fix(pipeline): repair Agerpres feed, archive lock starvation, enrichment cadence

packages/backend/convex/crons.ts

@@ -21,13 +21,14 @@ crons.interval(
 // automated refresh is ever wanted again.
 
 // ---------------------------------------------------------------------------
-// Article Enrichment (Embeddings) — Every 40 minutes
+// Article Enrichment (Embeddings) — Every 30 minutes
 // ---------------------------------------------------------------------------
 // Generates embeddings for unprocessed articles.
-// Runs more frequently than ingestion to keep the pipeline flowing.
+// Runs more frequently than ingestion to keep the pipeline flowing and to work
+// down the unprocessed-article backlog faster.
 crons.interval(
   "enrich-articles",
-  { minutes: 40 },
+  { minutes: 30 },
   internal.enrichmentNode.enrichUnprocessedArticles,
 );
 
@@ -68,12 +69,20 @@ crons.interval(
 );
 
 // ---------------------------------------------------------------------------
-// Stale Singleton Archive — Hourly
+// Stale Singleton Archive — Every 53 minutes (drifting)
 // ---------------------------------------------------------------------------
 // Archives stale processing singletons so they stop inflating the vector index.
+// The job yields (skips) whenever a clustering job holds a pipeline lock to
+// avoid concurrent mutation of hot event/embedding rows. Convex interval crons
+// are epoch-phase-aligned, so an *hourly* cadence is an exact multiple of the
+// 20-min merge and 30-min recluster cadences and fired in lockstep with them
+// every single time — guaranteeing a blocking lock and a 100% skip rate. A
+// 53-minute cadence is coprime with 20/30/40/60, so archive drifts across
+// phases and regularly lands in quiet windows without ever starving the core
+// clustering pipeline (which keeps priority).
 crons.interval(
   "archive-stale-singleton-events",
-  { hours: 1 },
+  { minutes: 53 },
   internal.singletonCleanup.archiveStaleSingletonEvents,
   {},
 );

packages/backend/convex/feeds.ts

@@ -90,10 +90,14 @@ const TIER_1: FeedDefinition[] = [
     mbfc: { category: "left-center", factual: "very-high", credibility: "high" },
   },
   {
-    // NOTE: endpoint returned 520/522 at verification time while the homepage
-    // was healthy — kept because Agerpres is the national wire service; the
-    // ingestion quarantine covers it if the RSS endpoint stays broken.
-    url: "https://www.agerpres.ro/rss/stiri",
+    // Agerpres has no native RSS: www.agerpres.ro/rss/stiri 301-redirects to an
+    // allorigins.win proxy wrapper that returns 520/522 (dead), which kept the
+    // feed permanently quarantined. The underlying bazqux "createfeed" extractor
+    // over the Agerpres widget returns valid RSS directly, so we point at it
+    // without the broken proxy layer. If this third-party extractor ever fails,
+    // the ingestion quarantine covers it. Kept because Agerpres is the national
+    // wire service.
+    url: "https://createfeed.bazqux.com/extract.php?url=https%3A%2F%2Fagerpres.ro%2Fwidget&max=50&order=document&guid=0",
     name: "Agerpres",
     domain: "agerpres.ro",
     tier: 1,

packages/backend/convex/pipeline.ts

@@ -952,6 +952,13 @@ export const checkPipelineAlerts = internalAction({
 
       const byJob = new Map<string, { total: number; ok: number }>();
       for (const log of logs as Array<Doc<"pipelineRunLogs">>) {
+        // A "skipped" run is a deliberate no-op (e.g. a job yielding to a
+        // pipeline lock, or short-circuiting because it was already running).
+        // It did no work, so it must not count as a failure toward the
+        // error-rate SLO — otherwise jobs that intentionally yield (like the
+        // stale-singleton archive) flap below 80%. The "job stuck skipping"
+        // failure mode is still caught by the per-job absent-ok-run checks.
+        if (log.status === "skipped") continue;
         const row = byJob.get(log.jobName) ?? { total: 0, ok: 0 };
         row.total++;
         if (log.status === "ok") row.ok++;