Stop reactor leaks, TLS handshake hangs and reactor spin

[SixiS]

A peer that accepts TCP but stalls TLS can leave Cassandra.cluster and reconnects hanging while the reactor spins. This change bounds TLS handshakes by connect_timeout, avoids polling idle sockets, and makes cluster shutdown and reactor restart release and restore their resources correctly.

Runtime target for review and validation: Ruby 3.4+.

Based on main (86b8ba00), retaining its Cassandra 5.x schema-fetcher mapping, vector support, and Ruby 3.4 compatibility updates.

Changes

• Add Cassandra::Cluster::IoReactor: share a deadline across TCP connect and TLS handshake, select pending TLS handshakes for readability when required, and close the underlying TCP socket with TLS. connect_timeout: Float::INFINITY remains supported.
• Sleep until IO, the next timer or connect deadline, or an explicit wake-up. Drain with a fixed tick and evict invalid descriptors without hiding unrelated select errors.
• Wake the reactor when connections or listeners close or drain, so idle descriptors are released promptly. Preserve close-listener ordering and contain socket IO errors without stopping healthy connections.
• Accept completed TCP connections and TLS handshakes when observed at the deadline; only still-pending attempts time out.
• Restore the wake-up pipe atomically with the start transition. Each reactor run owns its start/stop promises, and completion callbacks run outside the state lock.
• Keep reactor shutdown owned by ControlConnection. A distinct initial :disconnected state allows closing a never-connected cluster; repeated close and explicit reconnect use the correct lifecycle future.
• Add optional reconnection jitter, sampled uniformly within the bounded interval so retries do not concentrate at the ceiling.
• Pin ione to ~> 1.3.0, name the reactor thread io_reactor, and update the gem version and lockfile to 3.2.6.

Validation

Validated on Ruby 3.4.5 using the checked-in lockfile, including the ione fork at e665ea33 (1.3.1).

Run	Result
Native extension compilation (bundle exec rake compile)	Pass
Full unit suite	1,141 examples, 0 failures, 2 pending
Reactor specs, included in the full suite	50 examples, 0 failures
Isolated reactor specs on Linux, Ruby 3.4.5	50 examples, 0 failures
Additional local schema-picker checks for Cassandra 3.11.17, 4.1.7, and 5.0.9	3 examples, 0 failures

The full suite uses the real Snappy and LZ4 native extensions, with no compressor stubs. The two pending examples cover credentials-request equality and hashing. CCM integration tests and RuboCop were not run.

The extra schema-picker checks exercised driver dispatch and schema reconstruction with the existing system_schema fixtures; they did not use a live Cassandra server.

The shutdown/restart interleaving was reproduced deterministically before the fix and now passes. Regression coverage also checks an old stop future completing after a new run has started. Network tests use local loopback sockets.

Review findings resolved

All previously listed findings and the subsequent restart-race finding are addressed:

• Restore the unblocker after stop and crash.

• Use a fixed drain timeout tick even when timers are overdue.

• Remove the redundant nested reactor stop from Cluster#close_async.

• Guard a missing raw socket in TLS close.

• Bound the jitter window before drawing.

• Use a distinct initial control-connection state and fresh close promises after explicit reconnect.

• Update the lockfile PATH version to 3.2.6.

• Preserve infinite connect timeouts without constructing an infinite Time.

• Skip redundant timer wake-ups on the reactor thread.

• Bound select by the nearest finite connect deadline.

• Probe descriptors to handle real EBADF errors.

• Handle select TypeError from a missing to_io.

• Normalize invalid jitter errors to ArgumentError and correct the randomness-source documentation.

• Restrict ione to the 1.3.x series.

• Make pipe restoration and restart atomic with shutdown completion, preserving each run's promises.

• Wake idle select after closing connections or listeners, including drain paths, before notifying connection close listeners.

• Contain socket IOError and EBADF during dispatch, and tolerate transient accept errors.

• Prefer completed TCP connects and TLS handshakes over elapsed deadlines, including connections queued while stopped.

Follow-up

The reactor fixes are also proposed in ione PR #1. After that merges, update the ione dependency revision and replace the driver-local reactor with Ione::Io::IoReactor in the same change.