[SimonHeybrock]

Part of #746. Fixes #868. Supersedes #793, which fixed the same on the old design; its test is included here. Rewrites StreamProcessor on sciline.Stage as planned in sciline ADR 0003 (design doc section 6.6). Draft: needs a sciline release with Stage and scipp/sciline#256, so CI against released sciline will fail until then.

What changed

StreamProcessor splits its workflow into stages built in one prepare_stages call, so everything that depends on neither context nor chunks is still computed once at construction:

• context stages, computing from the context keys what the other stages read that depends on the context, and context-only targets. There is one per group of values depending on the same context keys, ordered so that a stage runs after those whose results it reads. set_context runs only the stages that read an updated key or a value recomputed before them;
• one chunk stage per group of accumulators that read the same dynamic keys;
• a finalize stage, computing the targets from the accumulated values.

Which context-derived values the context stage computes is found from the cached keys of the chunk and finalize stages built on their own (ADR section "Nested loops"). This replaces _FedWorkflow, _build_streaming_workflow, pruning by assigning None, and the graph walks that found context and finalize inputs. visualize uses sciline.visualize_stages with the same node groups and colors, which removes the custom styling and legend code. The public interface is unchanged.

Behavior changes

• A target that depends on a dynamic key other than through an accumulator raises at construction unless allow_bypass=True. Previously None was passed to the provider at finalize.
• An error about missing context now names the context keys that were never set.

The visualize tests that called removed private helpers now check the styles of the rendered nodes.

Performance of context updates

set_context timed on the esslivedata LOKI rear detector view in wavelength mode, with live transforms and ROIs:

Update	Before	Single context stage	This PR
detector carriage	86 ms	86 ms	86 ms
sample motor	11–13 ms	~80 ms	7–8 ms
ROI edit	77 ms	~80 ms	1.6 ms
wavelength lookup table	11 ms	~80 ms	1.7 ms

A single context stage, as the sciline design document proposed, rebuilt the detector projection on every sample motor message, although the projection does not depend on the sample. Grouping context-derived values by the context keys they depend on avoids that.

ROI edits got faster than before: the ROI values read the projection, and the old set_context computed them on a workflow that was fed only the context keys, so it rebuilt the projection as an intermediate on every edit and discarded it. Now the ROI stage takes the projection held from the last position update as input. The sample motor and lookup table improvements likely have the same cause; not traced.

Data updates on the same detector view: accumulate is unchanged (about 95 ms per chunk of 100 events, 175 ms for 1M events), while finalize went from 83 ms to 5 ms. The old finalize workflow was fed the context keys and only those context-derived values that are targets, so each finalize recomputed the ROI masks, ROI pixels, and pixel weights, including the projection they read. The finalize stage now takes them as input. Per-call graph building, which Stage.run avoids, cost only about 1 ms per call before, since inputs were already fed without rebuilding the pipeline.

Sciline findings

• Stage and prepare_stages rejected input keys the outputs do not need. That forced this driver to build stages only to find exact inputs, ruled out an accumulator feeding another accumulator, and made prepare_stages reject a context-only target that only the context stage outputs. Drop stage inputs the outputs do not need sciline#256 makes stages drop such keys instead, as Pipeline ignores unneeded parameters; run stays strict.
• Pipeline.provide and a reporter argument on Stage, listed as prerequisites in Tracking: migrate from sciline map/reduce to stages and aggregations (sciline ADR 0003) #746, are not needed: Stage.run takes the values directly, and StreamProcessor never used a reporter.

Test plan

• essreduce test suite passes with sciline main plus Drop stage inputs the outputs do not need sciline#256.
• esslivedata test suite against this branch: all pass except one test on an unmerged esslivedata branch that relies on the old None bypass and needs allow_bypass=True.
• Construction and processing time of the esslivedata workflow tests unchanged.

🤖 Generated with Claude Code