Merge curation-v2-step3 into curation-v2 (steps 2 and 3 united)

Union merge of the Claude editor (step 2) and the embeddings/signals/telemetry
work (step 3): both config trees, per-provider LLM budgets plus the Voyage
meter, one startup provider log, runs.config_json with ranking and voyage
settings, and selected picks' why lines landing in candidate_runs.editor_why.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe

README.md

@@ -15,10 +15,10 @@ publishes the lot over its own OPDS catalog — which doubles as a
 Each article chapter ends with Loved it / Good / Not for me links that feed back into tomorrow's curation.
 
 Steady-state cost is roughly **$1/day**: $0.05–0.30 in DeepSeek tokens plus
-~$0.50–0.80 for the Claude editor, each with its own per-UTC-day ceiling
+~$0.50–0.80 for the Claude editor and a few cents of Voyage AI embeddings, each
-(`max_daily_usd` and `anthropic.max_daily_usd`). Those ceilings are runaway
+with its own per-UTC-day ceiling (`max_daily_usd`, `anthropic.max_daily_usd` and
-guards, not accounting — set hard spend limits in both providers' dashboards as
+`voyage.max_daily_usd`). Those ceilings are runaway guards, not accounting — set
-the real backstop.
+hard spend limits in the providers' dashboards as the real backstop.
 
 - Full design: [`docs/plans/2026-08-15-the-daily-epub.md`](docs/plans/2026-08-15-the-daily-epub.md)
 - Implementation decisions: [`docs/plans/2026-08-15-implementation-notes.md`](docs/plans/2026-08-15-implementation-notes.md)
@@ -29,7 +29,8 @@ the real backstop.
 
 ```
 Miniflux ingest ─▶ dedupe ─▶ extraction ─▶ persist ─▶ social enrichment
-  ─▶ pre-filter ─▶ scoring (DeepSeek) ─▶ editor (Claude) ─▶ comments ─▶ editorial (Claude)
+  ─▶ hygiene ─▶ embeddings (Voyage) + cheap signals ─▶ pre-filter ─▶ scoring (DeepSeek)
+  ─▶ editor (Claude) ─▶ comments ─▶ editorial (Claude)
   ─▶ world briefing ─▶ EPUB (standard + X4) ─▶ XTC ─▶ publish ─▶ report
 ```
 
@@ -57,6 +58,7 @@ fallback (`fallbacks = "default"`) is enabled on every editor request.
 | **Miniflux** with an API key | the only content source | Settings → API Keys. The client is read-only and never mutates read state. |
 | **DeepSeek API key** | scoring, and the fallback for every editor call | <https://platform.deepseek.com>. Optional: `--skip-llm` runs the whole pipeline without it. |
 | **Anthropic API key** | the editor: selection, summaries, The Brief, the weekly profile rebuild | <https://console.anthropic.com>. Optional: without it every editor call runs on DeepSeek. Set a dashboard spend limit; `anthropic.max_daily_usd` is only a runaway guard. |
+| **Voyage AI API key** | article and interest embeddings behind the learned ranking signals | <https://www.voyageai.com>. Optional: without it (or with `--skip-embeddings`) the run uses cached vectors only and the learned signals are absent, never a penalty. |
 | A 32+ byte random secret | signs the article rating links | `openssl rand -hex 32` |
 | **BookOrbit** library + watched folder | *optional* — a richer library UI on top of the same folder | Delivery does not need it: `daily-epub serve` has its own OPDS catalog over `publish.epub_dir`. If you do run it, create a dedicated "The Daily EPUB" library, enable *Watch folders*, and point `publish.epub_dir` at it. |
 | **Node.js 18+** and a clone of [`epub-to-xtc-converter`](https://github.com/bigbag/epub-to-xtc-converter) | XTC/XTCH output for the Xteink X4 | Optional (`xtc.enabled = false` turns it off). Needs `npm install` **inside `cli/`**, and a settings JSON naming a real TTF/OTF — see below. It has **no global npm bin** — it is invoked as `node <repo>/cli/index.js convert …`, which is why `xtc.command`/`xtc.args` are fully general. |
@@ -79,12 +81,16 @@ sudo install -m0755 target/release/daily-epub /usr/local/bin/
 ### Commands
 
 ```
-daily-epub generate [--date YYYY-MM-DD] [--dry-run] [--out DIR] [--max-articles N] [--skip-llm]
+daily-epub generate [--date YYYY-MM-DD] [--dry-run] [--out DIR] [--max-articles N] [--skip-llm] [--skip-embeddings]
 daily-epub serve                # rating endpoints + OPDS catalog + downloads
 daily-epub profile rebuild      # regenerate learned profile adjustments
 daily-epub ratings list --days 90
 daily-epub ratings set --article 42 --label loved --note "excellent"
 daily-epub ratings clear --url https://example.com/article
+daily-epub explain --date YYYY-MM-DD (--article ID | --url URL) [--run-id N]
+daily-epub explain --date YYYY-MM-DD --near-misses [N]
+daily-epub features backfill [--days 30] [--rated-only] [--all] [--yes]
+daily-epub features prune       # stale embeddings + old candidate telemetry
 daily-epub backfill-social      # re-poll social scores for recent articles
 daily-epub db migrate           # run migrations (also automatic on every start)
 ```
@@ -94,6 +100,25 @@ articles, curates and **builds both EPUBs into `--out`**, but it does not copy t
 BookOrbit, does not run the retention sweep, does not write the `issues` row and
 does not advance the ingest watermark. It prints the lineup and the cost report.
 
+`--skip-embeddings` reads the embedding cache but makes zero Voyage calls.
+
+`explain` answers "why was this (not) in the paper" from the `candidate_runs`
+row the run persisted for every considered article: the stage it reached and the
+reason it stopped, every raw and normalized signal with its presence and
+effective weight, the top interests, the nearest rated neighbours, any cached
+LLM assessments, and the editor's reason for a pick. `--url` canonicalizes the
+address; an article that is not in the database at all is reported as never
+ingested (a feed problem, not a ranking one). `--near-misses` lists the highest
+ranked articles that were not selected.
+
+`features backfill` embeds the rated and published articles first (the learned
+set), then the standing interests, then — only with `--all` — every other
+article first seen in the window. It prints an estimate and asks before spending
+more than 5M tokens unless `--yes`; a warm cache makes zero calls. `features
+prune` drops embeddings of articles neither rated nor published that are older
+than `curation.ranking.embedding_retention_days`, and `candidate_runs` rows of
+runs older than `curation.ranking.telemetry_retention_days`.
+
 ---
 
 ## Configuration
@@ -152,6 +177,15 @@ Secrets belong in the environment file, never in the TOML.
 | `anthropic.price_output_per_mtok` | `25.0` | USD per 1M output tokens. |
 | `anthropic.max_daily_usd` | `3.0` | Claude ceiling per UTC day; tripping it moves the remaining editor work to DeepSeek. |
 | `anthropic.max_concurrent_requests` | `4` | Reserved for the parallel editor stages. |
+| `voyage.enabled` | `true` | Embed articles and interests with Voyage AI. `false` ⇒ cached vectors only. |
+| `voyage.base_url` | `https://api.voyageai.com/v1` | `POST {base_url}/embeddings`. |
+| `voyage.model` | `voyage-4-lite` | Embedding model; changing it invalidates the cache. |
+| `voyage.api_key` | — | **`DAILY_EPUB_VOYAGE__API_KEY`**. Absent ⇒ cached vectors only. |
+| `voyage.output_dimension` | `512` | One of 256, 512, 1024, 2048. |
+| `voyage.batch_size` | `32` | Texts per request. |
+| `voyage.max_concurrent_requests` | `4` | Requests in flight. |
+| `voyage.max_input_chars` | `60000` | Per-article cut, on a char boundary. |
+| `voyage.max_daily_usd` | `0.50` | Runaway guard at $0.02/M tokens. |
 | `curation.max_article_count` | `28` | Hard ceiling on issue size. `--max-articles N` lowers it to `min(28, N)` and drags the soft target down with it. Must be ≥ `target_article_count`. |
 | `curation.always_include_feeds` | `[]` | Miniflux feed ids or URL substrings that can never be dropped. |
 | `curation.blocked_domains` | `[]` | Hosts excluded outright. |
@@ -161,6 +195,7 @@ Secrets belong in the environment file, never in the TOML.
 | `curation.feedback.good_value` | `0.35` | Weight for a Good verdict. |
 | `curation.feedback.not_for_me_value` | `-1.0` | Weight for a Not for me verdict. |
 | `curation.feedback.verdicts_in_prompt` | `60` | Recent explicit verdicts included in the system prompt. |
+| `curation.ranking.*` | see below | Every weight, quota, gate and threshold of the personalized ranker. |
 | `editorial.summary_model` | `editor` | `editor` (Claude) or `bulk` (DeepSeek) for the per-article summaries. |
 | `editorial.summary_input_tokens` | `3000` | Article text offered to the summary prompt. |
 | `publish.epub_dir` | `/srv/bookorbit/libraries/daily-epub` | Both EPUB editions land here by atomic copy, and this is the directory the OPDS feed lists. The editions are distinguished by a `(X4)` tag in **both** the filename and `dc:title` — libraries and OPDS clients list books by title, so the filename alone would make them look identical. Point a BookOrbit watched folder at it if you want its UI too. **Renamed from `bookorbit_dir`**; the old key is a hard config error. |
@@ -175,6 +210,27 @@ Secrets belong in the environment file, never in the TOML.
 | `server.hmac_secret` | — | **`DAILY_EPUB_SERVER__HMAC_SECRET`** (or `DAILY_EPUB_SECRET`). Without it, generated links are rejected with 403. |
 | `server.basic_auth_user` / `_pass` | unset | Optional Basic auth for `/opds/*` and `/files/*`. |
 
+`[curation.ranking]` holds the ranker's tunables. The learned signals are
+gated: `knn` (rated-neighbour preference) ramps from `knn_floor` (8) to
+`knn_full` (25) rated articles with embeddings, `feed` (feed affinity) from
+`feed_floor` (15) to `feed_full` (40) attributable ratings; below the floor the
+signal is absent. Ratings decay with `rating_half_life_days` (60) over
+`rating_lookback_days` (180); `neighbour_k` (5) neighbours per side and
+`negative_coefficient` (0.75) shape the signal. `triage_max` (800),
+`deep_keep` (120), `shortlist_keep` (60), `assessment_reuse_days` (3),
+`semantic_min_words` (300), `exploration_slots` (5), `[curation.ranking.quotas]`
+(`triage` 60 · `interest` 20 · `knn` 20), `[curation.ranking.weights.utility]`
+and `[curation.ranking.diversity]` (`cluster_threshold` 0.85, `per_cluster_cap`
+2, `utility_protected` 10) are validated now and drive the LLM triage, deep
+assessment and diversification stages as they land.
+`[curation.ranking.weights.preliminary]` (`interest` 0.35 · `knn` 0.25 ·
+`heuristic` 0.20 · `feed` 0.10 · `social` 0.10) blends the cheap signals; weights
+are renormalized over the signals present for each article, so they need not sum
+to 1. `embedding_retention_days` (120) and `telemetry_retention_days` (180) are
+what `features prune` enforces. Validation: weights non-negative; `deep_keep ≥
+shortlist_keep ≥ target_article_count`; `*_full > *_floor`; `0 ≤
+cluster_threshold ≤ 1`; `per_cluster_cap ≥ 1`; Voyage batch size and concurrency ≥ 1.
+
 ---
 
 ## Deployment (systemd)
@@ -199,6 +255,7 @@ sudo tee /etc/daily-epub/env >/dev/null <<EOF
 DAILY_EPUB_MINIFLUX__API_KEY=…
 DAILY_EPUB_DEEPSEEK__API_KEY=…
 DAILY_EPUB_ANTHROPIC__API_KEY=…
+DAILY_EPUB_VOYAGE__API_KEY=…
 DAILY_EPUB_SERVER__HMAC_SECRET=$(openssl rand -hex 32)
 EOF
 sudo chown daily-epub:daily-epub /etc/daily-epub/env && sudo chmod 0600 /etc/daily-epub/env
@@ -440,7 +497,8 @@ server. The stages themselves:
 
 ```text
 miniflux.rs   ingest            curate/       scoring and selection
-dedupe.rs     clustering          prefilter, llm, score, select, editorial
+dedupe.rs     clustering          prefilter, llm, score, select, editorial,
+                                  embedding, signals, telemetry
 extract.rs    body text           profile/    the reader's taste profile
 images/       article images    comments.rs   discussion chapters
   normalize     usable <img>    world.rs      the world briefing
@@ -511,10 +569,15 @@ From spec §7, plus what implementation turned up:
   stack, so `curate/llm.rs` speaks the OpenAI-compatible wire protocol over the
   shared `reqwest` client instead, behind a `ChatBackend` trait. The dependency
   was removed.
-- **Two providers are wired**, DeepSeek (bulk) and Anthropic (editor), each a
+- **Two chat providers are wired**, DeepSeek (bulk) and Anthropic (editor),
-  `ChatBackend` impl with its own `UsageMeter` and price table. A third means
+  each a `ChatBackend` impl with its own `UsageMeter` and price table. A third
-  another impl.
+  means another impl. Voyage AI embeddings sit behind the analogous
-- **No embedding-based personal ranker yet** (spec §3.9 future work); the schema
+  `EmbeddingBackend` trait in `curate/embedding.rs`.
-  is ready for it once ~200 ratings exist.
+- **The learned signals are computed but do not yet gate selection.** Every
+  eligible article gets interest, rated-neighbour, feed-affinity, social and
+  heuristic signals persisted to `candidate_runs.signals_json` (read them with
+  `explain`), while the heuristic pre-filter still decides what the LLM sees.
+  The rated-neighbour and feed signals stay absent until their gates open
+  (8 and 15 ratings respectively).
 - **One reader, one issue per day.** There is no multi-user support and no
   weekly/retrospective edition (spec §6).

config.example.toml

@@ -5,6 +5,7 @@
 #   DAILY_EPUB_MINIFLUX__API_KEY=...
 #   DAILY_EPUB_DEEPSEEK__API_KEY=...
 #   DAILY_EPUB_ANTHROPIC__API_KEY=...
+#   DAILY_EPUB_VOYAGE__API_KEY=...
 #   DAILY_EPUB_SERVER__HMAC_SECRET=...
 #   DAILY_EPUB_LOOKBACK_HOURS=30
 
@@ -14,7 +15,7 @@ target_article_count = 20
 prefilter_keep = 120
 retention_days = 21                  # EPUBs, by age
 xtc_retention_count = 5              # XTC issues, by count (~80-100 MB each)
-max_daily_usd = 2.0                  # DeepSeek ceiling per UTC day; [anthropic] has its own
+max_daily_usd = 2.0                  # DeepSeek ceiling per UTC day; [anthropic] and [voyage] have their own
 world_briefing = true
 
 # SQLite database file. Parent directories are created on demand.
@@ -63,6 +64,20 @@ price_output_per_mtok = 25.0
 max_daily_usd = 3.0
 max_concurrent_requests = 4
 
+# Voyage AI embeddings behind the interest and rated-neighbour signals. Set
+# `enabled = false` (or leave the key unset) and the paper still builds: the
+# learned signals are simply absent, never a penalty.
+[voyage]
+enabled = true
+base_url = "https://api.voyageai.com/v1"
+model = "voyage-4-lite"
+# api_key via DAILY_EPUB_VOYAGE__API_KEY env
+output_dimension = 512               # 256 | 512 | 1024 | 2048
+batch_size = 32
+max_concurrent_requests = 4
+max_input_chars = 60000              # per article, cut on a char boundary
+max_daily_usd = 0.50                 # runaway guard ($0.02 / M tokens)
+
 [curation]
 max_article_count = 28               # hard ceiling; there is no minimum (§13)
 always_include_feeds = []            # miniflux feed ids or site urls
@@ -87,6 +102,55 @@ good_value = 0.35
 not_for_me_value = -1.0
 verdicts_in_prompt = 60
 
+# Every weight, quota, gate and threshold of the personalized ranker. The
+# learned signals (`knn`, `feed`) contribute nothing until their gates open:
+# the weight ramps linearly from `*_floor` to `*_full` rated articles.
+[curation.ranking]
+triage_max = 800                     # eligible articles the triage LLM reads
+deep_keep = 120                      # deep-assessment set
+shortlist_keep = 60                  # what the editor sees
+assessment_reuse_days = 3
+rating_lookback_days = 180
+rating_half_life_days = 60
+neighbour_k = 5
+negative_coefficient = 0.75
+knn_floor = 8
+knn_full = 25
+feed_floor = 15
+feed_full = 40
+semantic_min_words = 300
+exploration_slots = 5
+embedding_retention_days = 120       # `features prune`: unrated, unpublished vectors
+telemetry_retention_days = 180       # `features prune`: candidate_runs rows
+
+[curation.ranking.quotas]
+triage = 60
+interest = 20
+knn = 20
+
+# Weights need not sum to 1; they are renormalized over the present signals.
+[curation.ranking.weights.preliminary]
+interest = 0.35
+knn = 0.25
+heuristic = 0.20
+feed = 0.10
+social = 0.10
+
+[curation.ranking.weights.utility]
+quality = 0.40
+fit = 0.20
+knn = 0.15
+interest = 0.10
+feed = 0.05
+triage = 0.05
+social = 0.03
+heuristic = 0.02
+
+[curation.ranking.diversity]
+cluster_threshold = 0.85
+per_cluster_cap = 2
+utility_protected = 10
+
 [editorial]
 summary_model = "editor"             # editor (Claude) | bulk (DeepSeek)
 summary_input_tokens = 3000          # article text offered per summary

src/config.rs

@@ -75,6 +75,7 @@ pub struct Config {
     pub miniflux: MinifluxConfig,
     pub deepseek: DeepseekConfig,
     pub anthropic: AnthropicConfig,
+    pub voyage: VoyageConfig,
     pub curation: CurationConfig,
     pub editorial: EditorialConfig,
     pub publish: PublishConfig,
@@ -100,6 +101,7 @@ impl Default for Config {
             miniflux: MinifluxConfig::default(),
             deepseek: DeepseekConfig::default(),
             anthropic: AnthropicConfig::default(),
+            voyage: VoyageConfig::default(),
             curation: CurationConfig::default(),
             editorial: EditorialConfig::default(),
             publish: PublishConfig::default(),
@@ -229,6 +231,38 @@ impl Default for EditorialConfig {
     }
 }
 
+/// `[voyage]` — embedding endpoint and cache shape (§4.3).
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields, default)]
+pub struct VoyageConfig {
+    pub enabled: bool,
+    pub base_url: String,
+    pub model: String,
+    /// Supply via `DAILY_EPUB_VOYAGE__API_KEY`; never put it in the TOML.
+    pub api_key: Option<String>,
+    pub output_dimension: usize,
+    pub batch_size: usize,
+    pub max_concurrent_requests: usize,
+    pub max_input_chars: usize,
+    pub max_daily_usd: f64,
+}
+
+impl Default for VoyageConfig {
+    fn default() -> Self {
+        Self {
+            enabled: true,
+            base_url: "https://api.voyageai.com/v1".into(),
+            model: "voyage-4-lite".into(),
+            api_key: None,
+            output_dimension: 512,
+            batch_size: 32,
+            max_concurrent_requests: 4,
+            max_input_chars: 60_000,
+            max_daily_usd: 0.50,
+        }
+    }
+}
+
 /// `[curation]` — pre-filter and section palette (§3.5, §3.6).
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
 #[serde(deny_unknown_fields, default)]
@@ -245,6 +279,7 @@ pub struct CurationConfig {
     /// The only section names the LLM may use (§3.6 stage B).
     pub sections: Vec<String>,
     pub feedback: FeedbackConfig,
+    pub ranking: RankingConfig,
 }
 
 impl Default for CurationConfig {
@@ -268,6 +303,154 @@ impl Default for CurationConfig {
             .map(|s| s.to_string())
             .collect(),
             feedback: FeedbackConfig::default(),
+            ranking: RankingConfig::default(),
+        }
+    }
+}
+
+/// `[curation.ranking]` — every weight, quota, gate and threshold of the
+/// personalized ranker (plan §19). Steps 4–5 consume most of these; step 3
+/// uses the learned-signal gates, the preliminary weights and the retention
+/// windows.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields, default)]
+pub struct RankingConfig {
+    pub triage_max: usize,
+    pub deep_keep: usize,
+    pub shortlist_keep: usize,
+    pub assessment_reuse_days: i64,
+    pub rating_lookback_days: i64,
+    pub rating_half_life_days: f64,
+    pub neighbour_k: usize,
+    pub negative_coefficient: f64,
+    pub knn_floor: usize,
+    pub knn_full: usize,
+    pub feed_floor: usize,
+    pub feed_full: usize,
+    pub semantic_min_words: i64,
+    pub exploration_slots: usize,
+    pub embedding_retention_days: i64,
+    pub telemetry_retention_days: i64,
+    pub quotas: RankingQuotas,
+    pub weights: RankingWeights,
+    pub diversity: DiversityConfig,
+}
+
+impl Default for RankingConfig {
+    fn default() -> Self {
+        Self {
+            triage_max: 800,
+            deep_keep: 120,
+            shortlist_keep: 60,
+            assessment_reuse_days: 3,
+            rating_lookback_days: 180,
+            rating_half_life_days: 60.0,
+            neighbour_k: 5,
+            negative_coefficient: 0.75,
+            knn_floor: 8,
+            knn_full: 25,
+            feed_floor: 15,
+            feed_full: 40,
+            semantic_min_words: 300,
+            exploration_slots: 5,
+            embedding_retention_days: 120,
+            telemetry_retention_days: 180,
+            quotas: RankingQuotas::default(),
+            weights: RankingWeights::default(),
+            diversity: DiversityConfig::default(),
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields, default)]
+pub struct RankingQuotas {
+    pub triage: usize,
+    pub interest: usize,
+    pub knn: usize,
+}
+
+impl Default for RankingQuotas {
+    fn default() -> Self {
+        Self {
+            triage: 60,
+            interest: 20,
+            knn: 20,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields, default)]
+pub struct RankingWeights {
+    pub preliminary: PreliminaryWeights,
+    pub utility: UtilityWeights,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields, default)]
+pub struct PreliminaryWeights {
+    pub interest: f64,
+    pub knn: f64,
+    pub heuristic: f64,
+    pub feed: f64,
+    pub social: f64,
+}
+
+impl Default for PreliminaryWeights {
+    fn default() -> Self {
+        Self {
+            interest: 0.35,
+            knn: 0.25,
+            heuristic: 0.20,
+            feed: 0.10,
+            social: 0.10,
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields, default)]
+pub struct UtilityWeights {
+    pub quality: f64,
+    pub fit: f64,
+    pub knn: f64,
+    pub interest: f64,
+    pub feed: f64,
+    pub triage: f64,
+    pub social: f64,
+    pub heuristic: f64,
+}
+
+impl Default for UtilityWeights {
+    fn default() -> Self {
+        Self {
+            quality: 0.40,
+            fit: 0.20,
+            knn: 0.15,
+            interest: 0.10,
+            feed: 0.05,
+            triage: 0.05,
+            social: 0.03,
+            heuristic: 0.02,
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields, default)]
+pub struct DiversityConfig {
+    pub cluster_threshold: f64,
+    pub per_cluster_cap: usize,
+    pub utility_protected: usize,
+}
+
+impl Default for DiversityConfig {
+    fn default() -> Self {
+        Self {
+            cluster_threshold: 0.85,
+            per_cluster_cap: 2,
+            utility_protected: 10,
         }
     }
 }
@@ -479,6 +662,70 @@ impl Config {
                 "anthropic.effort must be one of low, medium, high, xhigh, max".into(),
             ));
         }
+        let ranking = &self.curation.ranking;
+        if ranking.deep_keep < ranking.shortlist_keep
+            || ranking.shortlist_keep < self.target_article_count
+        {
+            return Err(ConfigError::Invalid(
+                "curation.ranking must satisfy deep_keep >= shortlist_keep >= target_article_count"
+                    .into(),
+            ));
+        }
+        if ranking.knn_full <= ranking.knn_floor || ranking.feed_full <= ranking.feed_floor {
+            return Err(ConfigError::Invalid(
+                "curation.ranking *_full must be > *_floor >= 0".into(),
+            ));
+        }
+        if !(0.0..=1.0).contains(&ranking.diversity.cluster_threshold) {
+            return Err(ConfigError::Invalid(
+                "curation.ranking.diversity.cluster_threshold must be between 0 and 1".into(),
+            ));
+        }
+        if ranking.diversity.per_cluster_cap == 0 {
+            return Err(ConfigError::Invalid(
+                "curation.ranking.diversity.per_cluster_cap must be >= 1".into(),
+            ));
+        }
+        let preliminary = &ranking.weights.preliminary;
+        let utility = &ranking.weights.utility;
+        let weights = [
+            preliminary.interest,
+            preliminary.knn,
+            preliminary.heuristic,
+            preliminary.feed,
+            preliminary.social,
+            utility.quality,
+            utility.fit,
+            utility.knn,
+            utility.interest,
+            utility.feed,
+            utility.triage,
+            utility.social,
+            utility.heuristic,
+        ];
+        if weights
+            .iter()
+            .any(|weight| !weight.is_finite() || *weight < 0.0)
+        {
+            return Err(ConfigError::Invalid(
+                "curation.ranking weights must be finite and non-negative".into(),
+            ));
+        }
+        if self.voyage.batch_size == 0 || self.voyage.max_concurrent_requests == 0 {
+            return Err(ConfigError::Invalid(
+                "voyage.batch_size and voyage.max_concurrent_requests must be >= 1".into(),
+            ));
+        }
+        if ![256, 512, 1024, 2048].contains(&self.voyage.output_dimension) {
+            return Err(ConfigError::Invalid(
+                "voyage.output_dimension must be one of 256, 512, 1024, 2048".into(),
+            ));
+        }
+        if ranking.rating_half_life_days <= 0.0 || !ranking.rating_half_life_days.is_finite() {
+            return Err(ConfigError::Invalid(
+                "curation.ranking.rating_half_life_days must be > 0".into(),
+            ));
+        }
         if self.curation.sections.is_empty() {
             return Err(ConfigError::Invalid(
                 "curation.sections must not be empty".into(),
@@ -545,8 +792,12 @@ mod tests {
             jail.set_env("DAILY_EPUB_MINIFLUX__API_KEY", "secret-token");
             jail.set_env("DAILY_EPUB_TARGET_ARTICLE_COUNT", "12");
             jail.set_env("DAILY_EPUB_SERVER__HMAC_SECRET", "hunter2");
+            jail.set_env("DAILY_EPUB_VOYAGE__API_KEY", "voyage-key");
+            jail.set_env("DAILY_EPUB_VOYAGE__ENABLED", "false");
 
             let c = Config::load(None).map_err(|e| figment::Error::from(e.to_string()))?;
+            assert_eq!(c.voyage.api_key.as_deref(), Some("voyage-key"));
+            assert!(!c.voyage.enabled);
             // from file
             assert_eq!(c.lookback_hours, 30);
             assert!(!c.world_briefing);
@@ -639,6 +890,70 @@ mod tests {
         assert!(c.validate().is_err());
     }
 
+    #[test]
+    fn voyage_and_ranking_defaults_and_validation() {
+        let cfg = Config::default();
+        assert!(cfg.voyage.enabled);
+        assert_eq!(cfg.voyage.base_url, "https://api.voyageai.com/v1");
+        assert_eq!(cfg.voyage.model, "voyage-4-lite");
+        assert_eq!(cfg.voyage.output_dimension, 512);
+        assert_eq!(cfg.voyage.batch_size, 32);
+        assert_eq!(cfg.voyage.max_concurrent_requests, 4);
+        assert_eq!(cfg.voyage.max_input_chars, 60_000);
+        assert_eq!(cfg.voyage.max_daily_usd, 0.50);
+        let ranking = &cfg.curation.ranking;
+        assert_eq!(
+            (
+                ranking.triage_max,
+                ranking.deep_keep,
+                ranking.shortlist_keep
+            ),
+            (800, 120, 60)
+        );
+        assert_eq!((ranking.knn_floor, ranking.knn_full), (8, 25));
+        assert_eq!((ranking.feed_floor, ranking.feed_full), (15, 40));
+        assert_eq!(ranking.rating_half_life_days, 60.0);
+        assert_eq!(ranking.negative_coefficient, 0.75);
+        assert_eq!(ranking.weights.preliminary.interest, 0.35);
+        assert_eq!(ranking.weights.utility.quality, 0.40);
+        assert_eq!(ranking.diversity.per_cluster_cap, 2);
+        assert_eq!(ranking.embedding_retention_days, 120);
+        assert_eq!(ranking.telemetry_retention_days, 180);
+        cfg.validate().unwrap();
+
+        let mut bad = Config::default();
+        bad.voyage.output_dimension = 300;
+        assert!(bad.validate().is_err(), "dimension must be a Voyage size");
+        let mut bad = Config::default();
+        bad.voyage.batch_size = 0;
+        assert!(bad.validate().is_err());
+        let mut bad = Config::default();
+        bad.curation.ranking.weights.preliminary.knn = -0.1;
+        assert!(bad.validate().is_err(), "weights are non-negative");
+        let mut bad = Config::default();
+        bad.curation.ranking.knn_full = bad.curation.ranking.knn_floor;
+        assert!(bad.validate().is_err(), "*_full must exceed *_floor");
+        let mut bad = Config::default();
+        bad.curation.ranking.shortlist_keep = bad.curation.ranking.deep_keep + 1;
+        assert!(bad.validate().is_err(), "deep_keep >= shortlist_keep");
+        let mut bad = Config::default();
+        bad.curation.ranking.shortlist_keep = bad.target_article_count - 1;
+        assert!(bad.validate().is_err(), "shortlist_keep >= target");
+        let mut bad = Config::default();
+        bad.curation.ranking.diversity.cluster_threshold = 1.5;
+        assert!(bad.validate().is_err());
+        let mut bad = Config::default();
+        bad.curation.ranking.diversity.per_cluster_cap = 0;
+        assert!(bad.validate().is_err());
+
+        // Unknown keys inside a known section fail loudly.
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("config.toml");
+        std::fs::write(&path, "[voyage]\nenabled = true\nnot_a_key = 1\n").unwrap();
+        let err = Config::load(Some(&path)).expect_err("unknown voyage key must be rejected");
+        assert!(err.to_string().contains("not_a_key"), "{err}");
+    }
+
     #[test]
     fn validation_rejects_nonsense() {
         assert!(

src/curate/mod.rs

@@ -11,11 +11,14 @@
 //! client; selection and editorial on the editor with per-call bulk fallback.
 
 pub mod editorial;
+pub mod embedding;
 pub mod llm;
 pub mod prefilter;
 pub mod profile;
 pub mod score;
 pub mod select;
+pub mod signals;
+pub mod telemetry;
 
 use jiff::civil::Date;
 

src/curate/prefilter.rs

@@ -90,7 +90,7 @@ impl PrefilterContext {
         let since = today
             .checked_sub(jiff::Span::new().days(STALE_LOOKBACK_DAYS))
             .unwrap_or(today);
-        let already_published = db.previously_published_ids().await?;
+        let already_published = db.previously_published_ids_before(today).await?;
         let recently_rejected = db.recently_low_scored_ids(STALE_LOW_SCORE, since).await?;
         tracing::debug!(
             published = already_published.len(),
@@ -197,6 +197,29 @@ pub fn social_points(social_score: f64) -> f64 {
     MAX_SOCIAL_POINTS * (social_score / SOCIAL_SATURATION).min(1.0).sqrt()
 }
 
+/// Text-only heuristic used by personalized ranking (§9.3).
+pub fn text_heuristic(article: &Article) -> f64 {
+    longform_points(article.word_count)
+        - excerpt_only_penalty(article)
+        - roundup_penalty(&article.title)
+}
+
+pub fn excerpt_only_penalty(article: &Article) -> f64 {
+    if article.excerpt_only {
+        EXCERPT_ONLY_PENALTY
+    } else {
+        0.0
+    }
+}
+
+pub fn roundup_penalty(title: &str) -> f64 {
+    if looks_like_roundup(title) {
+        ROUNDUP_TITLE_PENALTY
+    } else {
+        0.0
+    }
+}
+
 /// Score one article 0–100 from word count, social proof, source signals,
 /// and the excerpt/roundup/blocklist penalties (§3.5).
 pub fn score_article(article: &Article, _ctx: &PrefilterContext, cfg: &Config) -> f64 {

src/curate/profile/mod.rs

@@ -125,6 +125,16 @@ pub fn load_profile(path: &Path) -> anyhow::Result<ProfileFile> {
     }
 }
 
+/// Load the exact standing-interest union used in the system prompt.
+pub fn load_standing_interests(
+    opml_path: &Path,
+    profile_path: &Path,
+) -> anyhow::Result<Vec<String>> {
+    let opml = parse_interests(opml_path)?;
+    let profile = load_profile(profile_path)?;
+    Ok(union_interests(opml, profile.interests))
+}
+
 fn union_interests(opml: Vec<String>, profile: Vec<String>) -> Vec<String> {
     let mut seen = BTreeSet::new();
     let mut out = Vec::new();

src/curate/signals.rs

@@ -0,0 +1,981 @@
+//! Cheap per-article ranking signals, the mid-rank percentile normalizer and the
+//! preliminary blend (plan §9, §12.2, §12.4).
+//!
+//! Every signal is an `Option<f64>`: `None` means *absent*, which is never a
+//! numeric zero. Absent signals are left out of the percentile computation and
+//! of the blend, whose remaining weights are renormalized (§12.2, §12.4).
+
+use std::collections::{BTreeMap, HashMap, HashSet};
+
+use jiff::Timestamp;
+use serde::{Deserialize, Serialize};
+
+use crate::config::{PreliminaryWeights, RankingConfig, VoyageConfig};
+use crate::curate::embedding::{dot, load_article_embeddings};
+use crate::curate::prefilter;
+use crate::db::Db;
+use crate::types::{Article, ArticleId, FeedId, SourceKind};
+
+/// Below this many embedded eligible articles the z-score is too noisy, so the
+/// interest signal falls back to the raw top-1 cosine (§9.1).
+pub const INTEREST_ZSCORE_MIN_ARTICLES: usize = 30;
+/// Standard-deviation floor for the per-interest z-score (§9.1).
+const ZSCORE_STD_FLOOR: f64 = 1e-3;
+/// How many interests and rated neighbours `signals_json` records (§7.5).
+const RECORDED_TOP: usize = 3;
+
+/// The signal names that go through the percentile normalizer, in the order
+/// they are rendered (§12.2). LLM scores (`triage`, `quality`, `fit`) are
+/// absolute and arrive in steps 4–5.
+pub const PERCENTILE_SIGNALS: [&str; 5] = ["interest", "knn", "feed", "social", "heuristic"];
+
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
+pub struct TopInterest {
+    pub name: String,
+    pub z: f64,
+    pub cos: f64,
+}
+
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
+pub struct Neighbour {
+    pub article_id: ArticleId,
+    pub label: String,
+    pub cos: f64,
+    pub title: String,
+}
+
+/// Every cheap signal for one article, plus what the normalizer and the blend
+/// derived from them (§9, §12.2, §12.4).
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
+pub struct Signals {
+    pub interest: Option<f64>,
+    /// Raw top-1 cosine behind `interest`, recorded for `explain` (§7.5).
+    pub interest_top1_cos: Option<f64>,
+    pub knn: Option<f64>,
+    pub feed: Option<f64>,
+    pub social: Option<f64>,
+    pub heuristic: Option<f64>,
+    /// Mid-rank percentiles of the present signals (§12.2).
+    #[serde(default)]
+    pub norm: BTreeMap<String, f64>,
+    /// Effective preliminary weights after gating and renormalization (§12.4).
+    #[serde(default)]
+    pub weights: BTreeMap<String, f64>,
+    #[serde(default)]
+    pub top_interests: Vec<TopInterest>,
+    #[serde(default)]
+    pub neighbours: Vec<Neighbour>,
+    #[serde(default)]
+    pub notes: Vec<String>,
+    /// Preliminary blend on a 0–100 scale; `None` when nothing is present.
+    pub preliminary: Option<f64>,
+    /// Gate ramps applied to the learned signals' weights (§9.2, §9.3).
+    #[serde(skip)]
+    pub knn_gate: f64,
+    #[serde(skip)]
+    pub feed_gate: f64,
+}
+
+impl Signals {
+    /// The raw value of a named signal, `None` when absent or unknown.
+    pub fn raw(&self, name: &str) -> Option<f64> {
+        match name {
+            "interest" => self.interest,
+            "interest_top1_cos" => self.interest_top1_cos,
+            "knn" => self.knn,
+            "feed" => self.feed,
+            "social" => self.social,
+            "heuristic" => self.heuristic,
+            _ => None,
+        }
+    }
+
+    pub fn present(&self, name: &str) -> bool {
+        self.raw(name).is_some()
+    }
+
+    /// The signals every eligible article gets without embeddings or ratings.
+    pub fn baseline(article: &Article) -> Self {
+        Self {
+            social: (!article.social.is_empty()).then(|| article.social_score()),
+            heuristic: Some(prefilter::text_heuristic(article)),
+            ..Self::default()
+        }
+    }
+}
+
+/// What the run log and the report say about the learned signals (§9.2, §15.4).
+#[derive(Debug, Clone, Copy, Default, PartialEq)]
+pub struct PreferenceSummary {
+    pub rated_with_embeddings: usize,
+    pub attributable_feed_ratings: usize,
+    pub knn_gate: f64,
+    pub feed_gate: f64,
+}
+
+/// One rated article with an embedding: the unit of the preference state (§9.2).
+#[derive(Debug, Clone, PartialEq)]
+pub struct RatedExample {
+    pub article_id: ArticleId,
+    pub label: String,
+    pub title: String,
+    /// The vote's value (`loved` 1.0, `good` 0.35, `not_for_me` −1.0).
+    pub value: f64,
+    /// `0.5 ^ (age_days / half_life_days)` at the time of the run.
+    pub decay: f64,
+    pub embedding: Vec<f32>,
+    /// Distinct direct feeds that carried the rated article (§9.3).
+    pub feeds: Vec<FeedId>,
+}
+
+impl RatedExample {
+    /// `weight_i = value_i × decay_i` (§9.2).
+    pub fn weight(&self) -> f64 {
+        self.value * self.decay
+    }
+}
+
+#[derive(Debug, Clone, Copy, Default, PartialEq)]
+struct FeedRate {
+    up: f64,
+    down: f64,
+}
+
+impl FeedRate {
+    /// Beta-smoothed rate `(up + 1) / (up + down + 2)` (§9.3).
+    fn rate(self) -> f64 {
+        (self.up + 1.0) / (self.up + self.down + 2.0)
+    }
+}
+
+/// Rated-neighbour and feed-affinity state, built once per run (§9.2, §9.3).
+#[derive(Debug, Clone, Default)]
+pub struct PreferenceState {
+    pub examples: Vec<RatedExample>,
+    feed_rates: HashMap<FeedId, FeedRate>,
+    pub attributable_feed_ratings: usize,
+    pub knn_gate: f64,
+    pub feed_gate: f64,
+}
+
+impl PreferenceState {
+    /// Build the state from already-loaded examples (pure; tests use this).
+    pub fn build(examples: Vec<RatedExample>, ranking: &RankingConfig) -> Self {
+        let (feed_rates, attributable_feed_ratings) = feed_rates(&examples);
+        Self {
+            knn_gate: gate(examples.len(), ranking.knn_floor, ranking.knn_full),
+            feed_gate: gate(
+                attributable_feed_ratings,
+                ranking.feed_floor,
+                ranking.feed_full,
+            ),
+            examples,
+            feed_rates,
+            attributable_feed_ratings,
+        }
+    }
+
+    /// Load `db::current_ratings(rating_lookback_days)` joined to
+    /// `article_embeddings`; ratings without an embedding are skipped (§9.2).
+    pub async fn load(
+        db: &Db,
+        voyage: &VoyageConfig,
+        ranking: &RankingConfig,
+        now: Timestamp,
+    ) -> anyhow::Result<Self> {
+        let ratings = db.current_ratings(ranking.rating_lookback_days).await?;
+        let ids = ratings
+            .iter()
+            .map(|rating| rating.article_id)
+            .collect::<Vec<_>>();
+        let embeddings = load_article_embeddings(db, voyage, &ids).await?;
+        let mut examples = Vec::new();
+        for rating in ratings {
+            let Some(embedding) = embeddings.get(&rating.article_id).cloned() else {
+                continue;
+            };
+            let feeds = db
+                .get_article(rating.article_id)
+                .await?
+                .as_ref()
+                .map(direct_feeds)
+                .unwrap_or_default();
+            let age_days = (now.as_second() - rating.event_at.as_second()).max(0) as f64 / 86_400.0;
+            examples.push(RatedExample {
+                article_id: rating.article_id,
+                label: rating.label,
+                title: rating.title,
+                value: rating.value,
+                decay: decay(age_days, ranking.rating_half_life_days),
+                embedding,
+                feeds,
+            });
+        }
+        Ok(Self::build(examples, ranking))
+    }
+
+    pub fn summary(&self) -> PreferenceSummary {
+        PreferenceSummary {
+            rated_with_embeddings: self.examples.len(),
+            attributable_feed_ratings: self.attributable_feed_ratings,
+            knn_gate: self.knn_gate,
+            feed_gate: self.feed_gate,
+        }
+    }
+
+    /// The once-per-run log line of §9.2.
+    pub fn log(&self, ranking: &RankingConfig) {
+        let feed_detail = if self.feed_gate > 0.0 {
+            format!("(n={})", self.attributable_feed_ratings)
+        } else {
+            format!(
+                "(n={} < {})",
+                self.attributable_feed_ratings, ranking.feed_floor
+            )
+        };
+        tracing::info!(
+            rated_with_embeddings = self.examples.len(),
+            knn_gate = self.knn_gate,
+            feed_gate = self.feed_gate,
+            "preference: {} rated articles with embeddings → knn gate {:.2}; feed gate {:.1} {}",
+            self.examples.len(),
+            self.knn_gate,
+            self.feed_gate,
+            feed_detail
+        );
+    }
+
+    /// Signed rated-neighbour preference and the three nearest rated articles
+    /// (§9.2). Absent when the gate is closed or there are no examples.
+    pub fn knn(&self, candidate: &[f32], ranking: &RankingConfig) -> (Option<f64>, Vec<Neighbour>) {
+        if self.knn_gate <= 0.0 || self.examples.is_empty() {
+            return (None, Vec::new());
+        }
+        let mut scored = self
+            .examples
+            .iter()
+            .filter_map(|example| {
+                dot(candidate, &example.embedding)
+                    .ok()
+                    .map(|s| (s, example))
+            })
+            .collect::<Vec<_>>();
+        scored.sort_by(|left, right| right.0.total_cmp(&left.0));
+
+        let side = |positive: bool| -> Option<f64> {
+            let chosen = scored
+                .iter()
+                .filter(|(_, example)| (example.weight() > 0.0) == positive)
+                .take(ranking.neighbour_k.max(1))
+                .collect::<Vec<_>>();
+            let denominator = chosen
+                .iter()
+                .map(|(_, example)| example.weight().abs())
+                .sum::<f64>();
+            (denominator > 0.0).then(|| {
+                chosen
+                    .iter()
+                    .map(|(similarity, example)| example.weight().abs() * similarity)
+                    .sum::<f64>()
+                    / denominator
+            })
+        };
+        let positive = side(true);
+        let negative = side(false);
+        let knn = (positive.is_some() || negative.is_some()).then(|| {
+            positive.unwrap_or(0.0) - ranking.negative_coefficient * negative.unwrap_or(0.0)
+        });
+        let neighbours = scored
+            .iter()
+            .take(RECORDED_TOP)
+            .map(|(cos, example)| Neighbour {
+                article_id: example.article_id,
+                label: example.label.clone(),
+                cos: *cos,
+                title: example.title.clone(),
+            })
+            .collect();
+        (knn, neighbours)
+    }
+
+    /// Mean Beta-smoothed rate over the article's rated direct feeds (§9.3).
+    pub fn feed(&self, article: &Article) -> Option<f64> {
+        if self.feed_gate <= 0.0 {
+            return None;
+        }
+        let rates = direct_feeds(article)
+            .into_iter()
+            .filter_map(|feed| self.feed_rates.get(&feed))
+            .map(|rate| rate.rate())
+            .collect::<Vec<_>>();
+        (!rates.is_empty()).then(|| rates.iter().sum::<f64>() / rates.len() as f64)
+    }
+
+    /// Per-feed `(up, down)` credit, exposed for tests of §9.3.
+    pub fn feed_credit(&self, feed: FeedId) -> Option<(f64, f64)> {
+        self.feed_rates.get(&feed).map(|rate| (rate.up, rate.down))
+    }
+}
+
+/// `0.5 ^ (age_days / half_life_days)` (§9.2).
+pub fn decay(age_days: f64, half_life_days: f64) -> f64 {
+    if half_life_days <= 0.0 {
+        return 1.0;
+    }
+    0.5f64.powf(age_days.max(0.0) / half_life_days)
+}
+
+/// `clamp((n − floor) / (full − floor), 0, 1)` (§9.2).
+pub fn gate(n: usize, floor: usize, full: usize) -> f64 {
+    if n <= floor {
+        0.0
+    } else if n >= full || full <= floor {
+        1.0
+    } else {
+        (n - floor) as f64 / (full - floor) as f64
+    }
+}
+
+/// Distinct `SourceKind::Feed` feeds that carried the article; the best entry's
+/// feed when there are none (§9.3).
+pub fn direct_feeds(article: &Article) -> Vec<FeedId> {
+    let mut feeds = article
+        .sources
+        .iter()
+        .filter(|source| source.kind == SourceKind::Feed)
+        .map(|source| source.feed_id)
+        .collect::<HashSet<_>>()
+        .into_iter()
+        .collect::<Vec<_>>();
+    if feeds.is_empty() && article.feed_id != 0 {
+        feeds.push(article.feed_id);
+    }
+    feeds.sort_unstable();
+    feeds
+}
+
+fn feed_rates(examples: &[RatedExample]) -> (HashMap<FeedId, FeedRate>, usize) {
+    let mut rates: HashMap<FeedId, FeedRate> = HashMap::new();
+    let mut attributable = 0;
+    for example in examples {
+        if example.feeds.is_empty() {
+            continue;
+        }
+        attributable += 1;
+        let credit = example.weight() / example.feeds.len() as f64;
+        for feed in &example.feeds {
+            let rate = rates.entry(*feed).or_default();
+            rate.up += credit.max(0.0);
+            rate.down += (-credit).max(0.0);
+        }
+    }
+    (rates, attributable)
+}
+
+/// The interest match of §9.1 for one article.
+#[derive(Debug, Clone, Default, PartialEq)]
+pub struct InterestMatch {
+    pub score: f64,
+    pub top1_cos: f64,
+    pub top_interests: Vec<TopInterest>,
+}
+
+/// Z-scored standing-interest match for every embedded article (§9.1).
+///
+/// Below [`INTEREST_ZSCORE_MIN_ARTICLES`] embedded articles the score is the raw
+/// top-1 cosine instead, and that is logged.
+pub fn interest_matches(
+    articles: &HashMap<ArticleId, Vec<f32>>,
+    interests: &HashMap<String, Vec<f32>>,
+) -> HashMap<ArticleId, InterestMatch> {
+    if articles.is_empty() || interests.is_empty() {
+        return HashMap::new();
+    }
+    let fallback = articles.len() < INTEREST_ZSCORE_MIN_ARTICLES;
+    if fallback {
+        tracing::info!(
+            embedded = articles.len(),
+            "fewer than {INTEREST_ZSCORE_MIN_ARTICLES} embedded articles; interest uses the raw top-1 cosine"
+        );
+    }
+
+    let mut matches: HashMap<ArticleId, Vec<TopInterest>> = HashMap::new();
+    for (name, interest) in interests {
+        let similarities = articles
+            .iter()
+            .filter_map(|(article_id, article)| {
+                dot(interest, article).ok().map(|cos| (*article_id, cos))
+            })
+            .collect::<Vec<_>>();
+        if similarities.is_empty() {
+            continue;
+        }
+        let n = similarities.len() as f64;
+        let mean = similarities.iter().map(|(_, cos)| cos).sum::<f64>() / n;
+        let variance = similarities
+            .iter()
+            .map(|(_, cos)| (cos - mean).powi(2))
+            .sum::<f64>()
+            / n;
+        let std = variance.sqrt().max(ZSCORE_STD_FLOOR);
+        for (article_id, cos) in similarities {
+            matches.entry(article_id).or_default().push(TopInterest {
+                name: name.clone(),
+                z: (cos - mean) / std,
+                cos,
+            });
+        }
+    }
+
+    matches
+        .into_iter()
+        .map(|(article_id, mut all)| {
+            all.sort_by(|left, right| {
+                right
+                    .z
+                    .total_cmp(&left.z)
+                    .then_with(|| left.name.cmp(&right.name))
+            });
+            let top1_cos = all
+                .iter()
+                .map(|item| item.cos)
+                .fold(f64::NEG_INFINITY, f64::max);
+            all.truncate(RECORDED_TOP);
+            let score = if fallback {
+                top1_cos
+            } else {
+                let top_mean = all.iter().map(|item| item.z).sum::<f64>() / all.len() as f64;
+                0.7 * all[0].z + 0.3 * top_mean
+            };
+            (
+                article_id,
+                InterestMatch {
+                    score,
+                    top1_cos,
+                    top_interests: all,
+                },
+            )
+        })
+        .collect()
+}
+
+/// Every cheap signal for the eligible set, normalized and blended (§9, §12.2,
+/// §12.4). Pure: the preference state is already loaded.
+pub fn compute(
+    articles: &[Article],
+    article_embeddings: &HashMap<ArticleId, Vec<f32>>,
+    interest_embeddings: &HashMap<String, Vec<f32>>,
+    preference: &PreferenceState,
+    ranking: &RankingConfig,
+) -> HashMap<ArticleId, Signals> {
+    let interests = interest_matches(article_embeddings, interest_embeddings);
+    let mut all = articles
+        .iter()
+        .map(|article| {
+            let mut signals = Signals::baseline(article);
+            signals.knn_gate = preference.knn_gate;
+            signals.feed_gate = preference.feed_gate;
+            if let Some(matched) = interests.get(&article.id) {
+                signals.interest = Some(matched.score);
+                signals.interest_top1_cos = Some(matched.top1_cos);
+                signals.top_interests = matched.top_interests.clone();
+            }
+            if let Some(embedding) = article_embeddings.get(&article.id) {
+                let (knn, neighbours) = preference.knn(embedding, ranking);
+                signals.knn = knn;
+                signals.neighbours = neighbours;
+            }
+            signals.feed = preference.feed(article);
+            if preference.knn_gate > 0.0 {
+                signals.notes.push(format!(
+                    "knn gate {:.2} (n={} rated with embeddings)",
+                    preference.knn_gate,
+                    preference.examples.len()
+                ));
+            }
+            signals
+        })
+        .collect::<Vec<_>>();
+
+    normalize(&mut all.iter_mut().collect::<Vec<_>>());
+    for signals in &mut all {
+        preliminary_blend(signals, &ranking.weights.preliminary);
+    }
+    articles.iter().map(|article| article.id).zip(all).collect()
+}
+
+/// [`compute`] with the preference state loaded from the database.
+pub async fn compute_all(
+    db: &Db,
+    articles: &[Article],
+    article_embeddings: &HashMap<ArticleId, Vec<f32>>,
+    interest_embeddings: &HashMap<String, Vec<f32>>,
+    voyage: &VoyageConfig,
+    ranking: &RankingConfig,
+    now: Timestamp,
+) -> anyhow::Result<(HashMap<ArticleId, Signals>, PreferenceSummary)> {
+    let preference = PreferenceState::load(db, voyage, ranking, now).await?;
+    preference.log(ranking);
+    Ok((
+        compute(
+            articles,
+            article_embeddings,
+            interest_embeddings,
+            &preference,
+            ranking,
+        ),
+        preference.summary(),
+    ))
+}
+
+/// Mid-rank percentiles over the present values of each signal (§12.2).
+///
+/// `p(x) = (count_below + (count_equal + 1) / 2) / n_present`; fewer than two
+/// present values or all-equal values give 0.5. Article id never breaks ties.
+pub fn normalize(signals: &mut [&mut Signals]) {
+    for name in PERCENTILE_SIGNALS {
+        let mut values = signals
+            .iter()
+            .filter_map(|signal| signal.raw(name))
+            .collect::<Vec<_>>();
+        if values.is_empty() {
+            continue;
+        }
+        values.sort_by(f64::total_cmp);
+        let n = values.len() as f64;
+        let constant = values.len() < 2 || values.first() == values.last();
+        for signal in signals.iter_mut() {
+            let Some(value) = signal.raw(name) else {
+                continue;
+            };
+            let percentile = if constant {
+                0.5
+            } else {
+                let below = values.partition_point(|other| *other < value);
+                let equal = values.partition_point(|other| *other <= value) - below;
+                (below as f64 + (equal as f64 + 1.0) / 2.0) / n
+            };
+            signal.norm.insert(name.to_string(), percentile);
+        }
+    }
+}
+
+/// The preliminary blend of §12.4 on a 0–100 scale: present-and-active
+/// signals only, learned weights multiplied by their gate, renormalized to 1.
+pub fn preliminary_blend(signals: &mut Signals, configured: &PreliminaryWeights) -> Option<f64> {
+    let candidates = [
+        ("interest", configured.interest, 1.0),
+        ("knn", configured.knn, signals.knn_gate),
+        ("heuristic", configured.heuristic, 1.0),
+        ("feed", configured.feed, signals.feed_gate),
+        ("social", configured.social, 1.0),
+    ];
+    let active = candidates
+        .into_iter()
+        .filter_map(|(name, weight, gate)| {
+            let norm = *signals.norm.get(name)?;
+            let effective = weight * gate;
+            (effective > 0.0).then_some((name, effective, norm))
+        })
+        .collect::<Vec<_>>();
+    let total = active.iter().map(|(_, weight, _)| weight).sum::<f64>();
+    if total <= 0.0 {
+        signals.weights.clear();
+        signals.preliminary = None;
+        return None;
+    }
+    signals.weights = active
+        .iter()
+        .map(|(name, weight, _)| ((*name).to_string(), weight / total))
+        .collect();
+    let blend = active
+        .iter()
+        .map(|(_, weight, norm)| weight / total * norm)
+        .sum::<f64>()
+        * 100.0;
+    signals.preliminary = Some(blend);
+    Some(blend)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::types::{ExtractMethod, SourceRef};
+
+    fn ranking() -> RankingConfig {
+        RankingConfig::default()
+    }
+
+    fn unit(values: &[f32]) -> Vec<f32> {
+        let norm = values.iter().map(|v| v * v).sum::<f32>().sqrt();
+        values.iter().map(|v| v / norm).collect()
+    }
+
+    fn example(id: ArticleId, label: &str, value: f64, embedding: &[f32]) -> RatedExample {
+        RatedExample {
+            article_id: id,
+            label: label.into(),
+            title: format!("rated {id}"),
+            value,
+            decay: 1.0,
+            embedding: unit(embedding),
+            feeds: vec![id],
+        }
+    }
+
+    fn article(id: ArticleId, feeds: &[FeedId]) -> Article {
+        Article {
+            id,
+            canonical_url: format!("https://example.com/{id}"),
+            title: format!("Article {id}"),
+            best_entry_id: id,
+            content_html: String::new(),
+            word_count: 1000,
+            excerpt_only: false,
+            image_count: 0,
+            sources: feeds
+                .iter()
+                .map(|feed| SourceRef {
+                    entry_id: id,
+                    feed_id: *feed,
+                    feed_title: format!("feed {feed}"),
+                    category: None,
+                    kind: SourceKind::Feed,
+                })
+                .collect(),
+            first_seen: "2026-08-15T00:00:00Z".parse().unwrap(),
+            url: format!("https://example.com/{id}"),
+            author: None,
+            feed_id: feeds.first().copied().unwrap_or(0),
+            feed_title: String::new(),
+            category: None,
+            published_at: None,
+            comments_url: None,
+            image_urls: vec![],
+            social: vec![],
+            extract_method: ExtractMethod::Miniflux,
+        }
+    }
+
+    fn with_heuristic(value: Option<f64>) -> Signals {
+        Signals {
+            heuristic: value,
+            ..Signals::default()
+        }
+    }
+
+    // --- §9.1 interest z-scores ---
+
+    fn interest_fixture(n: usize) -> (HashMap<ArticleId, Vec<f32>>, HashMap<String, Vec<f32>>) {
+        // Article 1 sits on axis x; the rest sit near axis y with a tiny spread.
+        let mut articles = HashMap::new();
+        articles.insert(1, unit(&[1.0, 0.0, 0.0]));
+        for id in 2..=n as ArticleId {
+            articles.insert(id, unit(&[0.0, 1.0, 0.001 * id as f32]));
+        }
+        // "Broad" is about equally close to everything; "Specific" matches only article 1.
+        let mut interests = HashMap::new();
+        interests.insert("Broad".to_string(), unit(&[1.0, 1.0, 0.0]));
+        interests.insert("Specific".to_string(), unit(&[1.0, 0.0, 0.0]));
+        (articles, interests)
+    }
+
+    #[test]
+    fn specific_interest_with_one_strong_match_beats_a_broad_one() {
+        let (articles, interests) = interest_fixture(40);
+        let matched = interest_matches(&articles, &interests);
+        let strong = &matched[&1];
+        assert_eq!(strong.top_interests[0].name, "Specific");
+        assert!(
+            strong.top_interests[0].z > 3.0,
+            "z = {}",
+            strong.top_interests[0].z
+        );
+        let others = (2..=40)
+            .map(|id| matched[&id].score)
+            .fold(f64::NEG_INFINITY, f64::max);
+        assert!(strong.score > others + 2.0, "{} vs {others}", strong.score);
+        // Raw cosine would have called Broad a near-tie everywhere (≈0.707).
+        assert!((matched[&2].top1_cos - 0.707).abs() < 0.01);
+    }
+
+    #[test]
+    fn interest_falls_back_to_raw_cosine_under_thirty_articles() {
+        let (articles, interests) = interest_fixture(10);
+        let matched = interest_matches(&articles, &interests);
+        for (id, m) in &matched {
+            assert!(
+                (m.score - m.top1_cos).abs() < 1e-9,
+                "article {id} should use raw top-1"
+            );
+        }
+        assert!((matched[&2].score - 0.707).abs() < 0.01);
+    }
+
+    // --- §9.2 preference ---
+
+    #[test]
+    fn one_loved_article_gives_a_positive_knn_to_a_near_neighbour() {
+        let mut ranking = ranking();
+        ranking.knn_floor = 0;
+        ranking.knn_full = 1;
+        let state = PreferenceState::build(vec![example(1, "loved", 1.0, &[1.0, 0.0])], &ranking);
+        let (knn, neighbours) = state.knn(&unit(&[0.9, 0.1]), &ranking);
+        assert!(knn.unwrap() > 0.9);
+        assert_eq!(neighbours.len(), 1);
+        assert_eq!(neighbours[0].label, "loved");
+        let (far, _) = state.knn(&unit(&[0.0, 1.0]), &ranking);
+        assert!(far.unwrap().abs() < 1e-6);
+    }
+
+    #[test]
+    fn two_unrelated_loved_clusters_both_score_high() {
+        let mut ranking = ranking();
+        ranking.knn_floor = 0;
+        ranking.knn_full = 1;
+        ranking.neighbour_k = 2;
+        let state = PreferenceState::build(
+            vec![
+                example(1, "loved", 1.0, &[1.0, 0.0, 0.0]),
+                example(2, "loved", 1.0, &[0.98, 0.02, 0.0]),
+                example(3, "loved", 1.0, &[0.0, 1.0, 0.0]),
+                example(4, "loved", 1.0, &[0.0, 0.98, 0.02]),
+            ],
+            &ranking,
+        );
+        let (near_a, _) = state.knn(&unit(&[1.0, 0.0, 0.0]), &ranking);
+        let (near_b, _) = state.knn(&unit(&[0.0, 1.0, 0.0]), &ranking);
+        assert!(near_a.unwrap() > 0.95, "{near_a:?}");
+        assert!(near_b.unwrap() > 0.95, "{near_b:?}");
+        // A centroid would have put both at ~0.7.
+    }
+
+    #[test]
+    fn good_carries_a_third_of_loved() {
+        let loved = example(1, "loved", 1.0, &[1.0, 0.0]);
+        let good = example(2, "good", 0.35, &[1.0, 0.0]);
+        assert!((good.weight() / loved.weight() - 0.35).abs() < 1e-9);
+
+        // A mixed neighbourhood: the far example pulls the mean down by 0.35× as
+        // much weight when it is merely "good" as when it is "loved".
+        let mut ranking = ranking();
+        ranking.knn_floor = 0;
+        ranking.knn_full = 1;
+        let near = example(1, "loved", 1.0, &[1.0, 0.0]);
+        let candidate = unit(&[1.0, 0.0]);
+        let both_loved = PreferenceState::build(
+            vec![near.clone(), example(2, "loved", 1.0, &[0.0, 1.0])],
+            &ranking,
+        );
+        let one_good =
+            PreferenceState::build(vec![near, example(2, "good", 0.35, &[0.0, 1.0])], &ranking);
+        let pull_loved = 1.0 - both_loved.knn(&candidate, &ranking).0.unwrap();
+        let pull_good = 1.0 - one_good.knn(&candidate, &ranking).0.unwrap();
+        assert!(pull_good < pull_loved);
+        // Weighted means: 0.5 vs 1/1.35 → pulls 0.5 vs 0.35/1.35.
+        assert!((pull_good / pull_loved - 0.35 / 1.35 / 0.5).abs() < 1e-9);
+    }
+
+    #[test]
+    fn negatives_subtract_with_the_negative_coefficient() {
+        let mut ranking = ranking();
+        ranking.knn_floor = 0;
+        ranking.knn_full = 1;
+        let state =
+            PreferenceState::build(vec![example(1, "not_for_me", -1.0, &[1.0, 0.0])], &ranking);
+        let (knn, _) = state.knn(&unit(&[1.0, 0.0]), &ranking);
+        assert!((knn.unwrap() + ranking.negative_coefficient).abs() < 1e-9);
+    }
+
+    #[test]
+    fn decay_halves_at_the_half_life() {
+        assert!((decay(60.0, 60.0) - 0.5).abs() < 1e-12);
+        assert!((decay(0.0, 60.0) - 1.0).abs() < 1e-12);
+        assert!((decay(120.0, 60.0) - 0.25).abs() < 1e-12);
+    }
+
+    #[test]
+    fn gate_is_zero_below_floor_one_at_full_and_linear_between() {
+        assert_eq!(gate(0, 8, 25), 0.0);
+        assert_eq!(gate(8, 8, 25), 0.0);
+        assert_eq!(gate(25, 8, 25), 1.0);
+        assert_eq!(gate(100, 8, 25), 1.0);
+        assert!((gate(16, 8, 24) - 0.5).abs() < 1e-9);
+        assert!((gate(9, 8, 25) - 1.0 / 17.0).abs() < 1e-9);
+    }
+
+    #[test]
+    fn knn_is_absent_when_the_gate_is_closed() {
+        let ranking = ranking(); // knn_floor 8
+        let state = PreferenceState::build(vec![example(1, "loved", 1.0, &[1.0, 0.0])], &ranking);
+        assert_eq!(state.knn_gate, 0.0);
+        assert_eq!(state.knn(&unit(&[1.0, 0.0]), &ranking), (None, Vec::new()));
+    }
+
+    // --- §9.3 feed affinity ---
+
+    #[test]
+    fn feed_credit_sums_to_one_across_direct_feeds() {
+        let mut rated = example(1, "loved", 1.0, &[1.0, 0.0]);
+        rated.feeds = vec![10, 20, 30];
+        let state = PreferenceState::build(vec![rated], &ranking());
+        let total: f64 = [10, 20, 30]
+            .iter()
+            .map(|feed| state.feed_credit(*feed).unwrap().0)
+            .sum();
+        assert!((total - 1.0).abs() < 1e-9);
+        assert!((state.feed_credit(10).unwrap().0 - 1.0 / 3.0).abs() < 1e-9);
+    }
+
+    #[test]
+    fn feed_affinity_is_the_mean_over_rated_feeds() {
+        let mut ranking = ranking();
+        ranking.feed_floor = 0;
+        ranking.feed_full = 1;
+        let mut loved = example(1, "loved", 1.0, &[1.0, 0.0]);
+        loved.feeds = vec![10];
+        let mut down = example(2, "not_for_me", -1.0, &[1.0, 0.0]);
+        down.feeds = vec![20];
+        let state = PreferenceState::build(vec![loved, down], &ranking);
+        // feed 10: (1+1)/(1+0+2) = 2/3; feed 20: (0+1)/(0+1+2) = 1/3; unrated 99 ignored.
+        let both = state.feed(&article(7, &[10, 20, 99])).unwrap();
+        assert!((both - 0.5).abs() < 1e-9, "{both}");
+        let best = state.feed(&article(8, &[10])).unwrap();
+        assert!((best - 2.0 / 3.0).abs() < 1e-9);
+        assert_eq!(state.feed(&article(9, &[99])), None);
+    }
+
+    #[test]
+    fn feed_is_absent_when_the_gate_is_closed() {
+        let ranking = ranking(); // feed_floor 15
+        let mut loved = example(1, "loved", 1.0, &[1.0, 0.0]);
+        loved.feeds = vec![10];
+        let state = PreferenceState::build(vec![loved], &ranking);
+        assert_eq!(state.feed_gate, 0.0);
+        assert_eq!(state.feed(&article(7, &[10])), None);
+    }
+
+    #[test]
+    fn direct_feeds_fall_back_to_the_best_entry_feed() {
+        let mut a = article(1, &[]);
+        a.feed_id = 42;
+        assert_eq!(direct_feeds(&a), vec![42]);
+        assert_eq!(direct_feeds(&article(2, &[5, 3, 5])), vec![3, 5]);
+    }
+
+    // --- §12.2 normalization, §12.4 blend ---
+
+    #[test]
+    fn constant_signal_normalizes_to_half_for_everyone() {
+        let mut values = [with_heuristic(Some(7.0)), with_heuristic(Some(7.0))];
+        let mut refs = values.iter_mut().collect::<Vec<_>>();
+        normalize(&mut refs);
+        assert!(values.iter().all(|v| v.norm["heuristic"] == 0.5));
+        let mut single = [with_heuristic(Some(3.0))];
+        let mut refs = single.iter_mut().collect::<Vec<_>>();
+        normalize(&mut refs);
+        assert_eq!(single[0].norm["heuristic"], 0.5);
+    }
+
+    #[test]
+    fn ties_get_equal_percentiles_without_an_id_ramp() {
+        let mut values = (0..400)
+            .map(|_| with_heuristic(Some(0.0)))
+            .collect::<Vec<_>>();
+        let mut refs = values.iter_mut().collect::<Vec<_>>();
+        normalize(&mut refs);
+        assert!(values.iter().all(|v| v.norm["heuristic"] == 0.5));
+
+        let mut mixed = [
+            with_heuristic(Some(1.0)),
+            with_heuristic(Some(2.0)),
+            with_heuristic(Some(2.0)),
+            with_heuristic(Some(3.0)),
+        ];
+        let mut refs = mixed.iter_mut().collect::<Vec<_>>();
+        normalize(&mut refs);
+        assert_eq!(mixed[0].norm["heuristic"], 0.25);
+        assert_eq!(mixed[1].norm["heuristic"], 0.625);
+        assert_eq!(mixed[2].norm["heuristic"], 0.625);
+        assert_eq!(mixed[3].norm["heuristic"], 1.0);
+    }
+
+    #[test]
+    fn absent_values_do_not_shift_present_values() {
+        let mut values = [
+            with_heuristic(Some(1.0)),
+            with_heuristic(Some(2.0)),
+            with_heuristic(None),
+        ];
+        let mut refs = values.iter_mut().collect::<Vec<_>>();
+        normalize(&mut refs);
+        assert!(!values[2].norm.contains_key("heuristic"));
+        // n_present = 2: the absent third value does not widen the scale.
+        assert_eq!(values[0].norm["heuristic"], 0.5);
+        assert_eq!(values[1].norm["heuristic"], 1.0);
+    }
+
+    #[test]
+    fn effective_weights_sum_to_one_and_missing_signals_are_skipped() {
+        let mut signals = Signals {
+            interest: Some(1.0),
+            heuristic: Some(2.0),
+            norm: BTreeMap::from([("interest".into(), 0.8), ("heuristic".into(), 0.4)]),
+            ..Signals::default()
+        };
+        let blend = preliminary_blend(&mut signals, &PreliminaryWeights::default()).unwrap();
+        assert!((signals.weights.values().sum::<f64>() - 1.0).abs() < 1e-9);
+        assert!(!signals.weights.contains_key("knn"));
+        assert!(!signals.weights.contains_key("social"));
+        // 0.35/0.55 × 0.8 + 0.20/0.55 × 0.4 = 0.6545…
+        assert!((blend - 65.4545).abs() < 0.01, "{blend}");
+
+        let mut only_heuristic = Signals {
+            heuristic: Some(2.0),
+            norm: BTreeMap::from([("heuristic".into(), 0.4)]),
+            ..Signals::default()
+        };
+        let blend = preliminary_blend(&mut only_heuristic, &PreliminaryWeights::default());
+        assert!((blend.unwrap() - 40.0).abs() < 1e-9);
+        assert_eq!(only_heuristic.weights["heuristic"], 1.0);
+    }
+
+    #[test]
+    fn learned_weights_are_multiplied_by_their_gate() {
+        let mut signals = Signals {
+            knn: Some(0.5),
+            heuristic: Some(2.0),
+            knn_gate: 0.5,
+            norm: BTreeMap::from([("knn".into(), 1.0), ("heuristic".into(), 0.0)]),
+            ..Signals::default()
+        };
+        preliminary_blend(&mut signals, &PreliminaryWeights::default());
+        // knn 0.25 × 0.5 = 0.125 against heuristic 0.20.
+        assert!((signals.weights["knn"] - 0.125 / 0.325).abs() < 1e-9);
+    }
+
+    #[test]
+    fn compute_scores_every_article_and_leaves_ungated_signals_absent() {
+        let articles = vec![article(1, &[10]), article(2, &[20]), article(3, &[30])];
+        let mut embeddings = HashMap::new();
+        embeddings.insert(1, unit(&[1.0, 0.0]));
+        embeddings.insert(2, unit(&[0.0, 1.0]));
+        let mut interests = HashMap::new();
+        interests.insert("Axis".to_string(), unit(&[1.0, 0.0]));
+        let state = PreferenceState::build(vec![example(9, "loved", 1.0, &[1.0, 0.0])], &ranking());
+        let signals = compute(&articles, &embeddings, &interests, &state, &ranking());
+        assert_eq!(signals.len(), 3);
+        assert!(signals[&1].interest.is_some());
+        assert!(signals[&3].interest.is_none(), "no embedding → absent");
+        assert!(
+            signals
+                .values()
+                .all(|s| s.knn.is_none() && s.feed.is_none())
+        );
+        assert!(
+            signals
+                .values()
+                .all(|s| s.heuristic.is_some() && s.preliminary.is_some())
+        );
+    }
+}

src/curate/telemetry.rs

@@ -0,0 +1,962 @@
+//! Per-run candidate telemetry: the `candidate_runs` writer, `signals_json`,
+//! the `explain` command and feature retention (plan §7.4–7.5, §15.2, §16).
+//!
+//! One row per considered article per run says where it stopped and why. Rows
+//! are upserted on every stage transition with every column set (never
+//! `COALESCE`), so the last write for a run is the whole truth.
+
+use std::collections::BTreeMap;
+use std::fmt::Write as _;
+
+use jiff::Timestamp;
+use jiff::civil::Date;
+use serde::{Deserialize, Serialize};
+use sqlx::Row as _;
+
+use crate::curate::signals::{Neighbour, Signals, TopInterest};
+use crate::db::{Db, fmt_ts};
+use crate::types::ArticleId;
+
+/// The stage vocabulary of §7.4, in pipeline order.
+pub const STAGES: [&str; 7] = [
+    "excluded",
+    "eligible",
+    "triaged",
+    "admitted",
+    "assessed",
+    "shortlisted",
+    "selected",
+];
+
+/// Signal names rendered by `explain`, including the LLM ones steps 4–5 add.
+const RENDERED_SIGNALS: [&str; 8] = [
+    "interest",
+    "knn",
+    "feed",
+    "social",
+    "heuristic",
+    "triage",
+    "quality",
+    "fit",
+];
+
+/// One `candidate_runs` row (§7.4).
+#[derive(Debug, Clone)]
+pub struct CandidateRun<'a> {
+    pub run_id: i64,
+    pub article_id: ArticleId,
+    pub stage: &'a str,
+    pub excluded_reason: Option<&'a str>,
+    /// JSON array of retriever names, first = the one that admitted it.
+    pub admitted_by: Option<&'a str>,
+    pub signals_json: &'a str,
+    pub utility: Option<f64>,
+    pub rank_utility: Option<i64>,
+    pub cluster_id: Option<i64>,
+    pub cluster_rank: Option<i64>,
+    pub editor_why: Option<&'a str>,
+}
+
+/// Upsert one row, setting every column (§7.4).
+pub async fn write(db: &Db, row: &CandidateRun<'_>) -> Result<(), sqlx::Error> {
+    sqlx::query(
+        "INSERT INTO candidate_runs
+             (run_id, article_id, stage, excluded_reason, admitted_by, signals_json,
+              utility, rank_utility, cluster_id, cluster_rank, editor_why)
+         VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+         ON CONFLICT(run_id, article_id) DO UPDATE SET
+             stage = excluded.stage,
+             excluded_reason = excluded.excluded_reason,
+             admitted_by = excluded.admitted_by,
+             signals_json = excluded.signals_json,
+             utility = excluded.utility,
+             rank_utility = excluded.rank_utility,
+             cluster_id = excluded.cluster_id,
+             cluster_rank = excluded.cluster_rank,
+             editor_why = excluded.editor_why",
+    )
+    .bind(row.run_id)
+    .bind(row.article_id)
+    .bind(row.stage)
+    .bind(row.excluded_reason)
+    .bind(row.admitted_by)
+    .bind(row.signals_json)
+    .bind(row.utility)
+    .bind(row.rank_utility)
+    .bind(row.cluster_id)
+    .bind(row.cluster_rank)
+    .bind(row.editor_why)
+    .execute(db.pool())
+    .await?;
+    Ok(())
+}
+
+/// The thin row of a hygiene exclusion: keys, `stage = 'excluded'`, the reason
+/// and `signals_json = '{}'` (§8.1).
+pub async fn thin_excluded(
+    db: &Db,
+    run_id: i64,
+    article_id: ArticleId,
+    reason: &str,
+) -> Result<(), sqlx::Error> {
+    write(
+        db,
+        &CandidateRun {
+            run_id,
+            article_id,
+            stage: "excluded",
+            excluded_reason: Some(reason),
+            admitted_by: None,
+            signals_json: "{}",
+            utility: None,
+            rank_utility: None,
+            cluster_id: None,
+            cluster_rank: None,
+            editor_why: None,
+        },
+    )
+    .await
+}
+
+/// `signals_json` (§7.5). Missing signals are absent from `raw`/`norm` and
+/// `false` in `present`; `weights` are the effective weights.
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+pub struct SignalsJson {
+    pub v: i64,
+    #[serde(default)]
+    pub raw: BTreeMap<String, f64>,
+    #[serde(default)]
+    pub norm: BTreeMap<String, f64>,
+    #[serde(default)]
+    pub present: BTreeMap<String, bool>,
+    #[serde(default)]
+    pub weights: BTreeMap<String, f64>,
+    #[serde(default)]
+    pub top_interests: Vec<TopInterest>,
+    #[serde(default)]
+    pub neighbours: Vec<Neighbour>,
+    #[serde(default)]
+    pub exploration: bool,
+    #[serde(default)]
+    pub auto_include: bool,
+    #[serde(default)]
+    pub notes: Vec<String>,
+}
+
+impl SignalsJson {
+    /// The blend implied by the stored effective weights, 0–100.
+    pub fn blend(&self) -> Option<f64> {
+        let mut score = 0.0;
+        let mut any = false;
+        for (name, weight) in &self.weights {
+            if let Some(value) = self.norm.get(name) {
+                score += weight * value;
+                any = true;
+            }
+        }
+        any.then_some(score * 100.0)
+    }
+}
+
+/// Serialize the signals of §7.5 for one article.
+pub fn serialize_signals(signals: &Signals, auto_include: bool) -> String {
+    let mut raw = BTreeMap::new();
+    for name in [
+        "interest",
+        "interest_top1_cos",
+        "knn",
+        "feed",
+        "social",
+        "heuristic",
+    ] {
+        if let Some(value) = signals.raw(name) {
+            raw.insert(name.to_string(), value);
+        }
+    }
+    let present = RENDERED_SIGNALS
+        .into_iter()
+        .map(|name| (name.to_string(), signals.present(name)))
+        .collect();
+    serde_json::to_string(&SignalsJson {
+        v: 1,
+        raw,
+        norm: signals.norm.clone(),
+        present,
+        weights: signals.weights.clone(),
+        top_interests: signals.top_interests.clone(),
+        neighbours: signals.neighbours.clone(),
+        exploration: false,
+        auto_include,
+        notes: signals.notes.clone(),
+    })
+    .unwrap_or_else(|_| "{}".into())
+}
+
+// ---------------------------------------------------------------------------
+// `explain` (§15.2)
+// ---------------------------------------------------------------------------
+
+/// A `candidate_runs` row joined to its article title.
+#[derive(Debug, Clone)]
+pub struct ExplainRow {
+    pub run_id: i64,
+    pub article_id: ArticleId,
+    pub title: String,
+    pub stage: String,
+    pub excluded_reason: Option<String>,
+    pub admitted_by: Option<String>,
+    pub signals_json: String,
+    pub utility: Option<f64>,
+    pub rank_utility: Option<i64>,
+    pub cluster_id: Option<i64>,
+    pub cluster_rank: Option<i64>,
+    pub editor_why: Option<String>,
+}
+
+impl ExplainRow {
+    fn from_row(row: &sqlx::sqlite::SqliteRow) -> Self {
+        Self {
+            run_id: row.get("run_id"),
+            article_id: row.get("article_id"),
+            title: row.get("title"),
+            stage: row.get("stage"),
+            excluded_reason: row.get("excluded_reason"),
+            admitted_by: row.get("admitted_by"),
+            signals_json: row.get("signals_json"),
+            utility: row.get("utility"),
+            rank_utility: row.get("rank_utility"),
+            cluster_id: row.get("cluster_id"),
+            cluster_rank: row.get("cluster_rank"),
+            editor_why: row.get("editor_why"),
+        }
+    }
+
+    pub fn signals(&self) -> Option<SignalsJson> {
+        serde_json::from_str(&self.signals_json).ok()
+    }
+
+    /// Utility when step 5 has written it, else the preliminary blend.
+    pub fn score(&self) -> Option<f64> {
+        self.utility
+            .or_else(|| self.signals().and_then(|signals| signals.blend()))
+    }
+}
+
+/// The run `explain` reads: `--run-id` when given (and of that date), else the
+/// latest non-dry run of the date.
+pub async fn resolve_run(
+    db: &Db,
+    date: Date,
+    requested: Option<i64>,
+) -> Result<Option<i64>, sqlx::Error> {
+    let row = match requested {
+        Some(run_id) => {
+            sqlx::query("SELECT id FROM runs WHERE id = ? AND date = ?")
+                .bind(run_id)
+                .bind(date.to_string())
+                .fetch_optional(db.pool())
+                .await?
+        }
+        None => {
+            sqlx::query(
+                "SELECT id FROM runs WHERE date = ? AND status != 'dry_run'
+                 ORDER BY id DESC LIMIT 1",
+            )
+            .bind(date.to_string())
+            .fetch_optional(db.pool())
+            .await?
+        }
+    };
+    Ok(row.map(|row| row.get("id")))
+}
+
+pub async fn explain_row(
+    db: &Db,
+    run_id: i64,
+    article_id: ArticleId,
+) -> Result<Option<ExplainRow>, sqlx::Error> {
+    let row = sqlx::query(
+        "SELECT cr.run_id, cr.article_id, COALESCE(a.title, '') AS title,
+                cr.stage, cr.excluded_reason, cr.admitted_by, cr.signals_json,
+                cr.utility, cr.rank_utility, cr.cluster_id, cr.cluster_rank, cr.editor_why
+         FROM candidate_runs cr JOIN articles a ON a.id = cr.article_id
+         WHERE cr.run_id = ? AND cr.article_id = ?",
+    )
+    .bind(run_id)
+    .bind(article_id)
+    .fetch_optional(db.pool())
+    .await?;
+    Ok(row.as_ref().map(ExplainRow::from_row))
+}
+
+/// The top `limit` rows by utility-or-blend that were not selected (§15.2).
+pub async fn near_misses(
+    db: &Db,
+    run_id: i64,
+    limit: usize,
+) -> Result<Vec<ExplainRow>, sqlx::Error> {
+    let rows = sqlx::query(
+        "SELECT cr.run_id, cr.article_id, COALESCE(a.title, '') AS title,
+                cr.stage, cr.excluded_reason, cr.admitted_by, cr.signals_json,
+                cr.utility, cr.rank_utility, cr.cluster_id, cr.cluster_rank, cr.editor_why
+         FROM candidate_runs cr JOIN articles a ON a.id = cr.article_id
+         WHERE cr.run_id = ? AND cr.stage != 'selected' AND cr.stage != 'excluded'",
+    )
+    .bind(run_id)
+    .fetch_all(db.pool())
+    .await?;
+    let mut output = rows.iter().map(ExplainRow::from_row).collect::<Vec<_>>();
+    output.sort_by(|left, right| {
+        right
+            .score()
+            .unwrap_or(f64::NEG_INFINITY)
+            .total_cmp(&left.score().unwrap_or(f64::NEG_INFINITY))
+            .then_with(|| left.article_id.cmp(&right.article_id))
+    });
+    output.truncate(limit);
+    Ok(output)
+}
+
+fn fmt_opt(value: Option<f64>) -> String {
+    value
+        .map(|v| format!("{v:.3}"))
+        .unwrap_or_else(|| "—".into())
+}
+
+/// Render one persisted row the way §15.2 lists it.
+pub async fn render_explain(db: &Db, row: &ExplainRow) -> Result<String, sqlx::Error> {
+    let mut out = String::new();
+    let _ = writeln!(out, "article {}: {}", row.article_id, row.title);
+    let _ = write!(out, "run {} · stage: {}", row.run_id, row.stage);
+    if let Some(reason) = &row.excluded_reason {
+        let _ = write!(out, " · reason: {reason}");
+    }
+    let _ = writeln!(out);
+    if let Some(signals) = row.signals() {
+        let _ = writeln!(out, "signals (raw · norm · weight):");
+        for name in RENDERED_SIGNALS {
+            let present = signals.present.get(name).copied().unwrap_or(false);
+            if present {
+                let _ = writeln!(
+                    out,
+                    "  {name:<10} {:>8} · {:>6} · {:>6}",
+                    fmt_opt(signals.raw.get(name).copied()),
+                    fmt_opt(signals.norm.get(name).copied()),
+                    fmt_opt(signals.weights.get(name).copied()),
+                );
+            } else {
+                let _ = writeln!(out, "  {name:<10}   absent");
+            }
+        }
+        if let Some(blend) = signals.blend() {
+            let _ = writeln!(out, "preliminary blend: {blend:.1}");
+        }
+        if let Some(cos) = signals.raw.get("interest_top1_cos") {
+            let _ = writeln!(out, "interest top-1 cosine: {cos:.3}");
+        }
+        if !signals.top_interests.is_empty() {
+            let _ = writeln!(out, "top interests:");
+            for interest in &signals.top_interests {
+                let _ = writeln!(
+                    out,
+                    "  {} · z {:.2} · cos {:.3}",
+                    interest.name, interest.z, interest.cos
+                );
+            }
+        }
+        if !signals.neighbours.is_empty() {
+            let _ = writeln!(out, "nearest rated neighbours:");
+            for neighbour in &signals.neighbours {
+                let _ = writeln!(
+                    out,
+                    "  {} · cos {:.3} · article {} · {}",
+                    neighbour.label, neighbour.cos, neighbour.article_id, neighbour.title
+                );
+            }
+        }
+        if signals.exploration || signals.auto_include {
+            let _ = writeln!(
+                out,
+                "flags: exploration={} auto_include={}",
+                signals.exploration, signals.auto_include
+            );
+        }
+        for note in &signals.notes {
+            let _ = writeln!(out, "note: {note}");
+        }
+    }
+
+    let assessments = sqlx::query(
+        "SELECT stage, model, score, fit, kind, facets_json, rationale, category,
+                paywalled_guess, assessed_at
+         FROM article_assessments WHERE article_id = ? ORDER BY stage",
+    )
+    .bind(row.article_id)
+    .fetch_all(db.pool())
+    .await?;
+    if !assessments.is_empty() {
+        let _ = writeln!(out, "assessments:");
+        for assessment in assessments {
+            let _ = writeln!(
+                out,
+                "  {} · {} · score {} · fit {} · kind {} · category {} · paywalled={} · {}",
+                assessment.get::<String, _>("stage"),
+                assessment.get::<String, _>("model"),
+                fmt_opt(assessment.get::<Option<f64>, _>("score")),
+                fmt_opt(assessment.get::<Option<f64>, _>("fit")),
+                assessment
+                    .get::<Option<String>, _>("kind")
+                    .unwrap_or_else(|| "—".into()),
+                assessment
+                    .get::<Option<String>, _>("category")
+                    .unwrap_or_else(|| "—".into()),
+                assessment.get::<i64, _>("paywalled_guess") != 0,
+                assessment
+                    .get::<Option<String>, _>("rationale")
+                    .unwrap_or_default(),
+            );
+            if let Some(facets) = assessment.get::<Option<String>, _>("facets_json") {
+                let _ = writeln!(out, "    facets: {facets}");
+            }
+        }
+    }
+    if row.utility.is_some() || row.rank_utility.is_some() {
+        let _ = writeln!(
+            out,
+            "utility: {} · rank {}",
+            fmt_opt(row.utility),
+            row.rank_utility
+                .map(|r| r.to_string())
+                .unwrap_or_else(|| "—".into())
+        );
+    }
+    if let Some(cluster) = row.cluster_id {
+        let _ = writeln!(
+            out,
+            "cluster: {cluster} · rank {}",
+            row.cluster_rank
+                .map(|r| r.to_string())
+                .unwrap_or_else(|| "—".into())
+        );
+    }
+    if let Some(admitted_by) = &row.admitted_by {
+        let _ = writeln!(out, "admitted by: {admitted_by}");
+    }
+    if let Some(why) = &row.editor_why {
+        let _ = writeln!(out, "editor: {why}");
+    }
+    Ok(out)
+}
+
+/// What `explain` was asked about.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum ExplainTarget {
+    Article(ArticleId),
+    Url(String),
+}
+
+/// `explain --date D (--article ID | --url URL) [--run-id N]` as text (§15.2).
+pub async fn explain(
+    db: &Db,
+    date: Date,
+    run_id: Option<i64>,
+    target: &ExplainTarget,
+) -> anyhow::Result<String> {
+    let article_id = match target {
+        ExplainTarget::Article(id) => *id,
+        ExplainTarget::Url(url) => {
+            let canonical = crate::dedupe::canonical_url(url)
+                .ok_or_else(|| anyhow::anyhow!("invalid article URL {url:?}"))?;
+            match db.article_id_for_url(&canonical).await? {
+                Some(id) => id,
+                None => {
+                    return Ok(format!(
+                        "{canonical} was never ingested: it is not in `articles`, so this is a feed problem, not a ranking problem."
+                    ));
+                }
+            }
+        }
+    };
+    if db.get_article(article_id).await?.is_none() {
+        return Ok(format!(
+            "article {article_id} was never ingested: it is not in `articles`."
+        ));
+    }
+    let Some(run_id) = resolve_run(db, date, run_id).await? else {
+        return Ok(match run_id {
+            Some(id) => format!("run {id} is not a run for {date}"),
+            None => format!("no non-dry run recorded for {date}"),
+        });
+    };
+    match explain_row(db, run_id, article_id).await? {
+        Some(row) => Ok(render_explain(db, &row).await?),
+        None => Ok(format!(
+            "article {article_id} was not considered by run {run_id} for {date} (outside its ingest window, or telemetry pruned)."
+        )),
+    }
+}
+
+/// `explain --date D --near-misses [N]` as text (§15.2).
+pub async fn explain_near_misses(
+    db: &Db,
+    date: Date,
+    run_id: Option<i64>,
+    limit: usize,
+) -> anyhow::Result<String> {
+    let Some(run_id) = resolve_run(db, date, run_id).await? else {
+        return Ok(format!("no non-dry run recorded for {date}"));
+    };
+    let rows = near_misses(db, run_id, limit).await?;
+    let mut out = String::new();
+    let _ = writeln!(
+        out,
+        "run {run_id} · {date} · top {} not selected, by {}:",
+        rows.len(),
+        if rows.iter().any(|row| row.utility.is_some()) {
+            "utility"
+        } else {
+            "preliminary blend"
+        }
+    );
+    for (index, row) in rows.iter().enumerate() {
+        let reason = row
+            .excluded_reason
+            .as_deref()
+            .map(|reason| format!(", {reason}"))
+            .unwrap_or_default();
+        let _ = writeln!(
+            out,
+            "{:>3}. {:>6} · {} · {}{} · article {}",
+            index + 1,
+            row.score()
+                .map(|score| format!("{score:.1}"))
+                .unwrap_or_else(|| "—".into()),
+            row.title,
+            row.stage,
+            reason,
+            row.article_id
+        );
+    }
+    Ok(out)
+}
+
+// ---------------------------------------------------------------------------
+// `features prune` (§7.1, §7.4)
+// ---------------------------------------------------------------------------
+
+/// Delete `article_embeddings` for articles neither rated nor published that
+/// are older than `embedding_retention_days`, and `candidate_runs` rows whose
+/// run started more than `telemetry_retention_days` ago. Returns the counts.
+pub async fn prune(
+    db: &Db,
+    embedding_retention_days: i64,
+    telemetry_retention_days: i64,
+    now: Timestamp,
+) -> Result<(u64, u64), sqlx::Error> {
+    let cutoff = |days: i64| {
+        now.checked_sub(jiff::Span::new().hours(days.max(0).saturating_mul(24)))
+            .unwrap_or(Timestamp::UNIX_EPOCH)
+    };
+    let embeddings = sqlx::query(
+        "DELETE FROM article_embeddings
+         WHERE article_id IN (
+             SELECT ae.article_id
+             FROM article_embeddings ae JOIN articles a ON a.id = ae.article_id
+             WHERE a.first_seen < ?
+               AND NOT EXISTS (SELECT 1 FROM rating_events re WHERE re.article_id = ae.article_id)
+               AND NOT EXISTS (SELECT 1 FROM issue_articles ia WHERE ia.article_id = ae.article_id)
+         )",
+    )
+    .bind(fmt_ts(cutoff(embedding_retention_days)))
+    .execute(db.pool())
+    .await?
+    .rows_affected();
+
+    let telemetry = sqlx::query(
+        "DELETE FROM candidate_runs
+         WHERE run_id IN (SELECT id FROM runs WHERE started_at < ?)",
+    )
+    .bind(fmt_ts(cutoff(telemetry_retention_days)))
+    .execute(db.pool())
+    .await?
+    .rows_affected();
+    Ok((embeddings, telemetry))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::curate::embedding::encode_blob;
+
+    async fn db_with_articles(ids: &[ArticleId]) -> (tempfile::TempDir, Db) {
+        let dir = tempfile::tempdir().unwrap();
+        let db = Db::open_and_migrate(&dir.path().join("telemetry.db"))
+            .await
+            .unwrap();
+        for id in ids {
+            sqlx::query(
+                "INSERT INTO articles (id, canonical_url, title, first_seen)
+                 VALUES (?, ?, ?, '2026-08-15T00:00:00Z')",
+            )
+            .bind(id)
+            .bind(format!("https://example.com/{id}"))
+            .bind(format!("Article {id}"))
+            .execute(db.pool())
+            .await
+            .unwrap();
+        }
+        (dir, db)
+    }
+
+    fn date() -> Date {
+        "2026-09-02".parse().unwrap()
+    }
+
+    fn signals(heuristic: f64, norm: f64) -> Signals {
+        Signals {
+            interest: Some(1.2),
+            interest_top1_cos: Some(0.61),
+            heuristic: Some(heuristic),
+            norm: BTreeMap::from([("heuristic".into(), norm), ("interest".into(), 0.9)]),
+            weights: BTreeMap::from([
+                ("heuristic".into(), 0.2 / 0.55),
+                ("interest".into(), 0.35 / 0.55),
+            ]),
+            top_interests: vec![TopInterest {
+                name: "Gaussian Splatting".into(),
+                z: 3.4,
+                cos: 0.61,
+            }],
+            neighbours: vec![Neighbour {
+                article_id: 812,
+                label: "loved".into(),
+                cos: 0.71,
+                title: "A rated piece".into(),
+            }],
+            notes: vec!["knn gate 0.60 (n=14 rated with embeddings)".into()],
+            ..Signals::default()
+        }
+    }
+
+    #[test]
+    fn signals_json_follows_the_plan_shape() {
+        let json = serialize_signals(&signals(41.0, 0.55), false);
+        let parsed: serde_json::Value = serde_json::from_str(&json).unwrap();
+        assert_eq!(parsed["v"], 1);
+        assert_eq!(parsed["raw"]["heuristic"], 41.0);
+        assert_eq!(parsed["raw"]["interest_top1_cos"], 0.61);
+        assert_eq!(parsed["present"]["heuristic"], true);
+        assert_eq!(parsed["present"]["knn"], false);
+        assert_eq!(parsed["present"]["quality"], false);
+        assert!(parsed["raw"].get("knn").is_none());
+        assert!(parsed["norm"].get("knn").is_none());
+        assert_eq!(parsed["exploration"], false);
+        assert_eq!(parsed["auto_include"], false);
+        assert_eq!(parsed["top_interests"][0]["name"], "Gaussian Splatting");
+        assert_eq!(parsed["neighbours"][0]["article_id"], 812);
+        assert_eq!(
+            parsed["notes"][0],
+            "knn gate 0.60 (n=14 rated with embeddings)"
+        );
+        let typed: SignalsJson = serde_json::from_str(&json).unwrap();
+        let weights: f64 = typed.weights.values().sum();
+        assert!((weights - 1.0).abs() < 1e-9);
+        assert!(typed.blend().is_some());
+    }
+
+    #[tokio::test]
+    async fn rows_are_upserted_with_every_column_replaced() {
+        let (_dir, db) = db_with_articles(&[1]).await;
+        let run_id = db.start_run(date(), Timestamp::now()).await.unwrap();
+        write(
+            &db,
+            &CandidateRun {
+                run_id,
+                article_id: 1,
+                stage: "eligible",
+                excluded_reason: Some("not_admitted"),
+                admitted_by: None,
+                signals_json: "{}",
+                utility: Some(1.0),
+                rank_utility: None,
+                cluster_id: None,
+                cluster_rank: None,
+                editor_why: None,
+            },
+        )
+        .await
+        .unwrap();
+        write(
+            &db,
+            &CandidateRun {
+                run_id,
+                article_id: 1,
+                stage: "selected",
+                excluded_reason: None,
+                admitted_by: Some("[\"prefilter\"]"),
+                signals_json: "{\"v\":1}",
+                utility: None,
+                rank_utility: None,
+                cluster_id: None,
+                cluster_rank: None,
+                editor_why: Some("because"),
+            },
+        )
+        .await
+        .unwrap();
+        let row = explain_row(&db, run_id, 1).await.unwrap().unwrap();
+        assert_eq!(row.stage, "selected");
+        assert_eq!(row.excluded_reason, None, "no COALESCE");
+        assert_eq!(row.utility, None);
+        assert_eq!(row.admitted_by.as_deref(), Some("[\"prefilter\"]"));
+        assert_eq!(row.editor_why.as_deref(), Some("because"));
+        let count: i64 = sqlx::query_scalar("SELECT COUNT(*) FROM candidate_runs")
+            .fetch_one(db.pool())
+            .await
+            .unwrap();
+        assert_eq!(count, 1);
+    }
+
+    #[tokio::test]
+    async fn explain_renders_persisted_rows_and_reports_never_ingested() {
+        let (_dir, db) = db_with_articles(&[1, 2]).await;
+        let run_id = db.start_run(date(), Timestamp::now()).await.unwrap();
+        thin_excluded(&db, run_id, 2, "blocked").await.unwrap();
+        let json = serialize_signals(&signals(41.0, 0.55), true);
+        write(
+            &db,
+            &CandidateRun {
+                run_id,
+                article_id: 1,
+                stage: "shortlisted",
+                excluded_reason: Some("not_selected"),
+                admitted_by: Some("[\"prefilter\"]"),
+                signals_json: &json,
+                utility: None,
+                rank_utility: None,
+                cluster_id: None,
+                cluster_rank: None,
+                editor_why: None,
+            },
+        )
+        .await
+        .unwrap();
+        // A later dry run must not shadow the real one.
+        let dry = db.start_run(date(), Timestamp::now()).await.unwrap();
+        sqlx::query("UPDATE runs SET status = 'dry_run' WHERE id = ?")
+            .bind(dry)
+            .execute(db.pool())
+            .await
+            .unwrap();
+        assert_eq!(resolve_run(&db, date(), None).await.unwrap(), Some(run_id));
+        assert_eq!(
+            resolve_run(&db, date(), Some(dry)).await.unwrap(),
+            Some(dry)
+        );
+        assert_eq!(
+            resolve_run(&db, "2026-01-01".parse().unwrap(), Some(dry))
+                .await
+                .unwrap(),
+            None
+        );
+
+        let text = explain(&db, date(), None, &ExplainTarget::Article(1))
+            .await
+            .unwrap();
+        assert!(text.contains("article 1: Article 1"), "{text}");
+        assert!(
+            text.contains("stage: shortlisted · reason: not_selected"),
+            "{text}"
+        );
+        let squashed = text.split_whitespace().collect::<Vec<_>>().join(" ");
+        assert!(
+            squashed.contains("heuristic 41.000 · 0.550 · 0.364"),
+            "{text}"
+        );
+        assert!(squashed.contains("knn absent"), "{text}");
+        assert!(squashed.contains("quality absent"), "{text}");
+        assert!(
+            text.contains("Gaussian Splatting · z 3.40 · cos 0.610"),
+            "{text}"
+        );
+        assert!(
+            text.contains("loved · cos 0.710 · article 812 · A rated piece"),
+            "{text}"
+        );
+        assert!(text.contains("admitted by: [\"prefilter\"]"), "{text}");
+        assert!(text.contains("auto_include=true"), "{text}");
+        assert!(text.contains("preliminary blend:"), "{text}");
+        assert!(text.contains("note: knn gate"), "{text}");
+
+        let by_url = explain(
+            &db,
+            date(),
+            None,
+            &ExplainTarget::Url("https://example.com/1?utm_source=x".into()),
+        )
+        .await
+        .unwrap();
+        assert_eq!(by_url, text, "--url canonicalizes and finds the same row");
+
+        let thin = explain(&db, date(), None, &ExplainTarget::Article(2))
+            .await
+            .unwrap();
+        assert!(thin.contains("stage: excluded · reason: blocked"), "{thin}");
+
+        let missing = explain(
+            &db,
+            date(),
+            None,
+            &ExplainTarget::Url("https://nowhere.example/post".into()),
+        )
+        .await
+        .unwrap();
+        assert!(missing.contains("never ingested"), "{missing}");
+        let missing_id = explain(&db, date(), None, &ExplainTarget::Article(99))
+            .await
+            .unwrap();
+        assert!(missing_id.contains("never ingested"), "{missing_id}");
+
+        let no_run = explain(
+            &db,
+            "2026-01-01".parse().unwrap(),
+            None,
+            &ExplainTarget::Article(1),
+        )
+        .await
+        .unwrap();
+        assert!(no_run.contains("no non-dry run"), "{no_run}");
+        let not_considered = explain(&db, date(), Some(dry), &ExplainTarget::Article(1))
+            .await
+            .unwrap();
+        assert!(
+            not_considered.contains("was not considered by run"),
+            "{not_considered}"
+        );
+    }
+
+    #[tokio::test]
+    async fn near_misses_rank_by_blend_and_skip_selected_and_excluded() {
+        let (_dir, db) = db_with_articles(&[1, 2, 3, 4, 5]).await;
+        let run_id = db.start_run(date(), Timestamp::now()).await.unwrap();
+        let rows = [
+            (1, "selected", None, 0.9),
+            (2, "shortlisted", Some("not_selected"), 0.7),
+            (3, "eligible", Some("not_admitted"), 0.95),
+            (4, "shortlisted", Some("not_selected"), 0.1),
+        ];
+        for (id, stage, reason, norm) in rows {
+            let json = serialize_signals(&signals(10.0, norm), false);
+            write(
+                &db,
+                &CandidateRun {
+                    run_id,
+                    article_id: id,
+                    stage,
+                    excluded_reason: reason,
+                    admitted_by: None,
+                    signals_json: &json,
+                    utility: None,
+                    rank_utility: None,
+                    cluster_id: None,
+                    cluster_rank: None,
+                    editor_why: None,
+                },
+            )
+            .await
+            .unwrap();
+        }
+        thin_excluded(&db, run_id, 5, "published_before")
+            .await
+            .unwrap();
+        let misses = near_misses(&db, run_id, 10).await.unwrap();
+        assert_eq!(
+            misses.iter().map(|row| row.article_id).collect::<Vec<_>>(),
+            vec![3, 2, 4]
+        );
+        let text = explain_near_misses(&db, date(), None, 2).await.unwrap();
+        assert!(
+            text.contains("top 2 not selected, by preliminary blend"),
+            "{text}"
+        );
+        assert!(
+            text.contains("Article 3 · eligible, not_admitted"),
+            "{text}"
+        );
+        assert!(!text.contains("Article 4"), "{text}");
+    }
+
+    #[tokio::test]
+    async fn prune_respects_rated_and_published() {
+        let (_dir, db) = db_with_articles(&[1, 2, 3, 4]).await;
+        let now = Timestamp::now();
+        let old = fmt_ts(now - jiff::Span::new().hours(200 * 24));
+        sqlx::query("UPDATE articles SET first_seen = ? WHERE id IN (1, 2, 3)")
+            .bind(&old)
+            .execute(db.pool())
+            .await
+            .unwrap();
+        let blob = encode_blob(&[0.5, 0.5]).unwrap();
+        for id in 1..=4 {
+            sqlx::query(
+                "INSERT INTO article_embeddings
+                     (article_id, model, dimension, input_hash, embedding, created_at)
+                 VALUES (?, 'voyage-4-lite', 2, 'h', ?, ?)",
+            )
+            .bind(id)
+            .bind(&blob)
+            .bind(&old)
+            .execute(db.pool())
+            .await
+            .unwrap();
+        }
+        sqlx::query(
+            "INSERT INTO rating_events (article_id, kind, source, label, value, event_at)
+             VALUES (1, 'explicit', 'cli', 'loved', 1.0, ?)",
+        )
+        .bind(&old)
+        .execute(db.pool())
+        .await
+        .unwrap();
+        sqlx::query(
+            "INSERT INTO issues (date, issue_number, generated_at) VALUES ('2026-02-01', 1, ?);
+             INSERT INTO issue_articles (issue_date, article_id, section) VALUES ('2026-02-01', 2, 'Top Stories');",
+        )
+        .bind(&old)
+        .execute(db.pool())
+        .await
+        .unwrap();
+
+        let old_run = db
+            .start_run("2026-02-01".parse().unwrap(), now)
+            .await
+            .unwrap();
+        sqlx::query("UPDATE runs SET started_at = ? WHERE id = ?")
+            .bind(&old)
+            .bind(old_run)
+            .execute(db.pool())
+            .await
+            .unwrap();
+        let new_run = db.start_run(date(), now).await.unwrap();
+        thin_excluded(&db, old_run, 1, "blocked").await.unwrap();
+        thin_excluded(&db, new_run, 1, "blocked").await.unwrap();
+
+        let (embeddings, telemetry) = prune(&db, 120, 180, now).await.unwrap();
+        assert_eq!(
+            embeddings, 1,
+            "only the old, unrated, unpublished article 3"
+        );
+        assert_eq!(telemetry, 1, "only the old run's rows");
+        let remaining: Vec<i64> =
+            sqlx::query_scalar("SELECT article_id FROM article_embeddings ORDER BY article_id")
+                .fetch_all(db.pool())
+                .await
+                .unwrap();
+        assert_eq!(remaining, vec![1, 2, 4]);
+        let runs: Vec<i64> = sqlx::query_scalar("SELECT run_id FROM candidate_runs")
+            .fetch_all(db.pool())
+            .await
+            .unwrap();
+        assert_eq!(runs, vec![new_run]);
+    }
+}

src/db.rs

@@ -328,6 +328,45 @@ impl Db {
         Ok(rows.iter().map(|r| r.get::<i64, _>("article_id")).collect())
     }
 
+    /// Article ids published before this issue date; same-date regeneration is allowed (§8.1).
+    pub async fn previously_published_ids_before(&self, date: Date) -> Result<Vec<ArticleId>> {
+        let rows =
+            sqlx::query("SELECT DISTINCT article_id FROM issue_articles WHERE issue_date < ?")
+                .bind(date.to_string())
+                .fetch_all(&self.pool)
+                .await?;
+        Ok(rows
+            .iter()
+            .map(|row| row.get::<i64, _>("article_id"))
+            .collect())
+    }
+
+    /// Published article ids first seen at or after `since` (`features backfill`).
+    pub async fn published_article_ids_since(&self, since: Timestamp) -> Result<Vec<ArticleId>> {
+        let rows = sqlx::query(
+            "SELECT DISTINCT ia.article_id FROM issue_articles ia
+             JOIN articles a ON a.id = ia.article_id
+             WHERE a.first_seen >= ?
+             ORDER BY ia.article_id",
+        )
+        .bind(fmt_ts(since))
+        .fetch_all(&self.pool)
+        .await?;
+        Ok(rows
+            .iter()
+            .map(|row| row.get::<i64, _>("article_id"))
+            .collect())
+    }
+
+    /// Every article id first seen at or after `since` (`features backfill --all`).
+    pub async fn article_ids_since(&self, since: Timestamp) -> Result<Vec<ArticleId>> {
+        let rows = sqlx::query("SELECT id FROM articles WHERE first_seen >= ? ORDER BY id")
+            .bind(fmt_ts(since))
+            .fetch_all(&self.pool)
+            .await?;
+        Ok(rows.iter().map(|row| row.get::<i64, _>("id")).collect())
+    }
+
     /// Articles the LLM scored below `threshold` within the last `days` (§3.5).
     pub async fn recently_low_scored_ids(
         &self,

src/main.rs

@@ -3,6 +3,7 @@
 //! Everything of substance lives in the library (`src/lib.rs`); this binary only
 //! parses flags, loads config, opens the database and dispatches.
 
+use std::io::Write as _;
 use std::path::PathBuf;
 
 use anyhow::{Context, Result};
@@ -10,6 +11,8 @@ use clap::{Parser, Subcommand, ValueEnum};
 use tracing_subscriber::EnvFilter;
 
 use daily_epub::config::Config;
+use daily_epub::curate::embedding::{self, BACKFILL_CONFIRM_TOKENS};
+use daily_epub::curate::telemetry;
 use daily_epub::db::Db;
 use daily_epub::pipeline::{self, GenerateOptions, GenerateOutcome};
 use daily_epub::report::RunReport;
@@ -40,6 +43,11 @@ enum Command {
     /// Inspect and edit explicit article verdicts.
     #[command(subcommand)]
     Ratings(RatingsCommand),
+    /// Why an article was (not) in the paper, from persisted run telemetry.
+    Explain(ExplainArgs),
+    /// Embedding cache and telemetry maintenance.
+    #[command(subcommand)]
+    Features(FeaturesCommand),
     /// Re-poll social scores for recent entries.
     BackfillSocial(BackfillSocialArgs),
     /// Database maintenance.
@@ -64,6 +72,9 @@ struct GenerateArgs {
     /// Skip every LLM call: prefilter order selects, excerpts stand in for summaries.
     #[arg(long)]
     skip_llm: bool,
+    /// Use cached embeddings only: zero Voyage calls.
+    #[arg(long)]
+    skip_embeddings: bool,
 }
 
 impl From<&GenerateArgs> for GenerateOptions {
@@ -74,6 +85,7 @@ impl From<&GenerateArgs> for GenerateOptions {
             out: args.out.clone(),
             max_articles: args.max_articles,
             skip_llm: args.skip_llm,
+            skip_embeddings: args.skip_embeddings,
         }
     }
 }
@@ -158,6 +170,55 @@ struct RatingsClearArgs {
     url: Option<String>,
 }
 
+/// `explain --date D (--article ID | --url URL) [--run-id N]` or
+/// `explain --date D --near-misses [N]` (plan §15.2).
+#[derive(Debug, clap::Args)]
+struct ExplainArgs {
+    /// Issue date whose run to read.
+    #[arg(long, value_name = "YYYY-MM-DD")]
+    date: String,
+    /// Article id, as printed by `ratings list` or `explain --near-misses`.
+    #[arg(
+        long,
+        required_unless_present_any = ["url", "near_misses"],
+        conflicts_with_all = ["url", "near_misses"]
+    )]
+    article: Option<ArticleId>,
+    /// Article URL; canonicalized before lookup.
+    #[arg(long, conflicts_with = "near_misses")]
+    url: Option<String>,
+    /// A specific run of that date instead of the latest non-dry one.
+    #[arg(long, value_name = "N")]
+    run_id: Option<i64>,
+    /// The top N articles that were considered but not selected (default 10).
+    #[arg(long, value_name = "N", num_args = 0..=1, default_missing_value = "10")]
+    near_misses: Option<usize>,
+}
+
+#[derive(Debug, Subcommand)]
+enum FeaturesCommand {
+    /// Embed rated and published articles, then interests, into the cache.
+    Backfill(BackfillArgs),
+    /// Drop stale embeddings and old candidate telemetry per the retention config.
+    Prune,
+}
+
+#[derive(Debug, clap::Args)]
+struct BackfillArgs {
+    /// Window for published (and, with --all, other) articles.
+    #[arg(long, default_value_t = 30)]
+    days: i64,
+    /// Only the rated set.
+    #[arg(long, conflicts_with = "all")]
+    rated_only: bool,
+    /// Also every other article first seen inside the window.
+    #[arg(long)]
+    all: bool,
+    /// Skip the confirmation prompt above the token threshold.
+    #[arg(long)]
+    yes: bool,
+}
+
 #[derive(Debug, clap::Args)]
 struct BackfillSocialArgs {
     /// How many days back to re-poll.
@@ -196,6 +257,14 @@ async fn main() -> Result<()> {
             let db = Db::open_and_migrate(&config.database_path).await?;
             cmd_ratings(&config, &db, command).await?;
         }
+        Command::Explain(args) => {
+            let db = Db::open_and_migrate(&config.database_path).await?;
+            cmd_explain(&db, args).await?;
+        }
+        Command::Features(command) => {
+            let db = Db::open_and_migrate(&config.database_path).await?;
+            cmd_features(&config, &db, command).await?;
+        }
         Command::BackfillSocial(args) => {
             let db = Db::open_and_migrate(&config.database_path).await?;
             cmd_backfill_social(&db, args.days).await?;
@@ -270,18 +339,27 @@ fn print_report(report: &RunReport) {
         report.counts.duplicates_merged,
         report.counts.entries_dropped,
     );
-    if report.counts.llm_unscored > 0 {
+    let unscored = if report.counts.llm_unscored > 0 {
+        format!(" ({} unscored)", report.counts.llm_unscored)
+    } else {
+        String::new()
+    };
     println!(
-            "curation: {} scored · {} unscored · {} selected",
+        "curation: {} eligible · {} embedded · {} rated w/ embeddings → {} candidates → {} scored{unscored} → {} selected",
-            report.counts.llm_scored, report.counts.llm_unscored, report.counts.selected,
+        report.counts.eligible,
+        report.counts.embedded,
+        report.counts.rated_with_embeddings,
+        report.counts.candidates,
+        report.counts.llm_scored,
+        report.counts.selected,
     );
-    }
     println!(
-        "tokens: {} input · {} cache read · {} cache write · {} output = ${:.4}",
+        "tokens: {} input · {} cache read · {} cache write · {} output · {} voyage = ${:.4}",
         report.usage.input_tokens,
         report.usage.cached_tokens,
         report.usage.cache_write_tokens,
         report.usage.output_tokens,
+        report.voyage_tokens,
         report.cost_usd,
     );
     for (provider, usage) in &report.provider_costs {
@@ -294,6 +372,10 @@ fn print_report(report: &RunReport) {
             usage.cost_usd,
         );
     }
+    println!(
+        "  voyage: {} tokens = ${:.4}",
+        report.voyage_tokens, report.voyage_cost_usd
+    );
     for warning in &report.warnings {
         println!("warning: {warning}");
     }
@@ -478,6 +560,101 @@ async fn cmd_ratings(config: &Config, db: &Db, command: RatingsCommand) -> Resul
     Ok(())
 }
 
+async fn cmd_explain(db: &Db, args: ExplainArgs) -> Result<()> {
+    let date: jiff::civil::Date = args
+        .date
+        .parse()
+        .with_context(|| format!("invalid --date {:?}, expected YYYY-MM-DD", args.date))?;
+    let text = if let Some(limit) = args.near_misses {
+        telemetry::explain_near_misses(db, date, args.run_id, limit).await?
+    } else {
+        let target = match (args.article, args.url) {
+            (Some(id), _) => telemetry::ExplainTarget::Article(id),
+            (None, Some(url)) => telemetry::ExplainTarget::Url(url),
+            (None, None) => anyhow::bail!("provide --article, --url or --near-misses"),
+        };
+        telemetry::explain(db, date, args.run_id, &target).await?
+    };
+    print!("{text}");
+    Ok(())
+}
+
+async fn cmd_features(config: &Config, db: &Db, command: FeaturesCommand) -> Result<()> {
+    match command {
+        FeaturesCommand::Backfill(args) => {
+            if !config.voyage.enabled {
+                anyhow::bail!("voyage.enabled is false; nothing to backfill");
+            }
+            let service = embedding::EmbeddingService::real(db.clone(), config.voyage.clone())
+                .context("building the Voyage client")?;
+            let opts = embedding::BackfillOptions {
+                days: args.days,
+                rated_only: args.rated_only,
+                all: args.all,
+            };
+            let plan = embedding::plan_backfill(db, config, &service, &opts).await?;
+            println!(
+                "backfill: {} articles ({} learned, {} other) + {} interests to embed, {} already cached",
+                plan.article_count(),
+                plan.learned.len(),
+                plan.others.len(),
+                plan.interests.len(),
+                plan.cached
+            );
+            if plan.is_empty() {
+                println!("cache is warm; nothing to do");
+                return Ok(());
+            }
+            println!(
+                "estimate: ~{} tokens ≈ ${:.4} with {} at ${:.2}/M",
+                plan.estimated_tokens,
+                plan.estimated_cost_usd(),
+                config.voyage.model,
+                embedding::VOYAGE_PRICE_PER_MTOK
+            );
+            if plan.estimated_tokens > BACKFILL_CONFIRM_TOKENS
+                && !args.yes
+                && !confirm("continue?")?
+            {
+                println!("aborted");
+                return Ok(());
+            }
+            let outcome = embedding::run_backfill(&service, &plan).await?;
+            println!(
+                "embedded {} articles and {} interests · {} tokens · ${:.4}",
+                outcome.articles_embedded,
+                outcome.interests_embedded,
+                outcome.tokens,
+                outcome.cost_usd
+            );
+        }
+        FeaturesCommand::Prune => {
+            let ranking = &config.curation.ranking;
+            let (embeddings, rows) = telemetry::prune(
+                db,
+                ranking.embedding_retention_days,
+                ranking.telemetry_retention_days,
+                jiff::Timestamp::now(),
+            )
+            .await?;
+            println!(
+                "pruned {embeddings} embeddings older than {} days and {rows} candidate rows older than {} days",
+                ranking.embedding_retention_days, ranking.telemetry_retention_days
+            );
+        }
+    }
+    Ok(())
+}
+
+/// A y/N question on stdin; anything but a leading `y` is a no.
+fn confirm(question: &str) -> Result<bool> {
+    print!("{question} [y/N] ");
+    std::io::stdout().flush()?;
+    let mut answer = String::new();
+    std::io::stdin().read_line(&mut answer)?;
+    Ok(answer.trim().to_lowercase().starts_with('y'))
+}
+
 async fn cmd_backfill_social(db: &Db, days: u32) -> Result<()> {
     let http = http::build_client(http::DEFAULT_TIMEOUT)?;
     let enricher = social::SocialEnricher::new(http, db.clone());
@@ -509,6 +686,7 @@ mod tests {
             "--max-articles",
             "6",
             "--skip-llm",
+            "--skip-embeddings",
         ])
         .unwrap();
         match cli.command {
@@ -518,10 +696,11 @@ mod tests {
                 assert_eq!(a.out, Some(PathBuf::from("./out")));
                 assert_eq!(a.max_articles, Some(6));
                 assert!(a.skip_llm);
+                assert!(a.skip_embeddings);
 
                 let opts = GenerateOptions::from(&a);
                 assert_eq!(opts.date.as_deref(), Some("2026-08-15"));
-                assert!(opts.dry_run && opts.skip_llm);
+                assert!(opts.dry_run && opts.skip_llm && opts.skip_embeddings);
                 assert_eq!(opts.max_articles, Some(6));
             }
             other => panic!("expected generate, got {other:?}"),
@@ -583,6 +762,129 @@ mod tests {
         assert_eq!(cli.config, Some(PathBuf::from("/tmp/x.toml")));
     }
 
+    #[test]
+    fn parses_explain_and_features() {
+        match Cli::try_parse_from([
+            "daily-epub",
+            "explain",
+            "--date",
+            "2026-09-02",
+            "--article",
+            "42",
+            "--run-id",
+            "7",
+        ])
+        .unwrap()
+        .command
+        {
+            Command::Explain(args) => {
+                assert_eq!(args.date, "2026-09-02");
+                assert_eq!(args.article, Some(42));
+                assert_eq!(args.run_id, Some(7));
+                assert_eq!(args.near_misses, None);
+            }
+            other => panic!("expected explain, got {other:?}"),
+        }
+        match Cli::try_parse_from([
+            "daily-epub",
+            "explain",
+            "--date",
+            "2026-09-02",
+            "--url",
+            "https://example.com/post",
+        ])
+        .unwrap()
+        .command
+        {
+            Command::Explain(args) => {
+                assert_eq!(args.url.as_deref(), Some("https://example.com/post"))
+            }
+            other => panic!("expected explain, got {other:?}"),
+        }
+        match Cli::try_parse_from([
+            "daily-epub",
+            "explain",
+            "--date",
+            "2026-09-02",
+            "--near-misses",
+        ])
+        .unwrap()
+        .command
+        {
+            Command::Explain(args) => assert_eq!(args.near_misses, Some(10)),
+            other => panic!("expected explain, got {other:?}"),
+        }
+        match Cli::try_parse_from([
+            "daily-epub",
+            "explain",
+            "--date",
+            "2026-09-02",
+            "--near-misses",
+            "3",
+        ])
+        .unwrap()
+        .command
+        {
+            Command::Explain(args) => assert_eq!(args.near_misses, Some(3)),
+            other => panic!("expected explain, got {other:?}"),
+        }
+        assert!(Cli::try_parse_from(["daily-epub", "explain", "--date", "2026-09-02"]).is_err());
+        assert!(
+            Cli::try_parse_from([
+                "daily-epub",
+                "explain",
+                "--date",
+                "2026-09-02",
+                "--article",
+                "1",
+                "--near-misses"
+            ])
+            .is_err()
+        );
+
+        match Cli::try_parse_from([
+            "daily-epub",
+            "features",
+            "backfill",
+            "--days",
+            "60",
+            "--all",
+            "--yes",
+        ])
+        .unwrap()
+        .command
+        {
+            Command::Features(FeaturesCommand::Backfill(args)) => {
+                assert_eq!(args.days, 60);
+                assert!(args.all && args.yes && !args.rated_only);
+            }
+            other => panic!("expected features backfill, got {other:?}"),
+        }
+        match Cli::try_parse_from(["daily-epub", "features", "backfill"])
+            .unwrap()
+            .command
+        {
+            Command::Features(FeaturesCommand::Backfill(args)) => assert_eq!(args.days, 30),
+            other => panic!("expected features backfill, got {other:?}"),
+        }
+        assert!(
+            Cli::try_parse_from([
+                "daily-epub",
+                "features",
+                "backfill",
+                "--rated-only",
+                "--all"
+            ])
+            .is_err()
+        );
+        assert!(matches!(
+            Cli::try_parse_from(["daily-epub", "features", "prune"])
+                .unwrap()
+                .command,
+            Command::Features(FeaturesCommand::Prune)
+        ));
+    }
+
     #[tokio::test]
     async fn cli_set_and_clear_append_cli_events_with_latest_issue_date() {
         use sqlx::Row as _;

src/pipeline.rs

@@ -23,7 +23,7 @@
 //! issue itself are upserted, `issue_articles` is replaced wholesale, and the
 //! published filenames are derived from the date.
 
-use std::collections::{BTreeMap, BTreeSet};
+use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet};
 use std::path::PathBuf;
 
 use anyhow::{Context, Result};
@@ -32,14 +32,15 @@ use jiff::{Timestamp, Zoned};
 
 use crate::config::Config;
 use crate::curate::llm::{Llms, PriceTable, UsageMeter};
-use crate::curate::{Curator, editorial, profile};
+use crate::curate::{Curator, editorial, embedding, prefilter, profile, signals, telemetry};
 use crate::db::Db;
 use crate::extract::Extractor;
 use crate::miniflux::MinifluxClient;
 use crate::publish::Published;
 use crate::report::{ProviderUsage, RunReport, RunStatus};
 use crate::types::{
-    Article, Artifact, Colophon, Edition, Issue, IssueMeta, Lineup, Models, reading_minutes,
+    Article, ArticleId, Artifact, Colophon, Edition, Issue, IssueMeta, Lineup, Models,
+    reading_minutes,
 };
 use crate::{comments, dedupe, epub, http, miniflux, publish, social, world};
 
@@ -56,6 +57,8 @@ pub struct GenerateOptions {
     pub max_articles: Option<usize>,
     /// `--skip-llm`: no DeepSeek call at all.
     pub skip_llm: bool,
+    /// `--skip-embeddings`: read the cache but make zero Voyage calls.
+    pub skip_embeddings: bool,
 }
 
 /// What one run produced, for the caller to print (§3.13).
@@ -196,10 +199,12 @@ pub async fn generate(config: &Config, db: &Db, opts: &GenerateOptions) -> Resul
         soft_target,
         hard_max,
         skip_llm = opts.skip_llm,
+        skip_embeddings = opts.skip_embeddings,
+        voyage_enabled = config.voyage.enabled,
         out = %out_dir.display(),
         "starting run"
     );
-    log_resolved_providers(config, opts.skip_llm);
+    log_resolved_providers(config, opts.skip_llm, opts.skip_embeddings);
 
     let run_id = db.start_run(date, started_at).await?;
     let mut report = RunReport::new(date, started_at);
@@ -213,6 +218,7 @@ pub async fn generate(config: &Config, db: &Db, opts: &GenerateOptions) -> Resul
     let ctx = StageContext {
         config,
         db,
+        run_id,
         date,
         soft_target,
         hard_max,
@@ -220,6 +226,7 @@ pub async fn generate(config: &Config, db: &Db, opts: &GenerateOptions) -> Resul
         out_dir,
         dry_run: opts.dry_run,
         skip_llm: opts.skip_llm,
+        skip_embeddings: opts.skip_embeddings,
     };
     let stages = match run_stages(&ctx, window_start, window_end, &mut report).await {
         Ok(stages) => {
@@ -275,6 +282,7 @@ struct StageOutput {
 struct StageContext<'a> {
     config: &'a Config,
     db: &'a Db,
+    run_id: i64,
     date: Date,
     soft_target: usize,
     hard_max: usize,
@@ -282,6 +290,7 @@ struct StageContext<'a> {
     out_dir: PathBuf,
     dry_run: bool,
     skip_llm: bool,
+    skip_embeddings: bool,
 }
 
 async fn run_stages(
@@ -369,7 +378,11 @@ async fn run_stages(
     report.counts.social_hits = enricher.enrich_all(&mut articles).await as i64;
     report.timings.record("social", elapsed_ms(stage));
 
-    // --- Stage 6: heuristic pre-filter (§3.5) ---
+    // --- Stage 6: hygiene, embeddings, and cheap signals (§8.1, §9) ---
+    let embeddings = build_embedding_service(ctx, report);
+    let feature_signals = prepare_features(ctx, &articles, &embeddings, report).await;
+
+    // --- Stage 6b: the old heuristic pre-filter still gates in this step (§21) ---
     let stage = Timestamp::now();
     let bulk_meter =
         UsageMeter::with_prices(PriceTable::deepseek(&config.deepseek), config.max_daily_usd);
@@ -399,6 +412,28 @@ async fn run_stages(
         .await
         .context("running the heuristic pre-filter")?;
     report.counts.candidates = candidates.len() as i64;
+    let admitted = candidates
+        .iter()
+        .map(|candidate| candidate.article.id)
+        .collect::<Vec<_>>();
+    let admitted_set = admitted.iter().copied().collect::<HashSet<_>>();
+    let not_admitted = feature_signals
+        .keys()
+        .copied()
+        .filter(|id| !admitted_set.contains(id))
+        .collect::<Vec<_>>();
+    record_stage(
+        ctx,
+        &feature_signals,
+        &not_admitted,
+        "eligible",
+        Some("not_admitted"),
+    )
+    .await
+    .context("recording prefilter telemetry")?;
+    record_stage(ctx, &feature_signals, &admitted, "admitted", None)
+        .await
+        .context("recording prefilter telemetry")?;
     report.timings.record("prefilter", elapsed_ms(stage));
 
     // --- Stage 7: LLM scoring, then selection (§3.6 A + B) ---
@@ -410,12 +445,54 @@ async fn run_stages(
     }
     report.counts.llm_scored = candidates.iter().filter(|c| c.llm.is_some()).count() as i64;
     report.counts.llm_unscored = report.counts.candidates - report.counts.llm_scored;
+    let assessed = candidates
+        .iter()
+        .filter(|candidate| candidate.llm.is_some())
+        .map(|candidate| candidate.article.id)
+        .collect::<Vec<_>>();
+    record_stage(ctx, &feature_signals, &assessed, "assessed", None)
+        .await
+        .context("recording assessment telemetry")?;
+    // Every prefilter survivor goes to the old selector, scored or not.
+    record_stage(ctx, &feature_signals, &admitted, "shortlisted", None)
+        .await
+        .context("recording shortlist telemetry")?;
 
     let mut lineup = curator
         .select(candidates, date)
         .await
         .context("selecting the lineup")?;
     report.counts.selected = lineup.picks.len() as i64;
+    let selected = lineup
+        .picks
+        .iter()
+        .map(|pick| pick.article.id)
+        .collect::<Vec<_>>();
+    let selected_set = selected.iter().copied().collect::<HashSet<_>>();
+    let not_selected = admitted
+        .iter()
+        .copied()
+        .filter(|id| !selected_set.contains(id))
+        .collect::<Vec<_>>();
+    // The editor's one-line `why` (§13) lands in `candidate_runs.editor_why` so
+    // `explain` can quote it; heuristic picks leave it NULL.
+    let selected_with_why = lineup
+        .picks
+        .iter()
+        .map(|pick| (pick.article.id, pick.why.as_deref()))
+        .collect::<Vec<_>>();
+    record_stage_with_why(ctx, &feature_signals, &selected_with_why, "selected", None)
+        .await
+        .context("recording selection telemetry")?;
+    record_stage(
+        ctx,
+        &feature_signals,
+        &not_selected,
+        "shortlisted",
+        Some("not_selected"),
+    )
+    .await
+    .context("recording selection telemetry")?;
     if lineup.picks.is_empty() {
         report.warn("the lineup is empty — check the lookback window and pre-filter");
     }
@@ -577,6 +654,253 @@ async fn run_stages(
     })
 }
 
+/// The cheap signals and hygiene outcome for one eligible article (§9).
+#[derive(Debug, Clone)]
+struct FeatureSignals {
+    signals: signals::Signals,
+    auto_include: bool,
+}
+
+/// The embedding cache with a Voyage client behind it, or cache-only under
+/// `--skip-embeddings`, `voyage.enabled = false` or a missing key (§16, §17).
+fn build_embedding_service(
+    ctx: &StageContext<'_>,
+    report: &mut RunReport,
+) -> embedding::EmbeddingService {
+    let (db, voyage) = (ctx.db.clone(), ctx.config.voyage.clone());
+    if ctx.skip_embeddings {
+        tracing::info!("--skip-embeddings: using cached vectors only, no Voyage calls");
+        return embedding::EmbeddingService::cached_only(db, voyage);
+    }
+    if !voyage.enabled {
+        tracing::info!("voyage disabled: using cached embeddings only");
+        return embedding::EmbeddingService::cached_only(db, voyage);
+    }
+    match embedding::EmbeddingService::real(db.clone(), voyage.clone()) {
+        Ok(service) => service,
+        Err(embedding::EmbeddingError::MissingApiKey) => {
+            tracing::warn!(
+                "voyage enabled but {} is unset; using cached embeddings only",
+                embedding::VOYAGE_API_KEY_ENV
+            );
+            embedding::EmbeddingService::cached_only(db, voyage)
+        }
+        Err(error) => {
+            report.warn(format!(
+                "Voyage unavailable; using cached embeddings only: {error}"
+            ));
+            embedding::EmbeddingService::cached_only(db, voyage)
+        }
+    }
+}
+
+/// Hygiene, embeddings and cheap signals for every article (§8.1, §9).
+///
+/// Hygiene-excluded articles get thin `candidate_runs` rows; every other
+/// article gets an `eligible` row with its `signals_json`. Nothing here can
+/// fail the run: embeddings and the learned signals degrade to absent (§17).
+async fn prepare_features(
+    ctx: &StageContext<'_>,
+    articles: &[Article],
+    service: &embedding::EmbeddingService,
+    report: &mut RunReport,
+) -> HashMap<ArticleId, FeatureSignals> {
+    let (config, db) = (ctx.config, ctx.db);
+    let hygiene = match prefilter::PrefilterContext::load(db, ctx.date).await {
+        Ok(context) => context,
+        Err(error) => {
+            report.warn(format!(
+                "could not load hygiene history; signals skipped: {error}"
+            ));
+            return HashMap::new();
+        }
+    };
+    let published = hygiene
+        .already_published
+        .iter()
+        .copied()
+        .collect::<HashSet<_>>();
+    let rejected = hygiene
+        .recently_rejected
+        .iter()
+        .copied()
+        .collect::<HashSet<_>>();
+    let mut eligible = Vec::new();
+    for article in articles {
+        let auto_include = prefilter::is_auto_include(article, &config.curation);
+        let reason = if published.contains(&article.id) {
+            Some("published_before")
+        } else if !auto_include && prefilter::is_blocked(article, &config.curation) {
+            Some("blocked")
+        } else if !auto_include && rejected.contains(&article.id) {
+            Some("recently_rejected")
+        } else {
+            None
+        };
+        match reason {
+            Some(reason) => {
+                if let Err(error) =
+                    telemetry::thin_excluded(db, ctx.run_id, article.id, reason).await
+                {
+                    report.warn(format!(
+                        "could not record excluded candidate {}: {error}",
+                        article.id
+                    ));
+                }
+            }
+            None => eligible.push(article.clone()),
+        }
+    }
+    report.counts.eligible = eligible.len() as i64;
+
+    // --- embed (§7.1, §7.2) ---
+    let stage = Timestamp::now();
+    let article_embeddings = match service.articles(&eligible).await {
+        Ok(embeddings) => embeddings,
+        Err(error) => {
+            report.warn(format!("article embedding stage degraded: {error}"));
+            HashMap::new()
+        }
+    };
+    report.counts.embedded = article_embeddings.len() as i64;
+    let interests =
+        match profile::load_standing_interests(&config.interests_opml, &config.profile_path) {
+            Ok(interests) => interests,
+            Err(error) => {
+                tracing::warn!(%error, "could not load standing interests for embeddings");
+                Vec::new()
+            }
+        };
+    let interest_embeddings = match service.interests(&interests).await {
+        Ok(embeddings) => embeddings,
+        Err(error) => {
+            report.warn(format!("interest embedding stage degraded: {error}"));
+            HashMap::new()
+        }
+    };
+    if let Some(meter) = service.meter() {
+        report.voyage_tokens = meter.total_tokens();
+        report.voyage_cost_usd = meter.cost_usd();
+    }
+    tracing::info!(
+        eligible = eligible.len(),
+        embedded = article_embeddings.len(),
+        interests = interest_embeddings.len(),
+        voyage_tokens = report.voyage_tokens,
+        "embeddings ready"
+    );
+    report.timings.record("embed", elapsed_ms(stage));
+
+    // --- signals (§9, §12.2, §12.4) ---
+    let stage = Timestamp::now();
+    let ranking = &config.curation.ranking;
+    let (mut computed, preference) = match signals::compute_all(
+        db,
+        &eligible,
+        &article_embeddings,
+        &interest_embeddings,
+        &config.voyage,
+        ranking,
+        Timestamp::now(),
+    )
+    .await
+    {
+        Ok(result) => result,
+        Err(error) => {
+            report.warn(format!("signal computation degraded: {error:#}"));
+            let state = signals::PreferenceState::default();
+            (
+                signals::compute(
+                    &eligible,
+                    &article_embeddings,
+                    &interest_embeddings,
+                    &state,
+                    ranking,
+                ),
+                state.summary(),
+            )
+        }
+    };
+    report.counts.rated_with_embeddings = preference.rated_with_embeddings as i64;
+    let mut output = HashMap::new();
+    for article in &eligible {
+        let auto_include = prefilter::is_auto_include(article, &config.curation);
+        let signals = computed
+            .remove(&article.id)
+            .unwrap_or_else(|| signals::Signals::baseline(article));
+        output.insert(
+            article.id,
+            FeatureSignals {
+                signals,
+                auto_include,
+            },
+        );
+    }
+    let eligible_ids = eligible
+        .iter()
+        .map(|article| article.id)
+        .collect::<Vec<_>>();
+    if let Err(error) = record_stage(ctx, &output, &eligible_ids, "eligible", None).await {
+        report.warn(format!("could not record eligible candidates: {error}"));
+    }
+    report.timings.record("signals", elapsed_ms(stage));
+    output
+}
+
+/// Upsert the `candidate_runs` row of every listed article at a new stage
+/// (§7.4). Articles without signals (hygiene-excluded) are left alone.
+async fn record_stage(
+    ctx: &StageContext<'_>,
+    features: &HashMap<ArticleId, FeatureSignals>,
+    ids: &[ArticleId],
+    stage: &str,
+    excluded_reason: Option<&str>,
+) -> Result<()> {
+    let rows = ids.iter().map(|id| (*id, None)).collect::<Vec<_>>();
+    record_stage_with_why(ctx, features, &rows, stage, excluded_reason).await
+}
+
+/// [`record_stage`] with the editor's `why` per article (§13, §7.4).
+async fn record_stage_with_why(
+    ctx: &StageContext<'_>,
+    features: &HashMap<ArticleId, FeatureSignals>,
+    rows: &[(ArticleId, Option<&str>)],
+    stage: &str,
+    excluded_reason: Option<&str>,
+) -> Result<()> {
+    let admitted = matches!(stage, "admitted" | "assessed" | "shortlisted" | "selected");
+    for (id, editor_why) in rows {
+        let Some(feature) = features.get(id) else {
+            continue;
+        };
+        let json = telemetry::serialize_signals(&feature.signals, feature.auto_include);
+        let admitted_by = admitted.then_some(if feature.auto_include {
+            "[\"auto\"]"
+        } else {
+            "[\"prefilter\"]"
+        });
+        telemetry::write(
+            ctx.db,
+            &telemetry::CandidateRun {
+                run_id: ctx.run_id,
+                article_id: *id,
+                stage,
+                excluded_reason,
+                admitted_by,
+                signals_json: &json,
+                utility: None,
+                rank_utility: None,
+                cluster_id: None,
+                cluster_rank: None,
+                editor_why: *editor_why,
+            },
+        )
+        .await
+        .with_context(|| format!("recording candidate {id} at stage {stage}"))?;
+    }
+    Ok(())
+}
+
 /// Insert/refresh the `articles` rows and stamp the returned ids back on (§3.13).
 async fn persist_articles(db: &Db, articles: &mut [Article]) -> Result<()> {
     for article in articles.iter_mut() {
@@ -701,9 +1025,10 @@ pub fn issue_size_bounds(config: &Config, max_articles: Option<usize>) -> (usize
 }
 
 /// Startup line naming the resolved models and whether each provider is on
-/// (§19): the root config ignores unknown sections, so an `[anthropics]` typo
+/// (§19): the root config ignores unknown sections, so an `[anthropics]` or
-/// would otherwise be silent. Keys are never logged, only their presence.
+/// `[voyages]` typo would otherwise be silent. Keys are never logged, only
-fn log_resolved_providers(config: &Config, skip_llm: bool) {
+/// their presence.
+fn log_resolved_providers(config: &Config, skip_llm: bool, skip_embeddings: bool) {
     let has_key = |key: Option<&str>| key.is_some_and(|k| !k.trim().is_empty());
     tracing::info!(
         bulk_model = %config.deepseek.model,
@@ -716,6 +1041,12 @@ fn log_resolved_providers(config: &Config, skip_llm: bool) {
         editor_effort = %config.anthropic.effort,
         editor_max_daily_usd = config.anthropic.max_daily_usd,
         summary_model = ?config.editorial.summary_model,
+        embedding_model = %config.voyage.model,
+        embedding_enabled = !skip_embeddings
+            && config.voyage.enabled
+            && has_key(config.voyage.api_key.as_deref()),
+        embedding_dimension = config.voyage.output_dimension,
+        embedding_max_daily_usd = config.voyage.max_daily_usd,
         "resolved providers"
     );
 }
@@ -730,20 +1061,25 @@ const PROMPT_VERSIONS: &[(&str, u32)] = &[
     ("profile", 2),
 ];
 
-/// The resolved `[curation]`, `[editorial]`, model names and prompt versions
+/// The resolved `[curation]` (ranking included), `[editorial]`, `[voyage]`,
-/// written to `runs.config_json` (§7.6, §19). Never includes keys.
+/// model names and prompt versions written to `runs.config_json` (§7.6, §19).
+/// Never includes keys.
 fn resolved_run_config(config: &Config, soft_target: usize, hard_max: usize) -> serde_json::Value {
     let mut curation = config.curation.clone();
     curation.max_article_count = hard_max;
+    let mut voyage = config.voyage.clone();
+    voyage.api_key = None;
     serde_json::json!({
         "target_article_count": soft_target,
         "prefilter_keep": config.prefilter_keep,
         "curation": curation,
         "editorial": config.editorial,
+        "voyage": voyage,
         "models": {
             "bulk": config.deepseek.model,
             "editor": if config.anthropic.enabled { config.anthropic.model.as_str() } else { "disabled" },
             "editor_effort": config.anthropic.effort,
+            "embedding": if config.voyage.enabled { config.voyage.model.as_str() } else { "disabled" },
         },
         "prompt_versions": PROMPT_VERSIONS
             .iter()
@@ -815,9 +1151,19 @@ mod tests {
         let mut config = Config::default();
         config.anthropic.api_key = Some("sk-secret".into());
         config.deepseek.api_key = Some("ds-secret".into());
+        config.voyage.api_key = Some("pa-secret".into());
         let value = resolved_run_config(&config, 6, 6);
         assert_eq!(value["target_article_count"], 6);
         assert_eq!(value["curation"]["max_article_count"], 6);
+        assert_eq!(value["curation"]["ranking"]["deep_keep"], 120);
+        assert_eq!(
+            value["curation"]["ranking"]["weights"]["utility"]["quality"],
+            0.40
+        );
+        assert_eq!(value["voyage"]["model"], "voyage-4-lite");
+        assert_eq!(value["voyage"]["output_dimension"], 512);
+        assert!(value["voyage"]["api_key"].is_null());
+        assert_eq!(value["models"]["embedding"], "voyage-4-lite");
         assert_eq!(value["editorial"]["summary_model"], "editor");
         assert_eq!(value["editorial"]["summary_input_tokens"], 3000);
         assert_eq!(value["models"]["bulk"], "deepseek-v4-flash");
@@ -865,4 +1211,344 @@ mod tests {
         assert_eq!(lineup.picks[0].summary.as_deref(), Some("An abstract."));
         assert!(lineup.picks[1..].iter().all(|p| p.summary.is_none()));
     }
+
+    use std::sync::Arc;
+
+    use crate::curate::embedding::{EmbeddingClient, EmbeddingService, MockBackend};
+    use crate::types::{Entry, ExtractMethod, SourceKind, SourceRef};
+    use sqlx::Row as _;
+
+    fn now() -> Timestamp {
+        "2026-09-02T09:00:00Z".parse().unwrap()
+    }
+
+    fn run_date() -> Date {
+        "2026-09-02".parse().unwrap()
+    }
+
+    fn fixture_article(entry_id: i64, host: &str, words: usize) -> Article {
+        let url = format!("https://{host}/post-{entry_id}");
+        let body = (0..words)
+            .map(|i| format!("word{i}"))
+            .collect::<Vec<_>>()
+            .join(" ");
+        Article {
+            id: 0,
+            canonical_url: url.clone(),
+            title: format!("Post {entry_id}"),
+            best_entry_id: entry_id,
+            content_html: format!("<p>{body}</p>"),
+            word_count: words as i64,
+            excerpt_only: false,
+            image_count: 0,
+            sources: vec![SourceRef {
+                entry_id,
+                feed_id: 100 + entry_id,
+                feed_title: format!("Feed {entry_id}"),
+                category: None,
+                kind: SourceKind::Feed,
+            }],
+            first_seen: now(),
+            url,
+            author: None,
+            feed_id: 100 + entry_id,
+            feed_title: format!("Feed {entry_id}"),
+            category: None,
+            published_at: None,
+            comments_url: None,
+            image_urls: vec![],
+            social: vec![],
+            extract_method: ExtractMethod::Miniflux,
+        }
+    }
+
+    fn entry_for(article: &Article) -> Entry {
+        Entry {
+            id: article.best_entry_id,
+            feed_id: article.feed_id,
+            feed_title: Some(article.feed_title.clone()),
+            category: None,
+            title: article.title.clone(),
+            url: article.url.clone(),
+            canonical_url: Some(article.canonical_url.clone()),
+            author: None,
+            published_at: None,
+            comments_url: None,
+            raw_content: article.content_html.clone(),
+            fetched_at: now(),
+        }
+    }
+
+    struct Harness {
+        _dir: tempfile::TempDir,
+        db: Db,
+        config: Config,
+        articles: Vec<Article>,
+        run_id: i64,
+    }
+
+    /// Four articles: two ordinary, one on a blocked host, one published yesterday.
+    async fn harness() -> Harness {
+        let dir = tempfile::tempdir().unwrap();
+        let db = Db::open_and_migrate(&dir.path().join("run.db"))
+            .await
+            .unwrap();
+        let mut config = Config::default();
+        config.curation.blocked_domains = vec!["blocked.example".into()];
+        config.voyage.output_dimension = 4;
+        config.target_article_count = 1;
+        config.interests_opml = dir.path().join("interests.opml");
+        std::fs::write(
+            &config.interests_opml,
+            "<opml><body><outline text=\"Writerdeck\"/></body></opml>",
+        )
+        .unwrap();
+        config.profile_path = dir.path().join("profile.md");
+        std::fs::write(&config.profile_path, "# Reader profile\n").unwrap();
+
+        let mut articles = vec![
+            fixture_article(1, "a.example", 1200),
+            fixture_article(2, "b.example", 900),
+            fixture_article(3, "blocked.example", 1500),
+            fixture_article(4, "d.example", 1400),
+        ];
+        let entries = articles.iter().map(entry_for).collect::<Vec<_>>();
+        db.upsert_entries(&entries).await.unwrap();
+        persist_articles(&db, &mut articles).await.unwrap();
+        sqlx::query(
+            "INSERT INTO issues (date, issue_number, generated_at)
+             VALUES ('2026-09-01', 1, '2026-09-01T12:00:00Z')",
+        )
+        .execute(db.pool())
+        .await
+        .unwrap();
+        sqlx::query(
+            "INSERT INTO issue_articles (issue_date, article_id, section)
+             VALUES ('2026-09-01', ?, 'Top Stories')",
+        )
+        .bind(articles[3].id)
+        .execute(db.pool())
+        .await
+        .unwrap();
+        let run_id = db.start_run(run_date(), now()).await.unwrap();
+        Harness {
+            _dir: dir,
+            db,
+            config,
+            articles,
+            run_id,
+        }
+    }
+
+    fn context<'a>(h: &'a Harness, skip_embeddings: bool) -> StageContext<'a> {
+        StageContext {
+            config: &h.config,
+            db: &h.db,
+            run_id: h.run_id,
+            date: run_date(),
+            soft_target: h.config.target_article_count,
+            hard_max: h.config.curation.max_article_count,
+            started_at: now(),
+            out_dir: PathBuf::from("."),
+            dry_run: true,
+            skip_llm: true,
+            skip_embeddings,
+        }
+    }
+
+    fn mock_service(h: &Harness, backend: Arc<MockBackend>) -> EmbeddingService {
+        let client = EmbeddingClient::with_backend(h.config.voyage.clone(), backend);
+        EmbeddingService::with_client(h.db.clone(), h.config.voyage.clone(), client)
+    }
+
+    async fn stage_rows(
+        db: &Db,
+        run_id: i64,
+    ) -> BTreeMap<i64, (String, Option<String>, Option<String>)> {
+        sqlx::query(
+            "SELECT article_id, stage, excluded_reason, admitted_by FROM candidate_runs
+             WHERE run_id = ? ORDER BY article_id",
+        )
+        .bind(run_id)
+        .fetch_all(db.pool())
+        .await
+        .unwrap()
+        .iter()
+        .map(|row| {
+            (
+                row.get::<i64, _>("article_id"),
+                (
+                    row.get::<String, _>("stage"),
+                    row.get::<Option<String>, _>("excluded_reason"),
+                    row.get::<Option<String>, _>("admitted_by"),
+                ),
+            )
+        })
+        .collect()
+    }
+
+    #[tokio::test]
+    async fn mocked_run_writes_a_candidate_runs_row_for_every_considered_article() {
+        let h = harness().await;
+        let ctx = context(&h, false);
+        let backend = Arc::new(MockBackend::auto(4));
+        let service = mock_service(&h, backend.clone());
+        let mut report = RunReport::new(run_date(), now());
+
+        let features = prepare_features(&ctx, &h.articles, &service, &mut report).await;
+        let [a, b, blocked, published] = [
+            h.articles[0].id,
+            h.articles[1].id,
+            h.articles[2].id,
+            h.articles[3].id,
+        ];
+        assert_eq!(
+            features.keys().copied().collect::<BTreeSet<_>>(),
+            BTreeSet::from([a, b])
+        );
+        assert_eq!(report.counts.eligible, 2);
+        assert_eq!(report.counts.embedded, 2);
+        assert_eq!(report.counts.rated_with_embeddings, 0);
+        assert!(report.timings.0.contains_key("embed") && report.timings.0.contains_key("signals"));
+        assert!(report.voyage_tokens > 0);
+        // One batch for the two articles, one for the interest.
+        assert_eq!(backend.calls(), 2);
+        let signals = &features[&a].signals;
+        assert!(signals.heuristic.is_some());
+        assert!(
+            signals.interest.is_some(),
+            "interest present under the raw fallback"
+        );
+        assert!(
+            signals.knn.is_none() && signals.feed.is_none(),
+            "gates closed"
+        );
+        assert!(signals.preliminary.is_some());
+
+        let rows = stage_rows(&h.db, h.run_id).await;
+        assert_eq!(rows.len(), 4, "one row per considered article");
+        assert_eq!(rows[&blocked].0, "excluded");
+        assert_eq!(rows[&blocked].1.as_deref(), Some("blocked"));
+        assert_eq!(rows[&published].0, "excluded");
+        assert_eq!(rows[&published].1.as_deref(), Some("published_before"));
+        assert_eq!(rows[&a].0, "eligible");
+        assert_eq!(rows[&a].1, None);
+        let thin: String =
+            sqlx::query_scalar("SELECT signals_json FROM candidate_runs WHERE article_id = ?")
+                .bind(blocked)
+                .fetch_one(h.db.pool())
+                .await
+                .unwrap();
+        assert_eq!(thin, "{}");
+
+        // The old prefilter and selector, with the stage transitions of step 3.
+        let curator = Curator::new(h.config.clone(), h.db.clone(), Llms::default());
+        let candidates = curator
+            .prefilter(h.articles.clone(), run_date())
+            .await
+            .unwrap();
+        let admitted = candidates.iter().map(|c| c.article.id).collect::<Vec<_>>();
+        assert_eq!(
+            admitted.iter().copied().collect::<BTreeSet<_>>(),
+            BTreeSet::from([a, b])
+        );
+        record_stage(&ctx, &features, &admitted, "admitted", None)
+            .await
+            .unwrap();
+        record_stage(&ctx, &features, &admitted, "shortlisted", None)
+            .await
+            .unwrap();
+        let lineup = curator.select(candidates, run_date()).await.unwrap();
+        let selected = lineup
+            .picks
+            .iter()
+            .map(|p| p.article.id)
+            .collect::<Vec<_>>();
+        assert_eq!(selected.len(), 1);
+        let not_selected = admitted
+            .iter()
+            .copied()
+            .filter(|id| !selected.contains(id))
+            .collect::<Vec<_>>();
+        record_stage(&ctx, &features, &selected, "selected", None)
+            .await
+            .unwrap();
+        record_stage(
+            &ctx,
+            &features,
+            &not_selected,
+            "shortlisted",
+            Some("not_selected"),
+        )
+        .await
+        .unwrap();
+
+        let rows = stage_rows(&h.db, h.run_id).await;
+        assert_eq!(rows.len(), 4);
+        let (winner, loser) = (selected[0], not_selected[0]);
+        assert_eq!(
+            rows[&winner],
+            ("selected".into(), None, Some("[\"prefilter\"]".into()))
+        );
+        assert_eq!(
+            rows[&loser],
+            (
+                "shortlisted".into(),
+                Some("not_selected".into()),
+                Some("[\"prefilter\"]".into())
+            )
+        );
+        let text = telemetry::explain(
+            &h.db,
+            run_date(),
+            Some(h.run_id),
+            &telemetry::ExplainTarget::Article(loser),
+        )
+        .await
+        .unwrap();
+        assert!(
+            text.contains("stage: shortlisted · reason: not_selected"),
+            "{text}"
+        );
+    }
+
+    #[tokio::test]
+    async fn skip_embeddings_makes_zero_voyage_calls_and_uses_the_cache() {
+        let h = harness().await;
+        let ctx = context(&h, true);
+        let mut report = RunReport::new(run_date(), now());
+        let service = build_embedding_service(&ctx, &mut report);
+        assert!(!service.has_client(), "--skip-embeddings is cache-only");
+        assert!(service.meter().is_none());
+        let features = prepare_features(&ctx, &h.articles, &service, &mut report).await;
+        assert_eq!(features.len(), 2);
+        assert_eq!(report.counts.embedded, 0, "nothing cached yet");
+        assert!(features.values().all(|f| f.signals.interest.is_none()));
+        assert!(features.values().all(|f| f.signals.heuristic.is_some()));
+        assert_eq!(report.voyage_tokens, 0);
+    }
+
+    #[tokio::test]
+    async fn a_voyage_failure_degrades_to_absent_signals_and_the_run_continues() {
+        let h = harness().await;
+        let ctx = context(&h, false);
+        let backend = Arc::new(MockBackend::new()); // nothing scripted: every call fails
+        let service = mock_service(&h, backend.clone());
+        let mut report = RunReport::new(run_date(), now());
+        let features = prepare_features(&ctx, &h.articles, &service, &mut report).await;
+        assert!(backend.calls() >= 1);
+        assert_eq!(features.len(), 2);
+        assert_eq!(report.counts.eligible, 2);
+        assert_eq!(report.counts.embedded, 0);
+        assert!(report.error.is_none());
+        for feature in features.values() {
+            assert!(feature.signals.interest.is_none() && feature.signals.knn.is_none());
+            assert!(feature.signals.heuristic.is_some());
+            assert!(
+                feature.signals.preliminary.is_some(),
+                "scored on what is present"
+            );
+        }
+        assert_eq!(stage_rows(&h.db, h.run_id).await.len(), 4);
+    }
 }

src/report.rs

@@ -67,6 +67,12 @@ pub struct StageCounts {
     pub excerpt_only: i64,
     /// Social lookups that returned a hit (§3.4).
     pub social_hits: i64,
+    /// Articles passing hygiene and eligible for personalized signals.
+    pub eligible: i64,
+    /// Eligible articles with a valid embedding.
+    pub embedded: i64,
+    /// Current rated articles with a valid embedding.
+    pub rated_with_embeddings: i64,
     /// Articles surviving the heuristic pre-filter (§3.5).
     pub candidates: i64,
     /// Articles scored by the LLM (§3.6 stage A).
@@ -113,10 +119,13 @@ pub struct RunReport {
     pub counts: StageCounts,
     /// Aggregate usage retained for the legacy `runs` columns.
     pub usage: TokenUsage,
-    /// Provider-keyed usage and cost written to `runs.provider_costs_json`.
+    /// LLM provider-keyed usage and cost written to `runs.provider_costs_json`.
     pub provider_costs: BTreeMap<String, ProviderUsage>,
     /// Resolved curation/editorial/model settings for this run.
     pub config_json: serde_json::Value,
+    /// Voyage document/query tokens and cost for this run.
+    pub voyage_tokens: i64,
+    pub voyage_cost_usd: f64,
     pub cost_usd: f64,
     pub timings: StageTimings,
     /// Ingest window actually used, RFC3339 (§3.1).
@@ -141,6 +150,8 @@ impl RunReport {
             usage: TokenUsage::default(),
             provider_costs: BTreeMap::new(),
             config_json: serde_json::Value::Null,
+            voyage_tokens: 0,
+            voyage_cost_usd: 0.0,
             cost_usd: 0.0,
             timings: StageTimings::default(),
             window_start: None,
@@ -163,11 +174,12 @@ impl RunReport {
         self.error = Some(err.to_string());
     }
 
-    /// Stamp the end time, total provider costs and settle the status.
+    /// Stamp the end time, total provider costs (LLM providers plus Voyage) and
+    /// settle the status.
     pub fn finish(&mut self, finished_at: Timestamp) {
         self.finished_at = Some(finished_at);
         self.usage = TokenUsage::default();
-        self.cost_usd = 0.0;
+        self.cost_usd = self.voyage_cost_usd;
         for provider in self.provider_costs.values() {
             self.usage.add(provider.usage);
             self.cost_usd += provider.cost_usd;
@@ -198,11 +210,12 @@ impl RunReport {
     /// Compact human-readable summary printed at the end of `generate`.
     pub fn summary_line(&self) -> String {
         format!(
-            "{} [{}] {} entries → {} articles → {} candidates → {} selected · ${:.4} · {}s",
+            "{} [{}] {} entries → {} articles → {} eligible → {} candidates → {} selected · ${:.4} · {}s",
             self.date,
             self.status,
             self.counts.entries_fetched,
             self.counts.articles,
+            self.counts.eligible,
             self.counts.candidates,
             self.counts.selected,
             self.cost_usd,
@@ -257,9 +270,12 @@ mod tests {
                 cost_usd: 0.05,
             },
         );
+        r.voyage_tokens = 250_000;
+        r.voyage_cost_usd = 0.005;
         r.finish(ts("2026-08-15T05:36:00Z"));
         assert_eq!(r.status, RunStatus::Ok);
-        assert!((r.cost_usd - 0.4728).abs() < 1e-9);
+        // LLM providers plus Voyage; Voyage tokens stay out of the LLM aggregate.
+        assert!((r.cost_usd - 0.4778).abs() < 1e-9);
         // The legacy aggregate columns are the sum across providers.
         assert_eq!(r.usage, usage(1_000_100, 1_003_000, 2_000, 1_000_800));
         assert_eq!(r.duration_secs(), Some(360));