Release v0.4.0

agora-etl-rs becomes Agora's optional
acceleration substrate for measured core hot paths while Python keeps runtime
semantics.

Added

• agora.core.acceleration as the single optional acceleration boundary.
• DeliveryConfig(acceleration_mode=...) with auto, off, and required.
• Declarative [performance] config with acceleration and profile.
• Acceleration status in Pipeline.explain(), run summaries, tracing
  attributes, agora run --plan, agora doctor, and agora doctor --json.
• Performance profiles that resolve to explicit runtime settings while keeping
  manual overrides authoritative.
• Benchmark matrix runner for 0.3.3 baseline, pure Python 0.4.0, and
  Rust-enabled 0.4.0 runs.

Changed

• Runtime modules no longer import agora_rs directly.
• Existing Rust primitives are selected through the acceleration facade.
• Direct flush, Rust prefetch, checkpoint state, metrics accumulator, and linear
  batch buffer decisions are visible in runtime metrics.
• throughput profile is the opt-in gate for Rust source prefetch outside
  buffered lanes until release benchmarks justify broader activation.

Removed

• agora.core.runtime.RecordDeliveryCoordinator
• agora.core.sink._WRITE_OK
• agora.core.context compatibility exports:
  _NOOP_SPAN_SCOPE, _BoundLogger, _NoopSpanScope, _PipelineSpanScope,
  and _normalize_trace_value
• Legacy source/sink data-plane bool flag inference
• BaseSource.runtime_counters() compatibility shim

Release Rule

Performance claims must be backed by benchmark JSON. Candidate Rust expansions,
including JSONL batch encoding and Arrow IPC transport, remain benchmark-gated.