fix(download-libs): rebuild against refreshed artifacts instead of reusing stale objects

.github/workflows/ci.yml

@@ -89,6 +89,34 @@ jobs:
       - name: Provision third-party headers
         run: RNET_HEADERS_ONLY=1 node scripts/download-libs.js
 
+      # third-party/include and the ExecuTorch built below must come from the
+      # same release. When they drifted, `make_tensor_ptr` gaining a trailing
+      # `Device` parameter surfaced as an undefined reference nine minutes into
+      # the dependency build, blamed on whichever code happened to call it.
+      # Comparing the two up front fails in seconds and names both values.
+      # Note this only spans releases: ET_VERSION comes from the source tree's
+      # version.txt, so it cannot tell the fork's patches apart from upstream,
+      # and it says nothing about TOKENIZERS_COMMIT.
+      - name: Check the header and test-dependency ExecuTorch versions agree
+        run: |
+          headers=$(sed -n 's/^#define ET_VERSION "\(.*\)"$/\1/p' \
+            third-party/include/executorch/runtime/core/version.h)
+          pinned=$(sed -n 's/^EXECUTORCH_VERSION="v\(.*\)"$/\1/p' \
+            scripts/build-native-test-deps.sh)
+          if [ -z "$headers" ] || [ -z "$pinned" ]; then
+            echo "::error::could not read the ExecuTorch versions" \
+                 "(headers='$headers' pinned='$pinned')"
+            exit 1
+          fi
+          if [ "$headers" != "$pinned" ]; then
+            echo "::error::third-party/include is ExecuTorch $headers but" \
+                 "scripts/build-native-test-deps.sh pins v$pinned." \
+                 "Bump EXECUTORCH_VERSION (and TOKENIZERS_COMMIT) to match" \
+                 "the artifacts release."
+            exit 1
+          fi
+          echo "ExecuTorch $headers on both sides"
+
       # Hermes and ExecuTorch are pinned to exact tags, so the cache only misses
       # when scripts/build-native-test-deps.sh changes those pins.
       #

packages/react-native-executorch/scripts/build-native-test-deps.sh

@@ -32,11 +32,11 @@ HERMES_REPO="https://github.com/facebook/hermes.git"
 # errors or, worse, ABI drift at runtime. cpp/extensions/llm additionally reads
 # private members of TextLLMRunner/MultimodalRunner, so a version skew there
 # fails to compile rather than silently misbehaving.
-EXECUTORCH_VERSION="v1.3.1"
+EXECUTORCH_VERSION="v1.4.1"
 EXECUTORCH_REPO="https://github.com/pytorch/executorch.git"
 
 # The shipped native libraries are built from software-mansion-labs/executorch
-# @rne-split-build, which is ExecuTorch 1.3.1 with the tokenizers submodule
+# @ms/separate-backends-1.4.1, which is ExecuTorch 1.4.1 with the tokenizers submodule
 # swapped for the fork below (it adds the WordPiece/Unigram models and the NFC
 # normalizer that upstream has not taken). third-party/include carries that
 # fork's headers, so linking upstream's libtokenizers.a here would compile
@@ -47,7 +47,7 @@ EXECUTORCH_REPO="https://github.com/pytorch/executorch.git"
 # Keep this commit in sync with the tokenizers submodule of the fork commit that
 # produced the current headers.tar.gz.
 TOKENIZERS_REPO="https://github.com/software-mansion-labs/pytorch-tokenizers.git"
-TOKENIZERS_COMMIT="56a30afbe2e6b4ca881d0fb7b961b9f9da156be4"
+TOKENIZERS_COMMIT="a03231a20a72036bf9a8e4a3b1d63494b90da1a6"
 
 cd "$(dirname "$0")/.."
 PACKAGE_DIR="$(pwd)"

packages/react-native-executorch/scripts/download-libs.js

@@ -390,17 +390,36 @@
   return result.toString().split(' ')[0].trim();
 }
 
-function isCacheValid(artifact) {
+async function isCacheValid(artifact) {
   if (!fs.existsSync(artifact.cacheFile)) return false;
-  if (!fs.existsSync(artifact.cacheChecksumFile)) return false;
+  // Refresh the checksum from the release before trusting the cache. Comparing
+  // a cached tarball against a CACHED checksum lets a stale cache validate
+  // itself: the cache directory is keyed on the libs version, so a release
+  // re-cut at the same version is never picked up -- every artifact reports a
+  // cache hit and the old files are reused indefinitely.
+  try {
+    await download(artifact.checksumUrl, artifact.cacheChecksumFile);
+  } catch {
+    // Unreachable checksum (offline, rate limited): fall back to the cached
+    // one so an already-populated cache still builds without a network.
+    if (!fs.existsSync(artifact.cacheChecksumFile)) return false;
+  }
   const expectedChecksum = fs.readFileSync(artifact.cacheChecksumFile, 'utf8').trim();
   const actualChecksum = sha256(artifact.cacheFile);
   return expectedChecksum === actualChecksum;
 }
 
 function extract(tarball, destDir) {
   ensureDir(destDir);
-  execSync(`tar -xzf "${tarball}" -C "${destDir}"`);
+  // `-m` stamps extracted files with the extraction time instead of the mtime
+  // recorded in the archive. Without it, headers keep the timestamp they had
+  // when the release was cut, so upgrading to a NEWER artifacts release can
+  // hand ninja headers that look OLDER than object files from a previous build.
+  // Ninja then treats those objects as up to date and never recompiles them,
+  // and they get archived and linked against the new libraries -- which shows
+  // up as an undefined symbol for whatever API changed between the two
+  // releases, far away from the actual cause.
+  execSync(`tar -xzmf "${tarball}" -C "${destDir}"`);
 }
 
 // ---- Main ------------------------------------------------------------------
@@ -432,7 +451,7 @@
   for (const artifact of artifacts) {
     console.log(`[react-native-executorch] Preparing ${artifact.name}...`);
 
-    if (isCacheValid(artifact)) {
+    if (await isCacheValid(artifact)) {
       console.log(`  ✓ Cache hit, skipping download`);
     } else {
       console.log(`  ↓ Downloading ${artifact.url}`);