fix(segment): propagate forward-store writer open failure instead of crashing

[mrcs64]

Problem

Opening and closing several collections concurrently in one process can crash. Under file-descriptor pressure, arrow::io::FileOutputStream::Open fails, so ChunkedFileWriter::Open returns nullptr. MemForwardStore::Open stores that null in writer_ but returns Status::OK() anyway, so the store looks initialized. Inserts proceed, and on close MemForwardStore::flush() dereferences the null writer_ at memory_forward_store.cc:283, crashing with SIGSEGV.

SegmentImpl::init_memory_components() compounds it: it assigns memory_store_ to the member before Open() succeeds. A failed init then leaves a non-null, half-built memory_store_ that satisfies the if (!memory_store_) re-init guard in internal_insert() and gets flushed on close.

Reproduction

Spawn N threads. Each creates and opens a collection in its own directory (one dense HNSW field plus an FTS field), upserts a few docs, runs a query, then drops it. Lowering ulimit -n induces the writer-open failure. This reliably SIGSEGVs, with the backtrace:

MemForwardStore::flush
SegmentImpl::flush
CollectionImpl::flush_unsafe
CollectionImpl::close_unsafe
CollectionImpl::~CollectionImpl

Single-threaded runs, and runs with a high fd limit, are unaffected.

Fix

• MemForwardStore::Open returns an error when the writer could not be created, instead of Status::OK().
• MemForwardStore::flush returns an error if writer_ is null, so close() cannot dereference it.
• SegmentImpl::init_memory_components rolls back partial state (memory_store_ and the vector indexers) on any failure, so a failed init leaves memory_store_ null and is never flushed on close.

Test

Adds a case to mem_store_test.cc that opens a MemForwardStore at a path whose parent directory does not exist. Before the fix, Open() returned OK with a null writer and later crashed; now it returns an error.

Related

Under much heavier concurrent-open load a separate crash appears inside RocksDB's filter registry (__hash_table::find("rocksdb.BuiltinBloomFilter")). That is independent of this change and is tracked in #580.

[CLAassistant]

All committers have signed the CLA.