fix(qwp): fix ingestion stalls after append failures and on sender startup

[bluestreak01]

Refs questdb/questdb#7143

This PR bundles two QWP fixes for the upcoming 1.3.1 client release. Both are small, localized, and validated by their own regression tests.

Microbatch buffer release on append failure

• QwpWebSocketSender.sealAndSwapBuffer() left the microbatch buffer in SENDING state when CursorSendEngine.appendBlocking() threw. No I/O thread ever recycles a buffer the engine never accepted, so the next flush would wait 30 s on the recycle latch and then throw Timeout waiting for buffer to be recycled.
• The catch block now puts the buffer back into a non-in-use state on the user thread before rethrowing: markRecycled() from SENDING, rollbackSealForRetry() from SEALED.
• The encoded payload itself is dropped, but flushPendingRows() aborts its post-enqueue state updates after sealAndSwapBuffer() throws, so the source rows in tableBuffers and the sent-schema watermark stay intact. The next batch re-emits the same rows together with the full schema and symbol-dict delta — matching the existing "self-sufficient frames" invariant the cursor SF pipeline already relies on.

Segment manager wakeup on register

• SegmentManager.register() now unparks the worker after publishing the new ring. Without the wakeup, if the worker thread acquires the manager lock before register() does, it snapshots an empty rings, services nothing, and parks for the full poll interval. A ring whose first append does not cross the high-water mark fires no producer-side wakeup either, so the spare never lands until the poll expires.
• The race is pre-existing in the store-and-forward code, latent since the SF feature first landed. It surfaced on the parent-repo CI run only because the mac-other + JaCoCo combination consistently scheduled the worker first; the prior fix (commit 19c5c65) widened the test budget to 2 s but the poll interval is 5 s, so no budget below 5 s could rescue that ordering.
• The fix is one call to wakeWorker() after setManagerWakeup(). LockSupport.unpark is cheap, is a no-op when the worker has not been started, and grants a permit that the next parkNanos consumes immediately — both interleavings are covered.

Tradeoffs

• The microbatch-buffer fix's close()-after-failed-flush retries the drain. If the underlying engine is still wedged (e.g. permanent backpressure, like the deliberately-undersized engine in the new test), that retry fails again and surfaces as a separate LineSenderException. The test acknowledges this and swallows the close-time rethrow. Real-world transient backpressure is expected to clear between the user's flush call and close(), in which case the drain succeeds.
• The segment-manager fix bundles a second unrelated change under a title that names only the microbatch fix. The release-note line for 1.3.1 will undercount the actual scope. Mitigated by clearly separating the two fixes in this description and in the commit history (two separate commits).
• No state-machine extension on either fix: MicrobatchBuffer.markRecycled() / rollbackSealForRetry() and SegmentManager.wakeWorker() already existed.

Test plan

• New QwpWebSocketSenderTest.testFlushAppendFailureDoesNotLeaveMicrobatchBufferInUse (the reproducer from QWP WS: cursor append failure leaves microbatch buffer stuck in SENDING questdb#7143). Fails on main with buffer0=SENDING, buffer1=FILLING and the 30 s recycle-timeout suppressed exception; passes in ~80 ms with this change.
• New SegmentManagerTest.testRegisterAfterWorkerParkedWakesWorker. Sleeps 250 ms between start() and register() to guarantee the worker has parked on an empty rings snapshot, then asserts the spare lands within 2 s. Without the wakeWorker() call this test fails reliably; with it, all 9 SegmentManagerTest cases pass.
• Full QwpWebSocketSenderTest and the broader QWP client suite still pass.

🤖 Generated with Claude Code

[mtopolnik]

[PR Coverage check]

😍 pass : 3 / 5 (60.00%)

file detail

	path	covered line	new line	coverage
🔵	io/questdb/client/cutlass/qwp/client/QwpWebSocketSender.java	2	4	50.00%
🔵	io/questdb/client/cutlass/qwp/client/sf/cursor/SegmentManager.java	1	1	100.00%