Fix SnapshotReader temp-file collision when concurrent readers open the same compressed snapshot

Summary:
For compressed `.pytb` inputs, `SnapshotReader::open()` decompresses into a temp file and mmaps it `MAP_SHARED` with `PROT_READ | PROT_WRITE` so the rest of the reader can resolve offsets through `data_`. The temp path was deterministic — `<path>.decompressed.tmp` — which means two `SnapshotReader::open(path)` calls against the same compressed file (commonly a Python process plus a sibling `tintype_debug_launcher` subprocess attached via dapper) collide on that path. The second open is `O_RDWR | O_CREAT | O_TRUNC` followed by `ftruncate(decompressedSize)`, which through the shared inode zero-fills the first reader's mmap. `getManifest()` then returns `decompressedSize` bytes of `\0` until the second `ZSTD_decompress` finishes repopulating the file (~tens of milliseconds). In our case `extract_revision()` saw the null buffer, raised `json.JSONDecodeError: Expecting value: line 1 column 1 (char 0)`, and the deepdump python postmortem bootstrapper aborted before the agent ever ran.

This diff switches the compressed-input path to `mkstemp(<path>.decompressed.XXXXXX)`, mirroring the pattern already used by `SnapshotWriter::open` (`mkstemp("/tmp/snapshot_XXXXXX")`). Each reader gets its own temp file, so concurrent opens cannot race on the same inode. `close()` already unlinks `tempFilePath_`, so cleanup is unchanged. The mode goes from `0644` to `0600` (mkstemp default), which is stricter and consistent with the writer. Adds `<cstdlib>` for `mkstemp` and `<vector>` for the writable template buffer.

Reviewed By: aperez

Differential Revision: D103755545

fbshipit-source-id: 64e44529801ac82d986c764ed119c6f8786b37a1

Author: bvu405

snapshot_lib/SnapshotReader.cpp

@@ -15,9 +15,11 @@
 
 #include <algorithm>
 #include <cerrno>
+#include <cstdlib>
 #include <cstring>
 #include <fstream>
 #include <stdexcept>
+#include <vector>
 
 namespace facebook::tintype::snapshot {
 
@@ -148,14 +150,22 @@ bool SnapshotReader::open(const std::string& path) {
       return false;
     }
 
-    // Create a temporary file for the decompressed data
-    tempFilePath_ = path + ".decompressed.tmp";
-    tempFd_ = ::open(tempFilePath_.c_str(), O_RDWR | O_CREAT | O_TRUNC, 0644);
+    // Create a unique temporary file for the decompressed data.
+    // mkstemp guarantees a fresh path so concurrent readers of the same
+    // compressed source file (e.g., a sibling process attaching
+    // tintype_debug_launcher to the same .pytb) cannot collide on this temp
+    // file. With a shared name, one reader's O_TRUNC + ftruncate corrupts
+    // another reader's MAP_SHARED mmap (zero-fills the buffer through the
+    // shared inode) — see SnapshotWriter::open which already uses mkstemp
+    // for the same reason.
+    std::string tempTemplate = path + ".decompressed.XXXXXX";
+    tempFd_ = ::mkstemp(tempTemplate.data());
     if (tempFd_ < 0) {
-      lastError_ = "Failed to create temporary file: " + tempFilePath_ +
+      lastError_ = "Failed to create temporary file: " + tempTemplate +
           " (errno: " + std::to_string(errno) + ")";
       return false;
     }
+    tempFilePath_ = std::move(tempTemplate);
 
     // Extend the file to the decompressed size
     if (ftruncate(tempFd_, decompressedSize) < 0) {