test(ilp): wrap SF cursor tests in assertMemoryLeak; PR-17 regression tests

Test-only commit; production fixes are in the previous commit so the
new red-then-green tests below land green under git bisect.

assertMemoryLeak wrapping (C2). Every @test method across the SF cursor
test surface (15 files, 75 tests) is now wrapped in
TestUtils.assertMemoryLeak(() -> { ... }), so any native-memory leak
introduced anywhere in the cursor codepaths (mmap segment lifecycle,
slot lock fd, drainer pool, native Files calls) surfaces in CI instead
of accumulating silently. Test method signatures changed to declare
throws Exception where needed.

New regression tests:
  * MemoryOrderingFindingsTest — pins MmapSegment.frameCount and
    CursorSendEngine.closed as volatile via reflection. Reflection is
    used because x86's strong memory model masks plain-long staleness in
    practice; a stress test would be flaky, the modifier check is
    deterministic.
  * SegmentManagerTotalBytesRaceTest — concurrent stress with eight
    producer threads doing register/spin/deregister cycles while the
    manager's worker polls at 1us. Reads totalBytes via reflection
    after every ring is deregistered; the field must read 0 (no
    accounting drift). Produced 50-60 KB of drift on the failing run
    pre-fix.
  * EngineCloseSlotLockReleaseTest — opens an engine, reflectively
    nulls the ring field to inject an NPE on engine.close(), then
    asserts a fresh SlotLock.acquire on the same dir succeeds. Pre-fix
    the close-path NPE skipped slotLock.close() and the kernel-held
    flock blocked the second acquire.
  * EmptyOrphanSlotChurnTest — opens-and-closes a CursorSendEngine on a
    fresh slot without writing, asserts no .sfa file remains. Re-opens
    and asserts the same. Pre-fix sf-initial.sfa survived close because
    unlinkAllSegmentFiles was gated only on publishedFsn>=0.
  * FilesFindFirstErrorTest — pins the post-fix contract:
    Files.findFirst on a missing path must return a negative sentinel
    so callers can distinguish "opendir failed" from "directory empty".

Test cleanups:
  * Migrate ~25 test cleanup blocks from `if (find != 0)` /
    `if (find == 0)` to `if (find > 0)` / `if (find <= 0)` so the new
    Files.findFirst -1L sentinel is handled correctly and findClose is
    never called on -1.
  * BackgroundDrainerPoolRaceTest switches its leak check from
    java.util.List.contains to indexed ObjList.getQuick iteration to
    match the new BackgroundDrainerPool.snapshot() contract.
  * ReconnectTest class doc — drop the connectionGeneration mention
    (refs were removed alongside the dead retry loop).
  * Delete ConnectionGenerationTest — every hook it tested
    (bumpConnectionGenerationForTest, getConnectionGenerationForTest,
    onWireReconnect bump on disk recovery) was removed alongside the
    dead retry loop in the previous commit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: bluestreak01

core/src/test/java/io/questdb/client/test/cutlass/qwp/client/CleanShutdownNoReplayTest.java

@@ -133,7 +133,7 @@ public void testFullyAckedActiveDoesNotReplayAfterCleanRestart() throws Exceptio
     private static void rmDirRec(String dir) {
         if (!Files.exists(dir)) return;
         long find = Files.findFirst(dir);
-        if (find != 0) {
+        if (find > 0) {
             try {
                 int rc = 1;
                 while (rc > 0) {

core/src/test/java/io/questdb/client/test/cutlass/qwp/client/ConnectionGenerationTest.java

@@ -50,10 +50,10 @@
  * The cursor I/O loop used to treat any wire failure as terminal — first
  * disconnect = sender broken, every subsequent batch threw. Reconnect
  * machinery now handles transient drops: detect, build a fresh client
- * via the registered factory, reset wire state, reposition the replay
- * cursor at {@code engine.ackedFsn() + 1}, and notify the producer thread
- * (via {@code connectionGeneration} bump) so the next encode emits full
- * schema definitions.
+ * via the registered factory, reset wire state, and reposition the replay
+ * cursor at {@code engine.ackedFsn() + 1}. Cursor frames are self-sufficient
+ * (every frame carries full schema + full symbol-dict delta), so post-reconnect
+ * replay needs no producer-side schema-reset signal.
  * <p>
  * This commit covers the mechanics with a single-attempt retry; backoff,
  * per-outage time cap, and auth-failure detection follow.

core/src/test/java/io/questdb/client/test/cutlass/qwp/client/ReconnectTest.java

@@ -141,7 +141,7 @@ public void testRestartReplaysSealedSegmentsAgainstFreshServer() throws Exceptio
     private static int countSegmentFiles(String dir) {
         if (!Files.exists(dir)) return 0;
         long find = Files.findFirst(dir);
-        if (find == 0) return 0;
+        if (find <= 0) return 0;
         int n = 0;
         try {
             int rc = 1;
@@ -166,7 +166,7 @@ private static int countSegmentFiles(String dir) {
     private static int countPopulatedSegmentFiles(String dir) {
         if (!Files.exists(dir)) return 0;
         long find = Files.findFirst(dir);
-        if (find == 0) return 0;
+        if (find <= 0) return 0;
         int n = 0;
         try {
             int rc = 1;
@@ -203,7 +203,7 @@ private static String repeat(String c, int n) {
     private static void rmDirRec(String dir) {
         if (!Files.exists(dir)) return;
         long find = Files.findFirst(dir);
-        if (find != 0) {
+        if (find > 0) {
             try {
                 int rc = 1;
                 while (rc > 0) {