ci: narrow dorny diff to current push + drop oversized peer caches

Two issues compounded into bare-rn rebuilding on a CV-only push:

1. dorny defaulted to comparing against the PR base, so each push re-evaluated
   the cumulative PR diff and pulled every app whose filter ever matched into
   the matrix. Setting `base: event.before || pull_request.base.sha` restricts
   the diff to just the current push (with PR-base fallback for opened /
   reopened / ready_for_review).

2. The Gradle (~3 GB) and CocoaPods (~500 MB × 5 iOS apps) caches saved during
   a single run, plus setup-java's redundant gradle cache, pushed the repo over
   GitHub's 10 GB cache quota and evicted the tiny (~250 B) build markers under
   LRU. Cells appeared in the matrix on subsequent pushes (legitimately, per
   PR-base diff) but missed the cache and rebuilt. Dropping these peer caches
   so markers survive — markers gate entire build cells, which is worth more
   than the 3-10 min those caches saved on warm-up.

Author: msluszniak

.github/actions/build-android-app/action.yml

@@ -65,7 +65,6 @@ runs:
       with:
         distribution: "zulu"
         java-version: 17
-        cache: "gradle"
 
     - name: Install Expo CLI
       if: steps.cache.outputs.cache-hit != 'true' && inputs.expo-prebuild == 'true'
@@ -82,18 +81,10 @@ runs:
         rm -rf android
         npx expo prebuild --platform android --no-install
 
-    - name: Cache Gradle
-      if: steps.cache.outputs.cache-hit != 'true'
-      uses: actions/cache@v5
-      with:
-        path: |
-          ~/.gradle/caches
-          ~/.gradle/wrapper
-          ${{ inputs.app-path }}/android/.gradle
-        key: ${{ runner.os }}-gradle-${{ hashFiles('**/*.gradle*', '**/gradle-wrapper.properties') }}
-        restore-keys: |
-          ${{ runner.os }}-gradle-
-
+    # Note: not caching ~/.gradle or apps/<app>/android/.gradle on purpose.
+    # A full Gradle cache for this monorepo runs ~3 GB and competes with the
+    # 10 GB repo cache quota against the much smaller (~250 B) build markers.
+    # Markers gating entire build cells beat ~5-10 min of Gradle warm-up.
     - name: Build app
       if: steps.cache.outputs.cache-hit != 'true'
       working-directory: ${{ inputs.app-path }}/android

.github/actions/build-ios-app/action.yml

@@ -67,17 +67,10 @@ runs:
         rm -rf ios
         npx expo prebuild --platform ios --no-install
 
-    - name: Cache CocoaPods
-      if: steps.cache.outputs.cache-hit != 'true'
-      uses: actions/cache@v5
-      with:
-        path: |
-          ~/Library/Caches/CocoaPods
-          ${{ inputs.app-path }}/ios/Pods
-        key: ${{ runner.os }}-pods-${{ inputs.filter-name }}-${{ hashFiles(format('{0}/ios/Podfile.lock', inputs.app-path)) }}
-        restore-keys: |
-          ${{ runner.os }}-pods-${{ inputs.filter-name }}-
-
+    # Note: not caching CocoaPods on purpose. A populated Pods/ runs ~500 MB
+    # per app — ~2.5 GB across the 5 iOS cells — and competes with the 10 GB
+    # repo cache quota against the much smaller (~250 B) build markers.
+    # Markers gating entire build cells beat ~3-5 min of pod-install warm-up.
     - name: Install CocoaPods dependencies
       if: steps.cache.outputs.cache-hit != 'true'
       working-directory: ${{ inputs.app-path }}/ios

.github/workflows/build-apps.yml

@@ -30,6 +30,12 @@ jobs:
         if: github.event_name != 'workflow_dispatch'
         uses: dorny/paths-filter@v3
         with:
+          # Diff against the previous commit (synchronize/push) or the PR base
+          # (opened/reopened/ready_for_review). Without this, dorny defaults to
+          # the PR base on every push, so each commit re-evaluates the
+          # cumulative PR diff and pulls every app whose filter ever matched
+          # into the matrix — even when the latest push only touched one app.
+          base: ${{ github.event.before || github.event.pull_request.base.sha }}
           filters: |
             # Cross-platform shared infrastructure
             core-shared: &core-shared