feat(transit-ops): persistent provider health, staleness alerts, OTEL metrics, admin UI

Wire up the observability + automation surface for the self-hosted
Transitous pipeline + transit provider chain. Six pieces land
together because they share data: a Redis-backed provider-health
tracker, per-feed staleness alerting, OpenTelemetry counters,
admin UI, plus two small infrastructure refactors that fall out of
the cleanup.

Persistent provider health
- Replace the in-memory `integrations/transit/health.ts` counter with
  a Redis-backed service at apps/api/src/services/provider-health/.
  State per provider: success/failure totals, sliding window of last
  100 outcomes, lastFailureAt + lastFailureReason, EMA latency
  (alpha 0.2). Auto-disable when window failure rate > 50% over at
  least 10 samples; 5-min cooldown.
- All state writes go through a single Lua script for atomicity. Keys
  `provider:health:<id>` with 30-day TTL.
- IntegrationContext gains optional providerHealth and metricsRecorder
  handles. apps/api wires the singletons; integrations/transit
  /orchestrator.ts wraps every provider call in `timed(health,
  metrics, providerId, method, fn)`.
- New endpoints on /api/data-manager:
  - GET /providers — sorted list with full state.
  - GET /providers/:id — single state or 404.
  - POST /providers/:id/reset — admin-session only.
- /api/integrations/transit/health is now 410 Gone with a pointer
  to the new endpoint.

Per-feed staleness alerts
- Add `consecutive_failures` column to data_manager.feed_state
  (migration 0005_feed_state_consecutive_failures.sql). Pipeline
  stages bump it on validation failure, reset on success.
- New services/data-manager/src/jobs/transitous/staleness-alerts.ts
  with detectStaleFeeds + emitFeedAlerts. Triggers:
  - stale: last_fetched_at older than 48h.
  - consecutive-failures: ≥ 3 in a row.
- Always logs a structured warn line. Optional GitHub Issue sink
  controlled by TRANSITOUS_ALERT_GH_TOKEN + TRANSITOUS_ALERT_GH_REPO;
  dedupes by exact title against open issues.
- New TRANSITOUS_STALENESS_CHECK_CRON (default `0 4 * * *`, daily
  after the 03:00 sync) + post-sync immediate invocation as a
  safety net.

OpenTelemetry metrics
- Add @opentelemetry/{api,sdk-metrics,exporter-prometheus} to
  apps/api. New apps/api/src/services/metrics/index.ts initialises
  a singleton MeterProvider with two instruments:
  - transit_provider_calls_total (counter)
  - transit_provider_call_duration_ms (histogram)
  Labels: provider_id, method, outcome (ok/empty/error/skipped).
- Mount GET /api/internal/metrics returning Prometheus text format.
  No auth — restrict via firewall / docker network.
- New Grafana dashboard at infra/docker/dashboards/transit-providers
  .json with three panels (call rate, p95 latency, error rate by
  provider).

Admin UI
- New apps/web/src/app/admin/transit/ page assembled from:
  - LockfileCard (Transitous ref + lockedAt + lockedBy, display-only).
  - CurrentJobCard (running state + "Run sync now" + "Restart
    MOTIS" buttons; live polling at 5s when in-flight).
  - RecentJobsTable (last 20 jobs + drawer with stage breakdown).
  - FeedsBreakdownChart (pure-CSS bars by region + status).
  - FeedsTable (paginated feed_state with region + status filters).
  - ProviderHealthTable (per-row Reset).
- New apps/web/src/lib/admin/transitHooks.ts (read + mutation
  hooks). Sidebar nav entry added in AdminSidebar.
- All English strings; admin shell has no i18n integration today.

Sweep cleanup
- Delete the deprecated useProviders hook + resolveProvider /
  providerAttribution / providerLabel helpers + the
  /api/integrations/transit/providers route + getTransitProviderAttribution
  orchestrator function + the now-empty attribution.test.ts. All
  callers were removed by the earlier attribution-canonicalisation
  pass; the residue is unused.
- Remove the duplicated ProviderAttribution interface from
  packages/core/src/constants/transit.ts + its barrel export.
  The framework's @openmapx/integration-framework version is now
  the single source. Zero external imports survived the cleanup.
- Extract services/motis-feed-proxy/lib/build.ts into a new
  @openmapx/motis-feed-proxy-config workspace package. Cross-service
  relative import in services/data-manager replaced by a typed
  workspace dep; services/data-manager/tsconfig narrows rootDir back
  to "." (was ".." to accommodate the relative import) and
  tsconfig.build.json emits to `dist/index.js` directly. Dockerfile
  CMD + start script updated to match.
- services/motis-feed-proxy/ now holds only service.json — the
  nginx config rendering lives in the new shared package.
- Strip plan references from code/doc comments (no §, no "plan §"
  citations). Comments now describe behaviour, not the plan they
  came from.

Tests
- 191 files, 2390 passed | 6 skipped. New suites: provider-health
  (10 unit tests covering window cap, auto-disable, cooldown
  expiry, reason truncation, EMA convergence), data-manager-providers
  route (8 cases for auth + 404 + reset), staleness-alerts (14
  cases incl. GitHub sink dedup), metrics module (6 cases incl.
  Prometheus text endpoint), admin transit page + FeedsTable +
  ProviderHealthTable (10 cases).

Author: Medformatik

apps/api/package.json

@@ -32,6 +32,9 @@
     "@openmapx/db-schema": "workspace:^",
     "@openmapx/integration-framework": "workspace:^",
     "@openmapx/place-ids": "workspace:^",
+    "@opentelemetry/api": "^1.9.1",
+    "@opentelemetry/exporter-prometheus": "^0.218.0",
+    "@opentelemetry/sdk-metrics": "^2.7.1",
     "better-auth": "^1.6.9",
     "better-auth-harmony": "^1.3.2",
     "cached-hafas-client": "^5.1.9",

apps/api/src/db/migrations/0005_feed_state_consecutive_failures.sql

@@ -0,0 +1 @@
+ALTER TABLE "data_manager"."feed_state" ADD COLUMN "consecutive_failures" integer DEFAULT 0 NOT NULL;