-off-by: Yossi Ovadia <[email protected]>

Author: yossiovadia

.github/workflows/test-and-build.yml

@@ -62,9 +62,9 @@ jobs:
         with:
           path: |
             models/
-          key: ${{ runner.os }}-models-v1-${{ hashFiles('tools/make/models.mk') }}
+          key: ${{ runner.os }}-models-v2-${{ hashFiles('tools/make/models.mk') }}
           restore-keys: |
-            ${{ runner.os }}-models-v1-
+            ${{ runner.os }}-models-v2-
         continue-on-error: true # Don't fail the job if caching fails
 
       - name: Check go mod tidy