[mrcsin]

Closes #116

Problem

The chart's live edge polled the archive. Simple-Scada flushes the archive 1 to 3 minutes late for a fast pen,
about 8 minutes for a slow one and up to a day for a constant one, so a change on the plant could not reach the
chart within 1 s.

What changed

The live edge reads the SCADA's OPC UA server; history stays on the archive.

• SemiPlot.DataSource.OpcUa, a new project:
  • OpcUaLiveSource is the second implementation of ILiveSource, registered by AddOpcUaLiveSource.
  • One 2.0.1 ManagedSession per process. It dials the configured opc.tcp://<host>:<port> and keeps it
    (UpdateBeforeConnect = false).
  • It takes the anonymous token of the None endpoint, and finds the project ID by browsing Objects for
    root, browsing again on every new session.
  • It subscribes each pen set at 100 ms, monitored-item queue 1. SourceTimestamp, which is true UTC, becomes
    Sample.TimestampUtc.
• Breaks:
  • A lost session writes one break per pen, 1 ms past its newest stamp.
  • A Bad notification writes a break. Uncertain draws as a value.
  • The chart (TrendPenState) writes a break ahead of a pen's first live sample past its last column, so a
    pen-set change draws no gap.
• Faults:
  • The pens the source cannot serve (unknown or non-numeric) are one Warning entry per subscription.
  • OPC UA has a second status-bar indicator, which is neutral while connecting. A Warning and a recovery
    Info go through ArchiveFailureMapper with the new LiveSourceError in SemiPlot.Core/Data/Errors.
  • A 30 s grace at start holds the first Unreachable back.
• Configuration: a new opc/ section (ConfigFiles/opc/opc.yaml, host: 127.0.0.1, port: 49875). A file
  that does not parse or carries a bad value is a startup failure; an unreachable server is not.
  DeliveredConfigurationTests covers the loader.
• Deleted: PostgresLiveSource, RealtimePoll, the RealtimePoll and RealtimeBaseline statements, and
  poll_interval_ms with its settings field, resx keys and tests. A leftover key in an installed file is
  ignored.
• Stand: converge gains a required --opc-port and writes opc/opc.yaml. The AppHost passes 55875 and runs
  the demo writer with --archive-lag 60.
• Docs:
  • [DEC:opc-live-edge] in sources.md, and the README locked decision.
  • data-integration.md#realtime, and its time boundary.
  • RT-1 in trend-feature-spec.md.
  • charting.md, overview.md, ui-text.md, testing-strategy.md, bench.md, AGENTS.md.
  • The OPC measurements in scada-archive-observed.md.

How it was verified

• dotnet build SemiPlot.slnx: 0 errors, only the AppHost's two NU1902.
• dotnet format --verify-no-changes exits 0. dotnet terse over every changed .cs file exits 0.
• SemiPlot.Tests.Unit: 2498/2498. SemiPlot.Tests.Integration: 205/205.
• The OpcUa integration filter passes 13/13 with no container, against the in-process demo server on free
  ports. Its coverage:
  • values under Asia/Tokyo;
  • a restart in place and a transport cut;
  • the grace and a late server;
  • unknown pens;
  • the bounded Dispose() under a dropping SynchronizationContext;
  • a pen-set change;
  • the journey to the chart.
• git grep proves the deletions and that no OPC test names 49875.
• The real-SCADA acceptance is recorded in a comment on this PR.
• Not run:
  • the manual steps that need a gesture or a stopped project (plan Post-Completion steps 4-7);
  • the stand checks;
  • the README screenshots, which still show the removed poll field and one indicator.

Known gaps left open

• StartupFailureViewModel still offers the settings dialog on an opc/ failure. The dialog cannot edit opc/
  until opc-settings lands.
• Four narrow races from the last review pass:
  • Close() can land between OnLost's check and Emit (PenSubscriptionSet.cs:93-95).
  • ApplyProjectAsync ignores a changed namespace index.
  • A standing setup fault can show while ConnectAsync retries after a successful discovery.
  • A drop in the few milliseconds before the state handlers attach gives no loss breaks.
• A brief browse failure on a live server turns a standing ProjectMissing into Unreachable and back, which
  writes two extra panel entries.

[mrcsin]

Acceptance on the real SCADA, 2026-10-21

Simple-Scada 2.7.6.0 on this machine, project running, OPC UA on 127.0.0.1:49875. The viewer ran from
c9201ef with --logging-level debug, the tracked opc/opc.yaml, for 5.5 minutes, window sticky, no
gesture.

The archive side was a throwaway Postgres container. It held a pg_dump of the SCADA's capture database,
taken at start, with the catalogue filled for the 21 SCADA variables (ids 2-22). Its semiplot_tags was
brought to the v0.5.0 columns. The capture database itself was not written.

Check	Result
Session	Connected 1.3 s after start, no OPC entry in the log, no Warning or Error from the live source
Apply moment minus the newest SourceTimestamp (debug line Live batch applied)	337 batches: median 94 ms, p99 203 ms, max 703 ms (the first batch, the initial values); 0 at or over 1000 ms
History reads while sticky	none after the opening reads; only the minimap's own band reads
Bool1 (21), Int64Array1 (22)	served as numbers (the server sends Byte, as measured in the roadmap); no PensUnserved entry

Not checked here:

• Post-Completion step 3, nothing changes when the archive catches up. The archive copy was static, so no
  rows arrived.
• Steps 4-7: a gesture, a stopped project, a killed server, a start while the project is stopped.
• The stand checks.