Stop routine excerpt fallbacks from degrading every run

Every issue since 2026-08-15 landed as `degraded`, because the extraction
stage warned on any fetch failure at all and a few always fail: the worst
of the 21 published days was 15% (33 of 220 articles), the median under
11%. A status every run carries says nothing, so warn only past a 30%
share -- roughly twice the worst day seen -- and log the rest. The exact
count was already in `counts.excerpt_only` either way.

`rebuild()` also stored the bumped profile version before reading the
interests OPML, so a rebuild that failed on a missing file would mark the
profile fresh for another week having never rewritten its text. Read the
prompt inputs first, ahead of both the model call and the writes, so a
failure stays due and costs nothing.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0164rQrMUZkf7uV1VGYuCFy4

src/curate/profile/mod.rs

@@ -432,6 +432,11 @@ pub async fn rebuild(
     profile_path: &Path,
     verdict_limit: usize,
 ) -> anyhow::Result<TasteProfile> {
+    // Read the prompt inputs first: a rebuild that dies on a missing OPML must
+    // stay due and must not have spent a model call getting there.
+    let opml = parse_interests(opml_path)?;
+    let profile_file = load_profile(profile_path)?;
+    let interests = union_interests(opml, profile_file.interests.clone());
     let ratings = db.current_ratings(RATINGS_LOOKBACK_DAYS).await?;
     let previous = db.kv_get(KV_LEARNED_ADJUSTMENTS).await?.unwrap_or_default();
     let learned = if ratings.is_empty() {
@@ -464,21 +469,17 @@ pub async fn rebuild(
     let built_at = Timestamp::now();
     store_version(db, next_version, built_at).await?;
 
-    let opml = parse_interests(opml_path)?;
-    let profile_file = load_profile(profile_path)?;
-    let interests = union_interests(opml, profile_file.interests.clone());
-    let current = db.current_ratings(RATINGS_LOOKBACK_DAYS).await?;
     let profile = TasteProfile {
         text: build(
             &profile_file.body,
             &interests,
             &learned,
-            &current,
+            &ratings,
             verdict_limit,
         ),
         version: next_version,
         built_at,
-        verdicts: current.len().min(verdict_limit),
+        verdicts: ratings.len().min(verdict_limit),
     };
     db.kv_set(KV_TASTE_PROFILE, &profile.text).await?;
     tracing::info!(
@@ -686,4 +687,43 @@ mod tests {
         assert_eq!(backend.calls(), 1);
         assert!(backend.prompts()[0].user.contains("LOVED | A deep report"));
     }
+
+    #[tokio::test]
+    async fn a_failed_rebuild_leaves_the_version_alone_and_spends_nothing() {
+        use std::sync::Arc;
+
+        use super::super::llm::{MockBackend, UsageMeter};
+        use crate::config::ProviderConfig;
+
+        let dir = tempfile::tempdir().unwrap();
+        let db = Db::open_and_migrate(&dir.path().join("profile.db"))
+            .await
+            .unwrap();
+        let opml = dir.path().join("interests.opml");
+        let profile_path = dir.path().join("profile.md");
+        std::fs::write(&opml, r#"<outline text="Rust"/>"#).unwrap();
+        std::fs::write(&profile_path, "# Reader profile\n\nLikes depth.\n").unwrap();
+        let initial = load_or_build(&db, &opml, &profile_path, 60).await.unwrap();
+        assert_eq!(initial.version, 1);
+
+        // The OPML goes missing the way a relative path does under a service
+        // whose working directory is not the checkout.
+        std::fs::remove_file(&opml).unwrap();
+
+        let backend = Arc::new(MockBackend::new());
+        let llm = LlmClient::with_backend(
+            "deepseek-v4-flash",
+            initial.text,
+            UsageMeter::for_provider(&ProviderConfig::deepseek()),
+            backend.clone(),
+        );
+        let error = rebuild(&db, &llm, &opml, &profile_path, 60)
+            .await
+            .expect_err("a missing OPML fails the rebuild");
+        assert!(format!("{error:#}").contains("reading the interests OPML"));
+
+        // Still version 1, so the profile stays stale and the rebuild is retried.
+        assert_eq!(stored_version(&db).await.unwrap().unwrap().0, 1);
+        assert_eq!(backend.calls(), 0);
+    }
 }

src/pipeline.rs

@@ -346,6 +346,11 @@ struct StageContext<'a> {
     rescore: bool,
 }
 
+/// Share of the day's articles that must fall back to a feed excerpt before the
+/// run is degraded over it. Single-digit percentages are routine; the exact count
+/// is always in `counts.excerpt_only`.
+const EXCERPT_FALLBACK_WARN_SHARE: f64 = 0.30;
+
 async fn run_stages(
     ctx: &StageContext<'_>,
     window_start: Timestamp,
@@ -412,11 +417,24 @@ async fn run_stages(
     let extract_stats = extractor.extract_all(&mut articles).await;
     report.counts.extracted = (extract_stats.from_miniflux + extract_stats.from_readability) as i64;
     report.counts.excerpt_only = extract_stats.excerpt_only as i64;
+    // A handful of fetch failures is the normal state of the open web, so only a
+    // day well past the usual rate is worth degrading the run over.
     if extract_stats.fetch_failures > 0 {
-        report.warn(format!(
-            "{} articles fell back to a feed excerpt",
-            extract_stats.fetch_failures
-        ));
+        let share = extract_stats.fetch_failures as f64 / articles.len().max(1) as f64;
+        if share >= EXCERPT_FALLBACK_WARN_SHARE {
+            report.warn(format!(
+                "{} of {} articles ({:.0}%) fell back to a feed excerpt",
+                extract_stats.fetch_failures,
+                articles.len(),
+                share * 100.0
+            ));
+        } else {
+            tracing::info!(
+                failures = extract_stats.fetch_failures,
+                articles = articles.len(),
+                "articles fell back to a feed excerpt"
+            );
+        }
     }
     report.timings.record("extract", elapsed_ms(stage));