rework plugin

crates/meta-feeder-sdk/src/cache.rs

@@ -53,6 +53,13 @@ const SUBTITLES_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("s
 /// OpenSubtitles search results keyed by `"<tmdb_id>\x01<lang3>"`. Value is the
 /// resolved subtitle's cid, or the literal `"null"` negative-cache sentinel.
 const OPENSUBTITLES_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("opensubtitles");
+/// Small durable plugin scratch values that must survive a restart but aren't
+/// content-derived — today, the torznab plugin's Prowlarr indexer-set
+/// fingerprint, so a set that changed *while the feeder was down* is still
+/// detected as a change on the next boot. Keys are plugin-namespaced
+/// (`prowlarr:indexer-set`). redb creates the table on open, so an older
+/// `.redb` file gains it transparently.
+const MISC_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("misc");
 
 /// Per-plugin midhash cache. Cheap to clone (the inner `Database` is shared
 /// via `Arc`).
@@ -105,6 +112,7 @@ impl MidhashCache {
             tx.open_table(TMDB_PRINCIPAL_TOPN_TABLE)?;
             tx.open_table(SUBTITLES_TABLE)?;
             tx.open_table(OPENSUBTITLES_TABLE)?;
+            tx.open_table(MISC_TABLE)?;
             tx.commit()?;
         }
         Ok(Self { db: Arc::new(db) })
@@ -112,6 +120,14 @@ impl MidhashCache {
 
     str_table_accessors!(get_midhash, put_midhash, MIDHASH_TABLE);
 
+    str_table_accessors!(
+        /// Read a durable plugin scratch value (see [`MISC_TABLE`]).
+        get_misc,
+        /// Write a durable plugin scratch value (see [`MISC_TABLE`]).
+        put_misc,
+        MISC_TABLE,
+    );
+
     pub fn entry_count(&self) -> Result<u64, redb::Error> {
         let tx = self.db.begin_read()?;
         let table = tx.open_table(MIDHASH_TABLE)?;

crates/meta-feeder-sdk/src/hash.rs

@@ -183,9 +183,9 @@ const NZB_RELEASE_MAX_DIGEST: usize = 64;
 ///
 /// Returns `None` when the id isn't hex or the digest would overflow the
 /// multihash budget (an oversized host) — the caller drops that row.
-pub fn compute_nzb_release_cid(host: &str, release_id: &str) -> Option<String> {
+pub fn compute_nzb_release_cid(api_base: &str, release_id: &str) -> Option<String> {
     let id_bytes = decode_hex_id(release_id)?;
-    let host_b = host.as_bytes();
+    let host_b = api_base.as_bytes();
 
     // identity-multihash digest = varint(host_len) ‖ host ‖ id_bytes
     let mut digest = Vec::with_capacity(2 + host_b.len() + id_bytes.len());

crates/meta-feeder-sdk/src/meta_core.rs

@@ -184,6 +184,48 @@ impl FeederStore {
         &self.meta_core_url
     }
 
+    /// Raw keys containing `contains` (`GET /api/kv/search`). Returns
+    /// `(keys, truncated)`; `truncated` means the server capped the result and
+    /// there may be more.
+    ///
+    /// This is a keyspace scan, so it belongs on rare control-plane paths only —
+    /// today, invalidating the gateway's search-coverage markers when a feeder's
+    /// upstream source set changes. Never call it per search.
+    pub async fn search_keys(&self, contains: &str, limit: usize) -> anyhow::Result<(Vec<String>, bool)> {
+        let url = format!("{}/api/kv/search", self.meta_core_url);
+        let resp = self
+            .http
+            .get(&url)
+            .query(&[("contains", contains), ("limit", &limit.to_string())])
+            .send()
+            .await?;
+        let status = resp.status();
+        if !status.is_success() {
+            anyhow::bail!("meta-core KV search `{contains}` returned {status}");
+        }
+        #[derive(Deserialize)]
+        struct KvSearchResponse {
+            #[serde(default)]
+            keys: Vec<String>,
+            #[serde(default)]
+            truncated: bool,
+        }
+        let body: KvSearchResponse = resp.json().await?;
+        Ok((body.keys, body.truncated))
+    }
+
+    /// Delete one raw KV key (`DELETE /api/kv/value?key=`). A missing key is not
+    /// an error.
+    pub async fn delete_kv(&self, key: &str) -> anyhow::Result<()> {
+        let url = format!("{}/api/kv/value", self.meta_core_url);
+        let resp = self.http.delete(&url).query(&[("key", key)]).send().await?;
+        let status = resp.status();
+        if status.is_success() || status == reqwest::StatusCode::NOT_FOUND {
+            return Ok(());
+        }
+        anyhow::bail!("meta-core KV DELETE {key} returned {status}");
+    }
+
     /// Self-publish a **metadata-only** outcome to meta-core, returning the
     /// **sparse** outcome (`record: None`) so the gateway's `(_, None)` branch
     /// skips it — no double write. Byte outcomes and already-sparse outcomes are

crates/meta-feeder-sdk/src/plugin.rs

@@ -129,6 +129,33 @@ pub fn upstream_id_field(upstream_id: &str) -> String {
     format!("{upstream_id}id")
 }
 
+/// Key-set prefix for the provenance `source/<label>` field (METADATA_KEYS §5).
+/// Each member is its own hash field (`source/gateway:nyaa.si = "true"`) so two
+/// producers that reach the same content by different routes union without a
+/// last-writer-wins clobber. The `<label>` uses `:` internally (not `/`, which
+/// is the key-set path separator) — e.g. `gateway:nyaa.si`, `gateway:tribler`.
+pub const SOURCE_KEYSET_PREFIX: &str = "source/";
+
+/// Stamp the SDK's default provenance member — `source/gateway:<upstream_id>` —
+/// onto a record's fields, but **only if it carries no `source/*` member yet**.
+/// This gives every gateway hit a provenance label for free (so meta-watch never
+/// falls back to "unknown" for a feeder record), while a plugin that knows a
+/// finer origin (torznab → the specific Prowlarr indexer) stamps its own flat
+/// `source/*` facets first (`source/gateway`, `source/<protocol>`,
+/// `source/<indexer-slug>`) and this default then no-ops.
+pub fn stamp_default_source(
+    fields: &mut std::collections::BTreeMap<String, String>,
+    upstream_id: &str,
+) {
+    if fields.keys().any(|k| k.starts_with(SOURCE_KEYSET_PREFIX)) {
+        return;
+    }
+    fields.insert(
+        format!("{SOURCE_KEYSET_PREFIX}gateway:{upstream_id}"),
+        "true".to_string(),
+    );
+}
+
 /// Static plugin contract. Each enabled `upstream_id` is implemented by
 /// exactly one `Box<dyn FeederPlugin>`. Lives behind `&self` in steady state —
 /// called concurrently, owns any interior mutability for caches.