feat: depth-aware channel derivation (leaf-preserving truncation)

[mbreissi]

The bug

MTConnect component paths go deeper than a UNS topic can carry. The live demo Mazak's stock data
item sits on Resources[resources]/Materials[materials]/Stock[stock], which with its id is four
channel tokens — an instance-scoped data topic (ecv1/{device}/{component}/{instance}/data/…) has
already spent four of the seven allowed / separators, so it has room for three. Under
selection: { "mode": "all" } the library refused the topic with DEPTH_EXCEEDED and the signal
never published at all.

The rule

A derived channel is the last k UNS-sanitized component-path segments plus the signal id, where
k is the largest value that fits the instance's real UNS budget.

Component path	Derived channel
(device level)	stock
Controller[controller]	controller-controller/estop
Axes/Linear[X]	axes/linear-x/xabs
Resources[resources]/Materials[materials]/Stock[stock]	materials-materials/stock-stock/stock
Systems[systems]/Hydraulic[hydraulic]/Pump[pump]/Motor[motor]/Sensor[sensor]	motor-motor/sensor-sensor/ptemp

• Root-side segments drop first. The leaf-most ones say what the signal is;
  Resources[resources] above Materials[materials]/Stock[stock] barely narrows anything.
• The id is terminal and never dropped. That is what makes truncation safe: signal ids are
  unique per instance (validate_bindings enforces it for the explicit half, the -2/-3 suffix
  chain for the derived half), so however much path is dropped two derived channels cannot collide.
  No extra suffix rule is needed and none was added — there is a test that asserts uniqueness at
  every k.
• The budget is measured, not assumed. app::channel_budget_of mints a probe topic through the
  library's own Uns builder and takes what the prefix did not spend, so MAX_TOPIC_SLASHES (7)
  and MAX_TOPIC_UTF8_BYTES (256) are never copied into this repo. It is per instance and
  includes the device/component/instance token lengths, so a verbose identity costs channel depth
  rather than silently breaking the topic. src/mtconnect/** stays EdgeCommons-free: it receives a
  plain ChannelBudget { max_tokens, max_bytes }.
• Nothing is lost. The full, untruncated component path stays in the ProbeModel and is served
  as signal.address.componentPath on sb/signals and on every sb/browse entry. Only the topic
  is shaped.

Observability

Ordinary truncation is normal derivation on a deep machine: counted in
ServedSet::channel_truncated and logged once per recompile at DEBUG — no event. The only
warning is the pathological floor, where not even the bare id fits (the identity itself has consumed
the topic): that raises the existing MtconnectSignalSetEvent with reason: "channelBudget", once
per distinct count, and leaves the channel as the bare id so the library's own validation refuses it
loudly on publish.

A hand-set channel on an explicit signals[] entry is untouched, however deep. An explicit
entry that omits channel takes the derived one and is shaped like any other.

Plumbing

ChannelBudgets resolves once at startup (identity is fixed for the process; adding an instance is
already RESTART_REQUIRED) and is stamped by both compile_mtconnect and
SignalRegistry::new, so the startup compile and every live reload derive identical channels. The
budget is hashed into the signal-set generation, because it changes what every derived channel is
and therefore what a browse cursor was minted against.

Live evidence

Run against the real demo.mtconnect.org Mazak with mode: "all", HOST/MQTT into a local broker.
The signal that used to be dropped now publishes:

ecv1/gw-01/MtconnectAdapter/cnc-1/data/materials-materials/stock-stock/stock

134 distinct topics observed, every one within budget; no publish failed / DEPTH_EXCEEDED in the
adapter log. Other deep paths on the same device shaped the same way
(systems-systems/coolant-coolant/cooltemp, controller-controller/path-path/execution, …).

Gates

• cargo test — 377 passed, 0 failed (+15 new: 8 unit in selection.rs, 5 in app.rs, 2
  end-to-end in poll_acquisition.rs; 2 more in reload.rs).
• cargo clippy --all-targets -- -D warnings — clean.
• cargo llvm-cov --fail-under-lines 90 (CI's exclusions) — 96.71% total;
  selection.rs 97.92%, app.rs 95.28%, reload.rs 98.30%.

New fixtures: tests/fixtures/devices_deep_2.7.xml (the demo Mazak stock shape reproduced
verbatim, plus a five-level path and a long-token path that exhausts bytes before levels) and
its /current companion.

Docs

DESIGN.md gains D-MtconnectAdapter-L12 (supersedes L10's channel clause);
docs/reference/configuration.md gains a Deep component paths section;
docs/explanation.md gains A machine model is deeper than a topic.