feat: always carry the canonical componentPath in the update extra

[mbreissi]

Every published SouthboundSignalUpdate now carries the signal's full, untruncated MTConnect component path as componentPath on the update-level extra map — unconditionally, for a truncated derived channel, an untruncated one and an explicitly configured signal alike. One consistent place, no conditional branch in reader code.

Placement — verified against the core, not assumed

Checked against proto/edgecommons/v1/telemetry.proto and the Rust codec at the pinned rev (a14a328):

• SouthboundSignalUpdate declares map<string, EcValue> extra = 100.
• to_telemetry collects every body key that is not signal/samples into that map (copy_extra(obj, &["signal", "samples"])); from_telemetry puts them back. So the update-level extra genuinely round-trips — it is what already carries the device block today.
• componentPath collides with nothing: the seven reserved sample keys are value, quality, qualityRaw, sourceTs, serverTs, sourceTsMs, serverTsMs; the update level reserves only signal and samples. No fallback to sourcePath was needed.
• EcValue has a NullValue arm, so a JSON null survives the codec — which is what makes unconditional presence possible in every case.

tests/component_path.rs asserts this on the decoded body after a real Message::to_vec() → Message::from_slice() round-trip, not on the JSON.

The probe-absent edge case

An explicit signals[] entry whose dataItemId is not in the probe model is the permanent-BAD case (MTC_NO_SUCH_DATAITEM). sb/signals does not serve nothing for it — unlearned_address serves address.componentPath: null. So the update stamps null and the key is still present. There is no exception: the field is on every update, always.

Three values, all of them present:

Value	When
"Axes[Axes]/Linear[X]"	The component chain, slash-joined, exactly as the probe declares it.
""	A device-level data item that hangs off no component (Lavail on the live demo Okuma).
null	No device model describes the signal — the same null sb/signals reports.

One source of truth

ProbeModel::component_path_of is the only formatter. address_of (what sb/signals and sb/browse serve) now reads it too, so the update and the control plane cannot drift apart. Reading.component_path carries it through the seam; the derived channel may be shortened to the UNS topic budget (L12), this never is.

No core change

The SignalUpdate builder has no update-level extra setter, so supervisor.rs uses the facade's own two-step form rather than hand-building anything:

DataFacade::build_body (the whole SOUTHBOUND §2.1 contract — quality defaulting, the qualityRaw marker, the serverTs fill, the samples wrapper) ▸ app::stamp_component_path (one additive key) ▸ DataFacade::publish_body_via (mints data/{channel} from the same effective_signal_path, stamps identity, honours the per-call channel override). Both structural checks publish() makes are re-made here rather than dropped.

Live evidence

demo.mtconnect.org was up; the real probe parsed through this adapter's own ProbeModel:

LX1actm     componentPath="Axes[Axes]/Linear[X]"
Lpexecution componentPath="Controller[Controller]"
Lavail      componentPath=""            <- the device-level empty-path case, on a real machine

The README's captured Okuma envelope now shows the field with that verified value.

Tests & gates

• cargo test — 390 passing, 0 failing (308 unit + 82 integration), of which 12 new: 7 unit in app.rs (unconditional presence, exact key, no reserved-name collision, "", null, truncated/untruncated/explicit parity, idempotent stamping), 4 seam tests in device.rs (value equals what sb/signals serves, device-level "", derived-signal untruncated path, probe-absent → None), plus 5 in the new tests/component_path.rs (protobuf round-trip at the update level, deep path, ""/null, batched window carries it once, immediate/flush paths keep it) and extended assertions in the deep-path poll_acquisition e2e and the live agent_integration case.
• cargo clippy --all-targets -- -D warnings — clean.
• cargo llvm-cov --fail-under-lines 90 — 96.68% lines (97.16% regions), gate 90.

Docs

• docs/reference/messaging-interface.md — new componentPath section with the value table, the "always present, no exception" statement, and the update-vs-channel distinction; the body example is now MTConnect-shaped.
• docs/explanation.md — why presence is unconditional rather than truncation-conditional.
• DESIGN.md — D-MtconnectAdapter-L13 records the user decision and the two rejected alternatives: (a) truncation-conditional stamping (conditional reader code, presence becomes a function of the topic budget) and (b) promoting the path to a first-class address block (a four-language wire-contract change for one protocol's shape; reconsider only if a second deep-path protocol appears). L12's "nothing is lost" bullet updated.