Offline ring buffer with reconnect drain

Keeps the producer hot path non-blocking regardless of daemon state: a bounded in-memory ring buffers records while the daemon is down and drains them FIFO on reconnect (a refinement of the DLT backend, taktora-log-dlt DLT-protoco... (FEAT_0072)). Overload manifests as the documented drop policy, never as a producer stall.

Feature: Offline ring buffer with reconnect drain FEAT_0076
status: approved
satisfies: FEAT_0072
is satisfied by: REQ_0812, REQ_0813, REQ_0814, REQ_0815

When the daemon socket is unavailable (early boot, daemon crash, daemon restart), the backend shall buffer records in a bounded in-memory ring and drain them on reconnect. The hot path stays non-blocking regardless of daemon state; overload manifests as the documented drop policy, never as a producer-thread stall.
Requirement: Emission shall not block the calling thread REQ_0812
status: implemented
satisfies: FEAT_0076
is verified by: TEST_0812
links outgoing: BB_0091, TEST_0812

Emitting a log record via log::* macros shall not block the calling thread. The producer side of taktora-log-dlt shall be a bounded-queue push only; all socket I/O shall happen on the backend’s own background flusher thread. The producer shall never wait on connect, write, or any other socket syscall.
Requirement: ERROR and FATAL emission shall not heap-allocate REQ_0813
status: approved
satisfies: FEAT_0076

Emitting an ERROR- or FATAL-level record shall not require heap allocation on the producer side. Pre-sized record buffers shall be used. DEBUG / TRACE / INFO may format-allocate under the formatter’s usual rules — only ERROR and FATAL are covered by the no-alloc guarantee, on the grounds that those are the records that must survive even under memory pressure.
Requirement: Bounded in-memory ring buffers records while daemon is down REQ_0814
status: implemented
satisfies: FEAT_0076
is verified by: TEST_0814
links outgoing: BB_0091, TEST_0814

When the daemon socket is unavailable, the DLT backend shall buffer records in a bounded in-memory ring (capacity configurable at init, with a documented safe default). On reconnect the backend shall flush the ring in FIFO order before resuming live emission. The ring’s storage shall be allocated once at init and shall not grow at runtime.
Requirement: Drop-oldest overflow policy with summary record on reconnect REQ_0815
status: approved
satisfies: FEAT_0076

When the in-memory ring is full and a new record arrives, the backend shall drop the oldest record, increment an internal drop counter, and continue. The backend shall not panic, shall not spin, and shall not block. On the next successful reconnect the backend shall emit exactly one summary record at the leading position of the drain (DLT App ID = the emitter’s, Context ID = reserved diagnostic context, message body = taktora.log.dropped count=N first_dropped_at=...) and reset the counter to zero.