fix(httpfs): initialize HTTP client on size-cache hit to avoid SIGSEGV on repeat opens

[adsharma]

Fixes the crash reported in LadybugDB/ladybug#880.

Problem

On 0.20.x, every remote (httpfs) read over http(s):// segfaults at bind time on the first ranged GET; 0.19.1 and earlier are clean. Local reads of the same files are fine, and URLs that 404/403 raise a clean RuntimeError instead of crashing.

Backtrace (matches the one in #880):

#0 std::_Function_handler<httplib::Result (), HTTPFileSystem::getRangeRequest(...)::{lambda()#1}>::_M_invoke(...)
#1 HTTPFileSystem::runRequestWithRetry(...)
#2 HTTPFileSystem::getRangeRequest(...)
#3 HTTPFileSystem::readFromFile(...)
...
#15 binder::Binder::bind(...)

Root cause

05c63dd ("perf(httpfs): pool HTTP connections and cache remote file sizes") added a process-wide cache of remote file sizes so openFile() skips the HEAD round trip on repeat opens of the same URL. The fast-path early return in HTTPFileInfo::initMetadata() was placed before initializeClient():

{
    std::lock_guard<std::mutex> lck{HTTPFileSystem::httpSizeCacheMtx};
    auto it = HTTPFileSystem::httpSizeCache.find(path);
    if (it != HTTPFileSystem::httpSizeCache.end()) {
        length = it->second;
        return;              // ← skips initializeClient()
    }
}
...
initializeClient();

httpClient is only assigned in initializeClient(), so the second openFile() of the same URL returns a handle whose httpClient is null. Every scan opens the same remote URL at least twice (bind + execution), and the first ranged GET on the second handle dereferences the null client — hence the SIGSEGV on all remote reads.

This also explains the reporter's observations:

• local parquet reads are unaffected (different file system),
• 404/403 URLs raise a clean error because they fail during the first open's HEAD (which does initialize the client),
• 0.19.1 (pre-05c63dd) is clean.

Reproduced locally with a python -m http.server file server and LOAD FROM 'http://127.0.0.1:8123/user.parquet' RETURN count(*): deterministic SIGSEGV (exit 139), with the crash point being mov (%r13),%rsi where r13 = 0 (the null client).

Fix

1. HTTPFileInfo::initMetadata(): call initializeClient() above the size-cache lookup. It only constructs an httplib::Client (no network I/O), so the invariant "an initialized HTTPFileInfo always has a client" is restored while keeping the HEAD deduplication.
2. Defensive null-client guards in headRequest / getRangeRequest / postRequest / putRequest so no code path can dereference an uninitialized client.

Testing

New regression test httpfs_httpfs_test (httpfs/test/httpfs_test.cpp), which serves a fixture file over a local range-capable httplib server and exercises the real VFS path:

• HttpFileSystemTest.ReadAfterRepeatOpenOfSizeCachedUrl: opens the same URL twice, reads (including at a non-zero offset) from the second handle. Segfaults without the fix; passes with it.
• HttpFileSystemTest.RepeatOpenDoesNotIssueSecondHead: asserts repeat opens still issue exactly one HEAD (the perf win of 05c63dd is preserved).

Verified:

• without fix: test binary dies with SIGSEGV (exit 139) in the regression test
• with fix: both new tests pass, existing httpfs_xetfs_test (4 tests) passes
• end-to-end: repeated remote parquet reads and a 20000-row multi-block remote CSV return correct counts with the fixed extension