fix(cache): make Zip-source caches atomic and robust to partial state

Follow-up to the GitHubStream cache fixes. The same poisoned-cache
class existed in the GitHubDownload path and a few related spots:

- GitHubDownload.download: wipe pre-existing state before extracting
  and write a .anon-complete marker only after a successful extract.
  On error, rm the partial cache so a retry starts clean. getFileContent
  and getFiles now gate on the marker instead of "any file/folder
  exists," so a half-extracted tree can never be served as canonical.
- GitHubDownload.getFileContent: validate cached file size against the
  upstream FileModel size (via the new AnonymizedFile.size()), same
  guard as GitHubStream. getFiles filters the marker from the listing.
- FileSystem.listFiles: drop the bogus stats.ino.toString() as sha.
  An inode isn't a content hash; anything comparing it to a Git blob
  sha would silently disagree. Leave undefined.
- S3.write: remove the fire-and-forget data.on("error") -> this.rm(...).
  Multipart Upload doesn't commit partial objects, so there was nothing
  to clean up, and the handler raced retries and could delete a
  previously-good object on a transient source-stream hiccup. The
  size-validated read path recovers from any other undersized objects.
- GitHubStream.resolveLfsPointer: drop the post-decision early-return
  in blobStream.on("error"). Currently redundant with the inner
  listener, but removes the future-refactor footgun.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/core/storage/FileSystem.ts

@@ -153,13 +153,17 @@ export default class FileSystem extends StorageBase {
               size: stats.size,
             });
           }
+          // Don't synthesise a sha here. The previous value (stats.ino)
+          // wasn't a content hash — just an inode number — and any code
+          // that compared it to an upstream Git blob sha would silently
+          // disagree. Leave it undefined so callers either look up the
+          // real sha from FileModel/GitHub or skip sha-keyed paths.
           output2.push(
             new FileModel({
               name: file,
               path: dir,
               repoId: repoId,
               size: stats.size,
-              sha: stats.ino.toString(),
             })
           );
         }

src/core/storage/S3.ts

@@ -175,12 +175,13 @@ export default class S3Storage extends StorageBase {
   ): Promise<void> {
     if (!config.S3_BUCKET) throw new Error("S3_BUCKET not set");
 
-    if (data instanceof Readable) {
-      data.on("error", (err) => {
-        console.error(`[ERROR] S3 write ${path}`, err);
-        this.rm(repoId, path);
-      });
-    }
+    // No fire-and-forget rm on stream error: Upload uses multipart and
+    // does not commit a partially-uploaded object, so there's nothing to
+    // clean up. The previous handler raced with retries and could delete
+    // a previously-good object on a transient source-stream hiccup. The
+    // size-validated read path in GitHubStream.getFileContentCache
+    // recovers from any object that does end up undersized for any
+    // reason.
 
     const params: PutObjectCommandInput = {
       Bucket: config.S3_BUCKET,