[FEAT] De-box `Source::next` — the boxed future buys erasure nobody uses

[lxsaah]

Is your feature request related to a problem? Please describe.

Source::next (aimdb-core/src/session/mod.rs:530) returns a BoxFut, so every implementor heap-allocates once per inbound message:

fn next(&mut self) -> BoxFut<'_, Option<(String, Payload)>>;

It is called from the pump_source read loop at aimdb-core/src/session/pump.rs:152, i.e. once per received telegram/message, on every inbound data path in the tree.

Unlike the outbound side, this allocation is not paying for anything. pump_sink holds Arc<dyn Connector> (pump.rs:40), so Connector::publish genuinely has to return a boxed future — that one is the price of the pluggable connector surface and should stay. But pump_source takes mut src: impl Source + 'static (pump.rs:141) — generic and monomorphized, no trait object anywhere. The only two dyn Source mentions in the workspace are the object-safety assertion itself (mod.rs:546) and a test (mod.rs:651). No production code erases a Source, so the box is satisfying a dyn-compatibility constraint that no caller exercises.

This is a code-cleanliness issue, not a performance one. At current connector rates (KNX TP1 is 9600 baud; IP tunnelling runs a single gateway channel) one small allocation per message is nowhere near a bottleneck, and on the inbound path it is outnumbered anyway — aimdb-knx-connector/src/client.rs:80 already does an addr.to_string() plus a Vec<u8> → Arc<[u8]> copy per telegram, both larger than the box. Filing so the reasoning is recorded rather than rediscovered; no urgency.

Describe the solution you'd like

Switch the trait to AFIT, matching the shape already used elsewhere in the tree:

fn next(&mut self) -> impl Future<Output = Option<(String, Payload)>> + Send + '_;

Two in-repo precedents show this works and needs no MSRV movement:

• CommandSource (aimdb-knx-connector/src/client.rs:43) is already written this way.
• EmbassySourceRaw (aimdb-embassy-adapter/src/connectors.rs:99) is already in the target shape, and its EmbassySource bridge boxes purely to satisfy core's BoxFut (connectors.rs:113). Since SendFutureWrapper<F> is unconditionally Send (aimdb-embassy-adapter/src/send_wrapper.rs:25), that line collapses to SendFutureWrapper(self.0.next()) with the box gone and nothing else moving — the allocation-sensitive MCU target is the one where the change is most mechanical.

Blast radius is three production impls plus core's own test scaffolding:

Site	What it is
aimdb-core/src/session/mod.rs:530	the trait declaration
aimdb-core/src/session/mod.rs:546	_source: &dyn Source in _assert_object_safe — must be dropped
aimdb-core/src/session/mod.rs:637,651	MockSource and its Box<dyn Source> test
aimdb-embassy-adapter/src/connectors.rs:112	EmbassySource<S>
aimdb-knx-connector/src/connector.rs:74	KnxSource
aimdb-mqtt-connector/src/tokio_client.rs:318	MqttEventLoopSource

pump_source itself needs no change — it is already generic over impl Source.

Note that a grep for BoxFut looks far more alarming than this change is: nearly all hits are Connection, Dialer, Listener and Dispatch, which are genuinely stored as Box<dyn …> throughout the session engine and are untouched here. Source is the one data-plane trait that declares a boxed future without anyone erasing it.

Describe alternatives you've considered

Leave it as is — the current decision, and defensible. Source mirroring Connector keeps the two data-plane halves symmetric, and keeping the trait dyn-compatible leaves the door open to a future dyn Source holder (a multiplexer over heterogeneous sources, say). If that door is one we want open, the box is the cost of holding it — worth making that an explicit decision rather than an inherited one, which is the main reason for this issue.

Reduce the surrounding allocations first — on the inbound KNX path, interning the group-address topic (it comes from a fixed set of GroupAddress values) and parsing straight into an Arc<[u8]> would each save more per telegram than de-boxing. Independent of this issue and a better payoff if allocation pressure ever actually matters.

Additional context

The blocking consideration is API breakage, not the in-tree edits. Source is publicly re-exported from aimdb-core/src/lib.rs:143, so making it non-dyn-compatible breaks any out-of-tree connector holding a Box<dyn Source>, with no deprecation path for that case. An out-of-tree impl that merely returns a future keeps compiling. In-tree this is an afternoon; as public API it wants a version boundary, so the natural time to do it is alongside the next breaking release of aimdb-core.

Worth revisiting sooner if a genuinely high-rate connector lands on pump_source, at which point the inbound allocation profile becomes worth real effort.