Centralise dynamic NeXus transforms for f144-driven geometry

[SimonHeybrock]

Why

Fixes #922. LOKI's i_of_q workflow crashes at construction against the chopper-aware geometry artifact: the rear-bank depends_on chain walks through /entry/instrument/detector_carriage/value, an empty NXlog placeholder, and essreduce's reject_time_dependent_transform filter fires because no live-stream override is installed.

The narrow fix would have copied DetectorViewFactory's existing per-source dynamic_transforms wiring into _i_of_q_factory. We chose not to. The same pattern was already silently missing from beam_monitor_m4 (the unfinished tail of #780/#782) and would need to be re-applied to every future SANS or monitor variant whose chain walks the same node. The mechanism's failure mode — "if you forget the wiring, the workflow crashes only when CI runs against the new artifact" — does not scale. This PR replaces it with a per-instrument registry and a single helper, so the project's "no silent wrong values" property covers dynamic geometry uniformly.

The fix is artifact-version-independent: the patched chain provider depends on a context key, so Sciline keeps it out of compute(last_static) at construction time. Neither the old length-1 NXlog nor the new length-0 NXlog crashes there. Once the chopper-workflow branch lands, the existing roundtrip test exercises the formerly-failing path automatically.

Design choices

Bindings keyed by NXlog path, not by source name. One NXlog can sit on the depends_on chain of many components — the carriage node drives multiple positions in LOKI today, and more elsewhere tomorrow. Path-keying with an explicit consumer set mirrors the depends_on graph and avoids each component re-declaring the same routing.

One author-declared TransformLog subclass per binding. The original NewType('TransformValueLog', sc.DataArray | None) was a typing-spec misuse (pylint/mypy grumble) and lost the per-binding distinction Sciline needs when multiple dynamic transforms apply in one workflow. Short subclasses of a shared dataclass are grep-able, IDE-visible, type-safe, and give Sciline distinct keys.

Apply patches per component type with a fused single-step provider. Sciline replaces providers by return type, so a patched provider for NeXusTransformationChain[T, SampleRun] cleanly displaces essreduce's generic implementation for that instantiation; multiple component types compose. Fusing the prior two-step TransformValueLog → TransformValue → chain into one provider drops a parallel hierarchy whose intermediate had a single consumer per binding; splitting back is a binding-local refactor if ever needed.

Aux-source merging centralised in Instrument.register_spec. The old design threaded dynamic_transforms=… through every factory and required each callsite to remember it. Centralising the merge means all specs benefit, including those that call register_spec directly (e.g. _i_of_q_factory). Scoping by the binding's consumer set keeps unrelated specs free of phantom aux inputs.

Registry sits on Instrument; helper is a method. Binding-matching is pure data on the instrument; only the workflow.insert(...) step is Sciline-aware. Exposing it as Instrument.apply_dynamic_transforms keeps the registry encapsulated and matches the spirit of register_spec. Callers pass an explicit source_name → component_type mapping, which doubles as documentation of which loaded components are subject to dynamic geometry.

Trust CI, not artifact introspection at runtime. A registry-validator test (per instrument) walks each binding's nxlog_path from every declared consumer in the actual artifact and scans for orphan empty NXlogs. With that in place, the runtime helper does pure registry matching — no file I/O, no h5py walker on the construction hot path. The scippnexus-based chain walker lives only in the test.

Instrument becomes a required argument of DetectorViewFactory and create_monitor_workflow. Every production callsite already had an Instrument in scope; optionality only allowed misuse to compile. With the argument required, InstrumentDetectorSource — which just wrapped the instrument to pass it through data_source — became redundant and is removed; data_source is optional and defaults to instrument-derived geometry. Also tightens the wavelength-mode guard so the instrument-derived path is rejected explicitly rather than failing deep in essreduce. This is a small detour from the issue's scope, but the asymmetry would otherwise have remained as a confusing artifact of the rollout order.

Loud failures everywhere. Forgetting to call the helper: reject_time_dependent_transform raises at construction. Helper called but routing not wired: UnsatisfiedRequirement at construction. Both wired but no f144 message yet: "no samples yet" at first finalize. Registry inconsistent with artifact: CI test fails. Empty NXlog on a consumer's chain that no binding covers: CI test fails. No path silently substitutes a default.

Deferred

• beam_monitor_m4 wavelength. Its position is represented by a separate empty NXlog (trans_20) rather than the carriage node. Resolving it needs either a make_geometry_nexus.py change so m4 shares the carriage chain (a one-line addition to the binding's consumers) or registering an independent m4-position f144 stream. Tracked in the CI validator's known-orphan list; the test trips the moment the entry is removed.
• Readiness gate at job start. The closure already raises a clear "no samples yet" error at first finalize. Converting that into an actionable error at job start is orthogonal to the correctness fix and worth its own work.

Test plan

• CI passes against main.
• After the chopper-workflow branch lands and registers the new geometry artifact, confirm test_workflow_roundtrip[loki/i_of_q/1] exercises the formerly-failing path under tox.
• Run LOKI i_of_q in dev mode against the new artifact and confirm the rear-bank workflow produces I(Q) once a detector_carriage f144 message arrives.
• Drop the detector_carriage f144 producer mid-run and confirm the workflow surfaces the "no samples yet" error rather than producing stale results.