feat(alarm): operator value-threshold alarms on the home (#158 Phase 2a)

[takashikasuya]

Summary

Implements ADR-0005 Phase 2a (#239). Per the maintainer decision: twin per-point thresholds, derived-on-read, stateless, no notification. The operator home now surfaces points whose current value breaches an opt-in threshold — a distinct axis from freshness (arrival, #158 Phase 1) and gateway connection (#230).

• Twin (opt-in per-point) — bos:alarmHigh / alarmLow / warnHigh / warnLow added to OxiGraphOntology, projected in BuildPointSelect, carried on the Point entity. Distinct from the control-write bounds (bos:min/maxValue, Close remaining gateway multi-connection test gaps from #114 #153) — these are the normal-operation value range. A point with no thresholds is simply not evaluated (unknown). Demo thresholds seeded on the sample twin (室温 30/28/10, CO2 1500/1000).
• Pure classifier lib/telemetry/alarm.ts — classifyPointAlarm → ok/warn/critical/unknown; critical (outer) limits win over warn (inner); the threshold boundary counts as a breach; no value / no thresholds → unknown. Snapshot semantics, no hysteresis (stateless).
• alarm-loader.ts — reads current values via the existing batch-latest (the API already returned value; the frontend was dropping it — now kept) and classifies against each point's thresholds.
• buildAttentionList — merges value alarms into the worst-first list (critical→warn→missing→stale). activeAlarms drops alarms on stale/missing points so an hour-old reading is never shown as a live 「異常値」 (a /code-review finding, fixed here).
• operator-home — an 異常 summary card + value-breach rows (「異常値（上限 40）」 / 「注意値（下限 …）」), linking to the point detail like the freshness rows.

Deferred (ADR-0005): notification / event history / ack (Phase 2b — stateful, per the maintainer's memory/stability concern) and per-unit rule editing (Phase 2c).

Test plan

• dotnet test — Shared.Test 688, ApiServer.Test 146 (Point entity + SPARQL projection additive)
• yarn typecheck / yarn lint (0 errors) / yarn test (484, +7: classifier, loader, aggregate merge + activeAlarms, mapping, operator-home) / yarn build
• Playwright home + admin-flows (11, route-mocked)
• /code-review (high) — the stale-value-alarm finding was fixed (activeAlarms); two remaining are documented (below)

Known / documented (not blockers):

1. Not verifiable offline (no OxiGraph/Docker): the real SPARQL projection of the new bos:alarm* triples end-to-end. Covered by the unit path (MapPoint, classifier, loader) + the seeded demo twin; the live twin read runs once the stack is up.
2. Aspida regen pending: mapping.ts reads the new Point.alarm* fields via a documented cast until ./sync-type.bash runs against a live API server (not runnable offline) — the regen is a no-op that formalizes them.
3. Two batch-latest reads per floor (freshness + alarms) — admin-only low-frequency; consolidating into one shared fetch is a noted follow-up.

Checklist

• I read CONTRIBUTING.md and followed the coding conventions in CLAUDE.md
• I added or updated tests for the change
• I updated relevant docs — ADR-0005 status (Phase 2a implemented + the D1/D2 decision)

🤖 Generated with Claude Code

---

Generated by Claude Code