Add OracleConnection.cancel() for server-side break

[iliasaz]

Summary

Adds a public OracleConnection.cancel() API that asynchronously interrupts an in-flight statement on the server — the OCI OCIBreak() / python-oracledb Connection.cancel() equivalent.

let task = Task {
    try await conn.execute("BEGIN dbms_session.sleep(60); END;", logger: logger)
}
// elsewhere…
conn.cancel()  // server raises ORA-01013, the awaiting Task throws .statementCancelled

Without this, Task.cancel() on the awaited promise doesn't propagate to the wire, so the server keeps running and the next execute(...) queues behind it.

Wire-level handshake

Mirrors python-oracledb's _break_external + _reset pair:

1. Client sends TNS_MARKER_TYPE_INTERRUPT.
2. Server replies with a BREAK acknowledgment marker.
3. The existing markerReceived toggle echoes a TNS_MARKER_TYPE_RESET back.
4. Server raises ORA-01013 on the in-flight statement.
5. StatementStateMachine.errorReceived(_:) routes through the cancellation branch — extended in this PR to also fail the awaiting execute promise from the .initialized, .describeInfoReceived, and .rowCountsReceived substates (the existing .streaming path was already wired).

Mid-fetch cancels delegate to the existing iterator-drop path (.forwardStreamError(clientCancelled: true) followed by sendMarker(RESET)), so the wire bytes for await stream.cancel() style cancellation are unchanged.

Why this also benefits non-cancel call sites

The PR makes the existing cancelStatementStream path slightly more robust:

• markerState is reset to .noMarkerSent in triggerBreak() so a cancel that follows another cancel doesn't get stuck on a stale toggle when the server elides one of the marker echoes.
• errorReceived(_:) now resolves the original execute promise on ORA-01013 from any pre-streaming substate, where previously only .streaming produced a usable failure.

Code changes

File	What
Sources/OracleNIO/Connection/OracleConnection.swift	Public cancel()
Sources/OracleNIO/OracleChannelHandler.swift	triggerBreak() entry point + .sendBreak action handler
Sources/OracleNIO/ConnectionStateMachine/ConnectionStateMachine.swift	New triggerBreak() and .sendBreak case
Sources/OracleNIO/ConnectionStateMachine/StatementStateMachine.swift	New markCancelledForBreak() returning a BreakOutcome; widened ORA-01013 handling
Sources/OracleNIO/Messages/Coding/OracleFrontendMessageEncoder.swift	marker(type:) overload
Sources/OracleNIO/Constants.swift	TNS_MARKER_TYPE_BREAK / _INTERRUPT typed as UInt8 for the encoder

Test coverage

Unit (5 new, all 363 existing still pass)

Tests/OracleNIOTests/ConnectionStateMachine/StatementStateMachineTests.swift:

• breakCancellationFromInitialized — the canonical dbms_session.sleep repro, statement is in .initialized.
• breakCancellationFromDescribeInfoReceived — between describeInfo and rowHeader.
• breakCancellationFromStreaming — mid-stream, asserts the existing forwardStreamError + sendMarker(read: true) path is still produced.
• breakOnIdleConnectionIsNoOp — .readyForStatement produces .wait.
• breakIsIdempotent — second triggerBreak() before the server responds is a no-op.

Integration (5 new, against a live DB)

Tests/IntegrationTests/CancelTests.swift:

• cancelDuringKernelSleep — BEGIN dbms_session.sleep(10); END;, cancel after 1s; follow-up SELECT runs in well under 8s (the kernel sleep checks for breaks at finite intervals so this is order-of-seconds, not instant).
• cancelDuringCPUBoundExecute — cancel a CPU-bound SELECT COUNT(*) while it's executing; follow-up runs under 2s.
• cancelMidFetch — cancel during a 10M-row stream; iterator throws .statementCancelled, follow-up immediate.
• cancelOnIdleConnection — no statement in flight, follow-up unaffected.
• doubleCancelIsIdempotent — two cancels in quick succession leave the connection reusable.

All 5 pass against the local Oracle Free 23ai test container at the standard ORA_HOSTNAME / ORA_PORT / ORA_SERVICE_NAME / ORA_USERNAME / ORA_PASSWORD env.

Notes

• scripts/soundness.sh passes (license headers, no unacceptable language, swift-format clean).
• DCO: commit is Signed-off-by:.
• API addition only — no breaking changes; all existing tests continue to pass.

Inspired by issue iliasaz/macintora#15 (Oracle SQL IDE) where the missing cancel was particularly visible.

[lovetodream]

This is a great addition. What I don't like about it is that it breaks the concept of structured concurrency with the addition of the cancel method. It should work via Task.cancel instead.

let task = Task {
   try await conn.execute("BEGIN dbms_session.sleep(60); END;", logger: logger)
}
task.cancel()

I didn't have time to review the internals yet. Do you want me to take it over?

[iliasaz]

It should work via Task.cancel instead.

That's a very good point! I didn't think of it. I thought of making it async but that alone doesn't get us any benefits.
Feel free to take it over :-)