Non-blocking hot path and offline buffering

Test cases for the non-blocking producer hot path and the offline ring buffer with reconnect drain (Offline ring buffer with re... (FEAT_0076)).

Test Case: Producer-side send never blocks the calling thread TEST_0812
status: implemented
verifies: REQ_0812
links incoming: REQ_0812

Goal. Confirm that pushing a record into the flusher’s bounded channel completes in well under 5 ms regardless of what the flusher / daemon socket is doing — the producer never waits on connect, write, or any socket syscall.

Fixture. crates/taktora-log-dlt/tests/flusher.rs:20-54 (gated #[cfg(unix)]). A UnixListener stands in for the daemon; the test asserts a wall-clock timing bound.

Steps.

  1. Spawn the flusher with a UDS transport pointed at the listener and a 64-slot offline ring.

  2. Record t0 = Instant::now() and call tx.send(b"hello".to_vec()).

  3. Assert t0.elapsed() < 5 ms.

  4. Join the listener thread and confirm it read back "hello".

Expected outcome. send returns in under 5 ms; the bytes arrive at the listener; the flusher shuts down cleanly. The producer hot path is bounded-queue push only — no socket I/O on the caller’s thread.
Test Case: Offline ring buffers and drains FIFO across reconnect TEST_0814
status: implemented
verifies: REQ_0814
links incoming: REQ_0814

Goal. Confirm that records pushed into the OfflineRing while the daemon socket is unavailable are drained in FIFO order on reconnect, and that mid-drain failure preserves the un-drained suffix in original order (no silent loss of un-sent records).

Fixture. Ring-only FIFO and capacity tests at crates/taktora-log-dlt/tests/ring.rs:4-34; flusher-level mid-drain failure rebuffers in FIFO order at crates/taktora-log-dlt/tests/flusher.rs:57-142 (Unix-only).

Steps. (ring-only)

  1. OfflineRing::with_capacity(3); push a, b; drain.

  2. OfflineRing::with_capacity(2); push a, b, c; drain.

Steps. (mid-drain failure)

  1. Pre-populate a capacity-16 ring with 5 large (256 KiB) records tagged '0'..'4' before the listener exists.

  2. Bind the listener, spawn the flusher; the listener accepts, reads one record, then shutdown(RDWR) on the connection.

  3. After 500 ms, handle.shutdown() and ring.drain_all().

Expected outcome. Ring-only: drained order is [a, b] then [b, c] (oldest dropped, FIFO preserved). Mid-drain: exactly the 4 records '1'..'4' remain re-buffered, in order — proving the un-drained suffix is rebuffered intact when the write to the daemon fails.