initial commit

.github/workflows/publish.yml

@@ -0,0 +1,59 @@
+name: Publish image
+
+# Build and push the feeder image to the GitHub Container Registry (ghcr.io):
+#   * every push to main  -> the moving `main` tag
+#   * every `v*` tag       -> semver tags (e.g. v1.0.0 -> 1.0.0, 1.0, 1)
+# The build context is the repo root (the workspace) so the vendored
+# `meta-feeder-sdk` path dependency resolves; the Dockerfile lives under
+# feeder-plugin/book-feeder/.
+on:
+  push:
+    branches: [main]
+    tags: ['v*']
+
+env:
+  REGISTRY: ghcr.io
+  IMAGE_NAME: ${{ github.repository }}
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      packages: write
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Log in to GitHub Container Registry
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Extract image metadata
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+          tags: |
+            type=ref,event=branch
+            type=semver,pattern={{version}}
+            type=semver,pattern={{major}}.{{minor}}
+            type=semver,pattern={{major}}
+            type=sha,format=short
+
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: ./feeder-plugin/book-feeder/Dockerfile
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max

Cargo.toml

@@ -0,0 +1,9 @@
+[workspace]
+resolver = "2"
+members = ["crates/*", "feeder-plugin/*"]
+
+[workspace.package]
+edition = "2021"
+rust-version = "1.85"
+license = "MIT"
+repository = "https://github.com/worph/meta-feeder-book"

crates/meta-feeder-sdk/Cargo.toml

@@ -0,0 +1,49 @@
+[package]
+name = "meta-feeder-sdk"
+version = "0.1.0"
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+publish = false
+
+# Source-agnostic foundation shared by the gateway core and every feeder
+# sidecar binary: the `FeederPlugin` trait, the content-addressing hash
+# families, the per-plugin redb cache, the shared HTTP/cache helpers, the
+# structured-query evaluator, and the axum `serve_feeders` harness that wraps
+# any set of `FeederPlugin`s in the feeder HTTP contract.
+#
+# Deliberately libp2p-free / blockstore-free: a feeder finds + fetches bytes;
+# the gateway core owns hashing-into-the-blockstore, the meta-core store-back,
+# and the libp2p wire. Keep heavy deps (librqbit, beetswap, cid, libp2p) OUT
+# of here so feeders that don't need them don't pay for them.
+
+[dependencies]
+# Async runtime + HTTP server (the serve_feeders harness)
+tokio = { version = "1.40", features = ["full"] }
+axum = { version = "0.7", features = ["json"] }
+hyper = { version = "1", features = ["full"] }
+reqwest = { version = "0.12", features = ["json", "rustls-tls", "stream"], default-features = false }
+
+# Serialization
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+base64 = "0.22"
+
+# Utilities
+futures = "0.3"
+async-trait = "0.1"
+anyhow = "1"
+thiserror = "1"
+tracing = "0.1"
+sha2 = "0.10"
+bytes = "1"
+# Title-tag parsing (shared filename_meta — season/episode/quality/codec/languages).
+regex = "1"
+
+# Per-plugin midhash/blob cache. Schema is identical to meta-share v1 — see
+# cache.rs. Pure-Rust single-process embedded store.
+redb = "2"
+
+[dev-dependencies]
+tempfile = "3"
+tokio = { version = "1.40", features = ["full", "test-util"] }

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

@@ -0,0 +1,337 @@
+//! Per-plugin midhash cache backed by redb.
+//!
+//! Each plugin gets its own redb file at
+//! `<state_dir>/gateway/<upstream_id>/cache.redb`. Keys are upstream
+//! `record_id`s (DOI for sci-hub, book id for gutenberg, …); values are the
+//! `Midhash` strings the plugin computed by fetching + hashing the upstream
+//! file.
+//!
+//! The schema is identical to meta-share v1's so old `.redb` files copy
+//! across. redb is synchronous and sub-millisecond for small ops, so calls
+//! run inline from async handlers without `spawn_blocking`.
+
+use std::path::Path;
+use std::sync::Arc;
+
+use redb::{Database, ReadableTableMetadata, TableDefinition};
+
+pub const CACHE_FILENAME: &str = "cache.redb";
+
+const MIDHASH_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("midhash");
+const BLOBS_TABLE: TableDefinition<'_, &str, &[u8]> = TableDefinition::new("blobs");
+const COVER_CID_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("cover_cid");
+const BIBREC_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("bibrec");
+const PREVIEW_CID_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("preview_cid");
+/// Full-file sha2-256 IPFS cid for a previously-fetched torrent payload
+/// (torznab's bt-fetch path). Distinct from `MIDHASH_TABLE` because the
+/// hash families differ: `midhash` is the synthetic midhash256-from-infohash
+/// fallback; `fullhash` records "we actually downloaded the bytes and this is
+/// the IPFS CIDv1 over them."
+const FULLHASH_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("fullhash");
+/// Enumerated torrent file list (JSON `Vec<TorrentFile>`) keyed by the
+/// torrent's `record_id` (infohash).
+const FILELIST_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("filelist");
+/// Cached TMDB *search* results keyed by a stable lookup key. The value is the
+/// JSON encoding of the top `TmdbHit`, or the literal `"null"` negative-cache
+/// sentinel.
+const TMDB_SEARCH_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("tmdb_search");
+/// Cached TMDB `GET /3/tv/{id}` structural details keyed by `tmdbid`.
+const TMDB_TVDETAILS_TABLE: TableDefinition<'_, &str, &str> =
+    TableDefinition::new("tmdb_tvdetails");
+/// Cached TMDB `GET /3/{tv,movie}/{id}/external_ids` keyed by `tmdbid`.
+const TMDB_EXTIDS_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("tmdb_extids");
+/// Cached TMDB `GET /3/movie/{id}` details keyed by `tmdbid`.
+const TMDB_MOVIEDETAILS_TABLE: TableDefinition<'_, &str, &str> =
+    TableDefinition::new("tmdb_moviedetails");
+/// Cached **ranked** TMDB `search/multi` anchors keyed by the normalized
+/// free-text query.
+const TMDB_PRINCIPAL_TOPN_TABLE: TableDefinition<'_, &str, &str> =
+    TableDefinition::new("tmdb_principal_topn");
+/// Subtitle linkage discovered at search-enrich time, keyed by the torrent's
+/// `record_id` (infohash). Value is the JSON encoding of a `Vec<SubtitleLink>`.
+const SUBTITLES_TABLE: TableDefinition<'_, &str, &str> = TableDefinition::new("subtitles");
+/// 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");
+
+/// Per-plugin midhash cache. Cheap to clone (the inner `Database` is shared
+/// via `Arc`).
+#[derive(Clone)]
+pub struct MidhashCache {
+    db: Arc<Database>,
+}
+
+/// Generate the standard `get`/`put` accessor pair for a `&str → &str` redb
+/// table.
+macro_rules! str_table_accessors {
+    ($(#[$gmeta:meta])* $get:ident, $(#[$pmeta:meta])* $put:ident, $table:ident $(,)?) => {
+        $(#[$gmeta])*
+        pub fn $get(&self, key: &str) -> Result<Option<String>, redb::Error> {
+            let tx = self.db.begin_read()?;
+            let table = tx.open_table($table)?;
+            Ok(table.get(key)?.map(|v| v.value().to_string()))
+        }
+
+        $(#[$pmeta])*
+        pub fn $put(&self, key: &str, value: &str) -> Result<(), redb::Error> {
+            let tx = self.db.begin_write()?;
+            {
+                let mut table = tx.open_table($table)?;
+                table.insert(key, value)?;
+            }
+            tx.commit()?;
+            Ok(())
+        }
+    };
+}
+
+impl MidhashCache {
+    pub fn open(cache_dir: &Path) -> Result<Self, redb::Error> {
+        let path = cache_dir.join(CACHE_FILENAME);
+        let db = Database::create(&path)?;
+        {
+            let tx = db.begin_write()?;
+            tx.open_table(MIDHASH_TABLE)?;
+            tx.open_table(BLOBS_TABLE)?;
+            tx.open_table(COVER_CID_TABLE)?;
+            tx.open_table(BIBREC_TABLE)?;
+            tx.open_table(PREVIEW_CID_TABLE)?;
+            tx.open_table(FULLHASH_TABLE)?;
+            tx.open_table(FILELIST_TABLE)?;
+            tx.open_table(TMDB_SEARCH_TABLE)?;
+            tx.open_table(TMDB_TVDETAILS_TABLE)?;
+            tx.open_table(TMDB_EXTIDS_TABLE)?;
+            tx.open_table(TMDB_MOVIEDETAILS_TABLE)?;
+            tx.open_table(TMDB_PRINCIPAL_TOPN_TABLE)?;
+            tx.open_table(SUBTITLES_TABLE)?;
+            tx.open_table(OPENSUBTITLES_TABLE)?;
+            tx.commit()?;
+        }
+        Ok(Self { db: Arc::new(db) })
+    }
+
+    str_table_accessors!(get_midhash, put_midhash, MIDHASH_TABLE);
+
+    pub fn entry_count(&self) -> Result<u64, redb::Error> {
+        let tx = self.db.begin_read()?;
+        let table = tx.open_table(MIDHASH_TABLE)?;
+        Ok(table.len()?)
+    }
+
+    pub fn get_blob(&self, cid: &str) -> Result<Option<Vec<u8>>, redb::Error> {
+        let tx = self.db.begin_read()?;
+        let table = tx.open_table(BLOBS_TABLE)?;
+        Ok(table.get(cid)?.map(|v| v.value().to_vec()))
+    }
+
+    pub fn put_blob(&self, cid: &str, bytes: &[u8]) -> Result<(), redb::Error> {
+        let tx = self.db.begin_write()?;
+        {
+            let mut table = tx.open_table(BLOBS_TABLE)?;
+            table.insert(cid, bytes)?;
+        }
+        tx.commit()?;
+        Ok(())
+    }
+
+    str_table_accessors!(get_cover_cid, put_cover_cid, COVER_CID_TABLE);
+
+    str_table_accessors!(
+        /// Read a previously-stored full-file IPFS cid for `record_id`.
+        get_fullhash,
+        /// Record the IPFS cid produced by a successful full fetch.
+        put_fullhash,
+        FULLHASH_TABLE
+    );
+
+    str_table_accessors!(
+        /// Read the cached torrent file list (JSON) for `record_id`.
+        get_filelist,
+        /// Cache the enumerated torrent file list (JSON `Vec<TorrentFile>`).
+        put_filelist,
+        FILELIST_TABLE
+    );
+
+    str_table_accessors!(
+        /// Read the cached subtitle linkage (JSON `Vec<SubtitleLink>`).
+        get_subtitles,
+        /// Cache the discovered subtitle linkage (JSON `Vec<SubtitleLink>`).
+        put_subtitles,
+        SUBTITLES_TABLE
+    );
+
+    str_table_accessors!(
+        /// Read a cached OpenSubtitles lookup by `"<tmdb_id>\x01<lang3>"`.
+        get_opensubtitles,
+        /// Cache an OpenSubtitles lookup result (cid or `"null"` sentinel).
+        put_opensubtitles,
+        OPENSUBTITLES_TABLE
+    );
+
+    str_table_accessors!(get_preview_cid, put_preview_cid, PREVIEW_CID_TABLE);
+
+    str_table_accessors!(
+        /// Read a cached TMDB search result by lookup key.
+        get_tmdb_search,
+        /// Cache a TMDB search result under `key`.
+        put_tmdb_search,
+        TMDB_SEARCH_TABLE
+    );
+
+    str_table_accessors!(
+        /// Read cached TMDB TV-details JSON by `tmdbid`.
+        get_tmdb_tvdetails,
+        /// Cache TMDB TV-details JSON under `tmdbid`.
+        put_tmdb_tvdetails,
+        TMDB_TVDETAILS_TABLE
+    );
+
+    str_table_accessors!(
+        /// Read cached TMDB external-ids JSON by `tmdbid`.
+        get_tmdb_extids,
+        /// Cache TMDB external-ids JSON under `tmdbid`.
+        put_tmdb_extids,
+        TMDB_EXTIDS_TABLE
+    );
+
+    str_table_accessors!(
+        /// Read cached TMDB movie-details JSON by `tmdbid`.
+        get_tmdb_moviedetails,
+        /// Cache TMDB movie-details JSON under `tmdbid`.
+        put_tmdb_moviedetails,
+        TMDB_MOVIEDETAILS_TABLE
+    );
+
+    str_table_accessors!(
+        /// Read the cached ranked anchor list by normalized query key.
+        get_tmdb_principal_topn,
+        /// Cache the ranked anchor list under the normalized query key.
+        put_tmdb_principal_topn,
+        TMDB_PRINCIPAL_TOPN_TABLE
+    );
+
+    pub fn get_bibrec(
+        &self,
+        record_id: &str,
+    ) -> Result<Option<std::collections::BTreeMap<String, String>>, redb::Error> {
+        let tx = self.db.begin_read()?;
+        let table = tx.open_table(BIBREC_TABLE)?;
+        let Some(v) = table.get(record_id)? else {
+            return Ok(None);
+        };
+        let s = v.value();
+        Ok(serde_json::from_str(s).ok())
+    }
+
+    pub fn put_bibrec(
+        &self,
+        record_id: &str,
+        fields: &std::collections::BTreeMap<String, String>,
+    ) -> Result<(), redb::Error> {
+        let json = serde_json::to_string(fields).unwrap_or_else(|_| "{}".to_string());
+        let tx = self.db.begin_write()?;
+        {
+            let mut table = tx.open_table(BIBREC_TABLE)?;
+            table.insert(record_id, json.as_str())?;
+        }
+        tx.commit()?;
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn tmp_cache_dir() -> tempfile::TempDir {
+        tempfile::tempdir().expect("tempdir")
+    }
+
+    #[test]
+    fn put_then_get_returns_value() {
+        let dir = tmp_cache_dir();
+        let cache = MidhashCache::open(dir.path()).expect("open");
+        assert_eq!(cache.get_midhash("10.1038/x").unwrap(), None);
+        cache.put_midhash("10.1038/x", "bafyXYZ").expect("put");
+        assert_eq!(
+            cache.get_midhash("10.1038/x").unwrap().as_deref(),
+            Some("bafyXYZ")
+        );
+    }
+
+    #[test]
+    fn put_overwrites_existing_value() {
+        let dir = tmp_cache_dir();
+        let cache = MidhashCache::open(dir.path()).expect("open");
+        cache.put_midhash("k", "v1").unwrap();
+        cache.put_midhash("k", "v2").unwrap();
+        assert_eq!(cache.get_midhash("k").unwrap().as_deref(), Some("v2"));
+    }
+
+    #[test]
+    fn values_persist_across_reopens() {
+        let dir = tmp_cache_dir();
+        {
+            let cache = MidhashCache::open(dir.path()).expect("open 1");
+            cache.put_midhash("k", "persistent").unwrap();
+        }
+        let cache = MidhashCache::open(dir.path()).expect("open 2");
+        assert_eq!(
+            cache.get_midhash("k").unwrap().as_deref(),
+            Some("persistent")
+        );
+    }
+
+    #[test]
+    fn tmdb_search_cache_distinguishes_miss_from_absent() {
+        let dir = tmp_cache_dir();
+        let cache = MidhashCache::open(dir.path()).expect("open");
+        assert_eq!(cache.get_tmdb_search("m\u{1}naruto\u{1}").unwrap(), None);
+        cache.put_tmdb_search("m\u{1}naruto\u{1}", "null").unwrap();
+        assert_eq!(
+            cache
+                .get_tmdb_search("m\u{1}naruto\u{1}")
+                .unwrap()
+                .as_deref(),
+            Some("null")
+        );
+        cache
+            .put_tmdb_search("t\u{1}spy x family\u{1}2019", "{\"tmdbid\":120089}")
+            .unwrap();
+        drop(cache);
+        let cache = MidhashCache::open(dir.path()).expect("reopen");
+        assert_eq!(
+            cache
+                .get_tmdb_search("t\u{1}spy x family\u{1}2019")
+                .unwrap()
+                .as_deref(),
+            Some("{\"tmdbid\":120089}")
+        );
+    }
+
+    #[test]
+    fn tmdb_tvdetails_cache_roundtrips() {
+        let dir = tmp_cache_dir();
+        let cache = MidhashCache::open(dir.path()).expect("open");
+        assert_eq!(cache.get_tmdb_tvdetails("120089").unwrap(), None);
+        cache
+            .put_tmdb_tvdetails("120089", "{\"number_of_seasons\":2}")
+            .unwrap();
+        assert_eq!(
+            cache.get_tmdb_tvdetails("120089").unwrap().as_deref(),
+            Some("{\"number_of_seasons\":2}")
+        );
+    }
+
+    #[test]
+    fn entry_count_reflects_inserts() {
+        let dir = tmp_cache_dir();
+        let cache = MidhashCache::open(dir.path()).expect("open");
+        assert_eq!(cache.entry_count().unwrap(), 0);
+        cache.put_midhash("a", "x").unwrap();
+        cache.put_midhash("b", "y").unwrap();
+        assert_eq!(cache.entry_count().unwrap(), 2);
+        cache.put_midhash("a", "x2").unwrap();
+        assert_eq!(cache.entry_count().unwrap(), 2);
+    }
+}

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

@@ -0,0 +1,288 @@
+//! Shared helpers for single-outcome source plugins (`arxiv`, `gutenberg`,
+//! `pubmed`, `wikicommons`, `giphy`, `scihub`).
+//!
+//! These plugins are structurally near-identical: build a reqwest client, map
+//! HTTP status → [`GatewayError`], and front a per-record midhash cache whose
+//! hit/put boilerplate is the same everywhere. Only the duplicated scaffolding
+//! lives here; the genuinely-different bodies stay in the plugin modules.
+
+use std::path::Path;
+use std::time::Duration;
+
+use tracing::{debug, warn};
+
+use crate::cache::MidhashCache;
+use crate::plugin::{ConfigError, HashKind, HashOutcome};
+use crate::types::{DiscoveryRecord, GatewayError, Hash};
+
+/// Build the shared reqwest client every source plugin uses.
+pub fn build_http_client(
+    timeout_secs: u64,
+    user_agent: &str,
+    redirect: Option<reqwest::redirect::Policy>,
+) -> reqwest::Client {
+    let mut builder = reqwest::Client::builder()
+        .timeout(Duration::from_secs(timeout_secs))
+        .user_agent(user_agent);
+    if let Some(policy) = redirect {
+        builder = builder.redirect(policy);
+    }
+    builder
+        .build()
+        .expect("reqwest client build is infallible with rustls-tls")
+}
+
+/// Map a non-success HTTP status to a [`GatewayError`].
+///
+/// - `404` → [`GatewayError::NotFound`]
+/// - `429` → [`GatewayError::RateLimited`] (honours `Retry-After`, default 60s)
+/// - `5xx` → [`GatewayError::Transient`]
+/// - anything else → [`GatewayError::Permanent`]
+pub fn map_status(resp: &reqwest::Response) -> Result<(), GatewayError> {
+    let status = resp.status();
+    if status.is_success() {
+        return Ok(());
+    }
+    let code = status.as_u16();
+    if code == 404 {
+        return Err(GatewayError::NotFound);
+    }
+    if code == 429 {
+        let retry_after = resp
+            .headers()
+            .get(reqwest::header::RETRY_AFTER)
+            .and_then(|v| v.to_str().ok())
+            .and_then(|v| v.parse::<u32>().ok())
+            .unwrap_or(60);
+        return Err(GatewayError::RateLimited {
+            retry_after_s: retry_after,
+        });
+    }
+    if status.is_server_error() {
+        return Err(GatewayError::Transient(format!(
+            "upstream {} returned {status}",
+            resp.url()
+        )));
+    }
+    Err(GatewayError::Permanent(format!(
+        "upstream {} returned {status}",
+        resp.url()
+    )))
+}
+
+/// Open the per-plugin redb midhash cache, wrapping the error with the plugin
+/// name for a clean startup diagnostic.
+pub fn open_midhash_cache(
+    cache_dir: &Path,
+    plugin: &'static str,
+) -> Result<MidhashCache, ConfigError> {
+    MidhashCache::open(cache_dir).map_err(|e| ConfigError::Other {
+        plugin,
+        source: anyhow::anyhow!("open redb cache: {e}"),
+    })
+}
+
+/// Cache-hit fast path for single-outcome plugins.
+///
+/// On hit, returns `Some(vec![sparse outcome])` — hash only, no bytes/record —
+/// so the core's `(_, None)` branch skips re-storing content. `Ok(None)` is a
+/// cache miss; the caller proceeds to fetch + hash.
+pub fn cached_outcome(
+    cache: &MidhashCache,
+    record_id: &str,
+    upstream_id: &'static str,
+) -> Result<Option<Vec<HashOutcome>>, GatewayError> {
+    let Some(cached) = cache
+        .get_midhash(record_id)
+        .map_err(|e| GatewayError::Internal(anyhow::anyhow!("cache get: {e}")))?
+    else {
+        return Ok(None);
+    };
+    debug!(
+        target: "meta-share::gateway",
+        upstream = upstream_id,
+        record_id,
+        "hash cache hit"
+    );
+    Ok(Some(vec![HashOutcome {
+        hash: Hash(cached),
+        hash_kind: HashKind::Sha2_256,
+        bytes: None,
+        record: None,
+        file_extension: None,
+    }]))
+}
+
+/// Persist a freshly-computed `(record_id → cid)` mapping. A write failure is
+/// non-fatal, so it is logged and swallowed.
+pub fn store_midhash(cache: &MidhashCache, record_id: &str, upstream_id: &'static str, cid: &str) {
+    if let Err(e) = cache.put_midhash(record_id, cid) {
+        warn!(
+            target: "meta-share::gateway",
+            upstream = upstream_id,
+            record_id,
+            error = %e,
+            "hash cache put failed (non-fatal)"
+        );
+    }
+}
+
+/// Resolve the per-plugin cache handle, or produce the standard "not
+/// configured" internal error.
+pub fn require_cache<'a>(
+    cache: Option<&'a MidhashCache>,
+    plugin: &str,
+) -> Result<&'a MidhashCache, GatewayError> {
+    cache.ok_or_else(|| {
+        GatewayError::Internal(anyhow::anyhow!(
+            "{plugin} plugin not configured (configure() never called)"
+        ))
+    })
+}
+
+/// Build the single-element outcome vec a single-outcome plugin returns on a
+/// fresh (cache-miss) compute: the full bytes + record, a Sha2_256 IPFS CID,
+/// and an optional file extension.
+pub fn single_outcome(
+    cid: String,
+    bytes: bytes::Bytes,
+    record: DiscoveryRecord,
+    file_extension: Option<String>,
+) -> Vec<HashOutcome> {
+    vec![HashOutcome {
+        hash: Hash(cid),
+        hash_kind: HashKind::Sha2_256,
+        bytes: Some(bytes),
+        record: Some(record),
+        file_extension,
+    }]
+}
+
+/// Percent-encode a query-string value using the unreserved set
+/// (RFC 3986 `A-Za-z0-9-_.~`) plus `+` for space.
+pub fn urlencode(s: &str) -> String {
+    let mut out = String::with_capacity(s.len());
+    for b in s.as_bytes() {
+        match *b {
+            b'A'..=b'Z' | b'a'..=b'z' | b'0'..=b'9' | b'-' | b'_' | b'.' | b'~' => {
+                out.push(*b as char);
+            }
+            b' ' => out.push('+'),
+            other => out.push_str(&format!("%{:02X}", other)),
+        }
+    }
+    out
+}
+
+/// GET `url` and return the body bytes, mapping transport + status errors to
+/// [`GatewayError`].
+pub async fn fetch_bytes(http: &reqwest::Client, url: &str) -> Result<bytes::Bytes, GatewayError> {
+    let resp = http
+        .get(url)
+        .send()
+        .await
+        .map_err(|e| GatewayError::Transient(format!("GET {url}: {e}")))?;
+    map_status(&resp)?;
+    resp.bytes()
+        .await
+        .map_err(|e| GatewayError::Transient(format!("read body {url}: {e}")))
+}
+
+/// Like [`fetch_bytes`] but decodes the body as UTF-8 text.
+pub async fn fetch_text(http: &reqwest::Client, url: &str) -> Result<String, GatewayError> {
+    let resp = http
+        .get(url)
+        .send()
+        .await
+        .map_err(|e| GatewayError::Transient(format!("GET {url}: {e}")))?;
+    map_status(&resp)?;
+    resp.text()
+        .await
+        .map_err(|e| GatewayError::Transient(format!("read body {url}: {e}")))
+}
+
+/// Decode a v1 BitTorrent infohash string into its raw 20 bytes.
+///
+/// Accepts 40 hex chars (case-insensitive) or 32 base32 chars (RFC 4648,
+/// case-insensitive). Returns `None` for anything else.
+pub fn decode_infohash(s: &str) -> Option<[u8; 20]> {
+    let bytes = s.as_bytes();
+    let raw: Vec<u8> = if bytes.len() == 40 && bytes.iter().all(u8::is_ascii_hexdigit) {
+        let mut out = Vec::with_capacity(20);
+        for pair in bytes.chunks(2) {
+            out.push((hex_nibble(pair[0])? << 4) | hex_nibble(pair[1])?);
+        }
+        out
+    } else if bytes.len() == 32 && bytes.iter().all(u8::is_ascii_alphanumeric) {
+        decode_base32_20(bytes)?
+    } else {
+        return None;
+    };
+    raw.try_into().ok()
+}
+
+fn hex_nibble(b: u8) -> Option<u8> {
+    match b {
+        b'0'..=b'9' => Some(b - b'0'),
+        b'a'..=b'f' => Some(b - b'a' + 10),
+        b'A'..=b'F' => Some(b - b'A' + 10),
+        _ => None,
+    }
+}
+
+/// Decode 32 RFC-4648 base32 chars (case-insensitive, no padding) into the 20
+/// bytes they encode.
+fn decode_base32_20(chars: &[u8]) -> Option<Vec<u8>> {
+    let mut out = Vec::with_capacity(20);
+    let mut buffer: u64 = 0;
+    let mut bits: u32 = 0;
+    for &c in chars {
+        let val = match c {
+            b'A'..=b'Z' => c - b'A',
+            b'a'..=b'z' => c - b'a',
+            b'2'..=b'7' => c - b'2' + 26,
+            _ => return None,
+        } as u64;
+        buffer = (buffer << 5) | val;
+        bits += 5;
+        if bits >= 8 {
+            bits -= 8;
+            out.push((buffer >> bits) as u8);
+        }
+    }
+    Some(out)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn decode_infohash_hex_roundtrips() {
+        let hex = "0123456789abcdef0123456789abcdef01234567";
+        let raw = decode_infohash(hex).expect("hex decodes");
+        assert_eq!(raw[0], 0x01);
+        assert_eq!(raw[19], 0x67);
+    }
+
+    #[test]
+    fn decode_infohash_is_case_insensitive_hex() {
+        let lower = decode_infohash("aabbccddeeff00112233445566778899aabbccdd");
+        let upper = decode_infohash("AABBCCDDEEFF00112233445566778899AABBCCDD");
+        assert!(lower.is_some());
+        assert_eq!(lower, upper);
+    }
+
+    #[test]
+    fn decode_infohash_accepts_base32() {
+        let raw = decode_infohash("MFRGGZDFMZTWQ2LKNNWG23TPOBYXE43U");
+        assert!(raw.is_some());
+    }
+
+    #[test]
+    fn decode_infohash_rejects_wrong_shape() {
+        assert!(decode_infohash("tooshort").is_none());
+        assert!(decode_infohash("zz23456789abcdef0123456789abcdef01234567").is_none());
+        assert!(decode_infohash("").is_none());
+    }
+}

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

@@ -0,0 +1,312 @@
+//! Self-describing per-plugin configuration: schema, redaction, merge, and the
+//! generic schema-driven HTML page the feeder serves at `/config`.
+//!
+//! The point of this module is that **the gateway carries no per-plugin config
+//! knowledge** and **the SDK carries no per-plugin field names**. A plugin
+//! declares its fields via [`FeederPlugin::config_schema`](crate::FeederPlugin::config_schema);
+//! everything here (redaction of secrets, merge-on-save, and the rendered form)
+//! is driven entirely by that schema. Add a feeder → it shows up with a working
+//! config form and zero UI/gateway changes.
+//!
+//! ## Persistence & precedence
+//!
+//! Config lives at `cache_dir/config.json` (the per-plugin dir handed to
+//! `configure()`). The plugin reads **file-or-env** at startup: env is the
+//! first-boot seed and stays authoritative until the operator saves through the
+//! UI, at which point `config.json` exists and wins. There is no runtime reload
+//! — a save returns `restart_required: true` (mirrors the gateway's
+//! `gateway-config.json` posture).
+//!
+//! ## Secrets
+//!
+//! Fields of kind [`FieldKind::Secret`] are **write-only**: never returned to
+//! the page (redacted to a `<key>_set: bool`), and a blank incoming value on
+//! save means "keep the stored secret".
+
+use serde::{Deserialize, Serialize};
+use serde_json::{Map, Value};
+
+/// A plugin's full config surface — an ordered list of fields.
+#[derive(Serialize, Deserialize, Clone, Debug, Default)]
+pub struct ConfigSchema {
+    pub fields: Vec<ConfigField>,
+}
+
+impl ConfigSchema {
+    /// Convenience: `true` when the plugin declares no configurable fields (the
+    /// page then renders a "no configuration required" notice).
+    pub fn is_empty(&self) -> bool {
+        self.fields.is_empty()
+    }
+}
+
+/// One configurable field. `record_array` fields nest `fields` for the
+/// per-record sub-schema (e.g. torznab's `indexers`).
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct ConfigField {
+    /// JSON key in the config object.
+    pub key: String,
+    /// Human label for the form.
+    pub label: String,
+    #[serde(rename = "type")]
+    pub kind: FieldKind,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub help: Option<String>,
+    #[serde(default)]
+    pub required: bool,
+    /// Sub-fields for [`FieldKind::RecordArray`]. The FIRST sub-field is treated
+    /// as the record identifier for secret back-fill on merge (e.g. `name`).
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub fields: Option<Vec<ConfigField>>,
+}
+
+impl ConfigField {
+    pub fn text(key: &str, label: &str) -> Self {
+        Self::new(key, label, FieldKind::Text)
+    }
+    pub fn secret(key: &str, label: &str) -> Self {
+        Self::new(key, label, FieldKind::Secret)
+    }
+    pub fn bool(key: &str, label: &str) -> Self {
+        Self::new(key, label, FieldKind::Bool)
+    }
+    pub fn number(key: &str, label: &str) -> Self {
+        Self::new(key, label, FieldKind::Number)
+    }
+    pub fn list(key: &str, label: &str) -> Self {
+        Self::new(key, label, FieldKind::List)
+    }
+    pub fn record_array(key: &str, label: &str, fields: Vec<ConfigField>) -> Self {
+        let mut f = Self::new(key, label, FieldKind::RecordArray);
+        f.fields = Some(fields);
+        f
+    }
+    fn new(key: &str, label: &str, kind: FieldKind) -> Self {
+        Self {
+            key: key.to_string(),
+            label: label.to_string(),
+            kind,
+            help: None,
+            required: false,
+            fields: None,
+        }
+    }
+    pub fn with_help(mut self, help: &str) -> Self {
+        self.help = Some(help.to_string());
+        self
+    }
+    pub fn required(mut self) -> Self {
+        self.required = true;
+        self
+    }
+}
+
+/// Field widget / value type. Drives both rendering and the secret/merge rules.
+#[derive(Serialize, Deserialize, Clone, Copy, Debug, PartialEq, Eq)]
+#[serde(rename_all = "snake_case")]
+pub enum FieldKind {
+    /// Free text → JSON string.
+    Text,
+    /// Write-only secret → JSON string; redacted on read, "blank keeps" on save.
+    Secret,
+    /// Checkbox → JSON bool.
+    Bool,
+    /// Numeric input → JSON number.
+    Number,
+    /// Comma/line list → JSON array of strings.
+    List,
+    /// Repeatable record → JSON array of objects keyed by nested `fields`.
+    RecordArray,
+}
+
+fn is_blank(v: Option<&Value>) -> bool {
+    match v {
+        None | Some(Value::Null) => true,
+        Some(Value::String(s)) => s.is_empty(),
+        _ => false,
+    }
+}
+
+/// Produce a UI-safe view of `values`: secret fields are dropped and replaced
+/// with a `<key>_set` boolean; everything else passes through. Recurses into
+/// `record_array` records for secret sub-fields.
+pub fn redact(values: &Value, schema: &ConfigSchema) -> Value {
+    let src = values.as_object().cloned().unwrap_or_default();
+    Value::Object(redact_object(&src, &schema.fields))
+}
+
+fn redact_object(src: &Map<String, Value>, fields: &[ConfigField]) -> Map<String, Value> {
+    let mut out = Map::new();
+    for f in fields {
+        match f.kind {
+            FieldKind::Secret => {
+                let set = !is_blank(src.get(&f.key));
+                out.insert(format!("{}_set", f.key), Value::Bool(set));
+            }
+            FieldKind::RecordArray => {
+                let subs = f.fields.as_deref().unwrap_or(&[]);
+                let arr = src
+                    .get(&f.key)
+                    .and_then(Value::as_array)
+                    .map(|recs| {
+                        recs.iter()
+                            .map(|r| {
+                                let m = r.as_object().cloned().unwrap_or_default();
+                                Value::Object(redact_object(&m, subs))
+                            })
+                            .collect::<Vec<_>>()
+                    })
+                    .unwrap_or_default();
+                out.insert(f.key.clone(), Value::Array(arr));
+            }
+            _ => {
+                if let Some(v) = src.get(&f.key) {
+                    out.insert(f.key.clone(), v.clone());
+                }
+            }
+        }
+    }
+    out
+}
+
+/// Merge `incoming` (from the UI) onto `base` (the stored/effective config),
+/// schema-aware. Returns the full, UNREDACTED object to persist:
+/// - secret: blank/absent incoming keeps `base`; non-blank overwrites.
+/// - text/number/bool/list: present in incoming overwrites; absent keeps base.
+/// - record_array: incoming array replaces base, but blank secret sub-fields are
+///   back-filled from the base record matched by the first sub-field (identifier).
+pub fn merge(base: &Value, incoming: &Value, schema: &ConfigSchema) -> Value {
+    let base_obj = base.as_object().cloned().unwrap_or_default();
+    let inc_obj = incoming.as_object().cloned().unwrap_or_default();
+    Value::Object(merge_object(&base_obj, &inc_obj, &schema.fields))
+}
+
+fn merge_object(
+    base: &Map<String, Value>,
+    inc: &Map<String, Value>,
+    fields: &[ConfigField],
+) -> Map<String, Value> {
+    let mut out = base.clone();
+    for f in fields {
+        match f.kind {
+            FieldKind::Secret => {
+                if !is_blank(inc.get(&f.key)) {
+                    out.insert(f.key.clone(), inc.get(&f.key).cloned().unwrap());
+                }
+                // blank/absent → keep base (no-op)
+            }
+            FieldKind::RecordArray => {
+                let subs = f.fields.as_deref().unwrap_or(&[]);
+                let id_key = subs.first().map(|s| s.key.clone());
+                let base_recs = base.get(&f.key).and_then(Value::as_array).cloned();
+                let inc_recs = match inc.get(&f.key).and_then(Value::as_array) {
+                    Some(r) => r.clone(),
+                    None => continue, // absent → keep base array untouched
+                };
+                let merged: Vec<Value> = inc_recs
+                    .iter()
+                    .map(|rec| {
+                        let rec_obj = rec.as_object().cloned().unwrap_or_default();
+                        // Find the matching base record by identifier sub-field.
+                        let base_match = match (&id_key, base_recs.as_ref()) {
+                            (Some(idk), Some(brs)) => brs
+                                .iter()
+                                .find(|br| br.get(idk) == rec_obj.get(idk))
+                                .and_then(Value::as_object)
+                                .cloned()
+                                .unwrap_or_default(),
+                            _ => Map::new(),
+                        };
+                        Value::Object(merge_object(&base_match, &rec_obj, subs))
+                    })
+                    .collect();
+                out.insert(f.key.clone(), Value::Array(merged));
+            }
+            _ => {
+                if let Some(v) = inc.get(&f.key) {
+                    out.insert(f.key.clone(), v.clone());
+                }
+            }
+        }
+    }
+    out
+}
+
+/// The generic, schema-driven config page served at `GET /config`. Self-contained
+/// (inline CSS + JS, no external assets) so it loads identically when hit
+/// directly on the feeder OR reverse-proxied through the gateway. All API calls
+/// are relative to `location.pathname`, so the proxy mount prefix is irrelevant.
+pub const CONFIG_PAGE_HTML: &str = include_str!("config_page.html");
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    fn torznab_like() -> ConfigSchema {
+        ConfigSchema {
+            fields: vec![
+                ConfigField::secret("tmdb_token", "TMDB token"),
+                ConfigField::bool("resolve_torrent_files", "Resolve .torrent"),
+                ConfigField::record_array(
+                    "indexers",
+                    "Indexers",
+                    vec![
+                        ConfigField::text("name", "Name"),
+                        ConfigField::secret("api_key", "API key"),
+                    ],
+                ),
+            ],
+        }
+    }
+
+    #[test]
+    fn redact_hides_secrets_and_flags_set() {
+        let v = json!({
+            "tmdb_token": "abc",
+            "resolve_torrent_files": true,
+            "indexers": [{"name":"prowlarr","api_key":"k"}]
+        });
+        let r = redact(&v, &torznab_like());
+        assert_eq!(r["tmdb_token_set"], json!(true));
+        assert!(r.get("tmdb_token").is_none());
+        assert_eq!(r["resolve_torrent_files"], json!(true));
+        assert_eq!(r["indexers"][0]["name"], json!("prowlarr"));
+        assert_eq!(r["indexers"][0]["api_key_set"], json!(true));
+        assert!(r["indexers"][0].get("api_key").is_none());
+    }
+
+    #[test]
+    fn merge_blank_secret_keeps_base() {
+        let base = json!({"tmdb_token": "keepme"});
+        let inc = json!({"tmdb_token": ""});
+        let m = merge(&base, &inc, &torznab_like());
+        assert_eq!(m["tmdb_token"], json!("keepme"));
+    }
+
+    #[test]
+    fn merge_nonblank_secret_overwrites() {
+        let base = json!({"tmdb_token": "old"});
+        let inc = json!({"tmdb_token": "new"});
+        let m = merge(&base, &inc, &torznab_like());
+        assert_eq!(m["tmdb_token"], json!("new"));
+    }
+
+    #[test]
+    fn merge_record_array_backfills_blank_secret_by_identifier() {
+        let base = json!({"indexers":[{"name":"prowlarr","api_key":"secret"}]});
+        // UI re-submits the indexer with a blank api_key → keep the stored one.
+        let inc = json!({"indexers":[{"name":"prowlarr","api_key":""}]});
+        let m = merge(&base, &inc, &torznab_like());
+        assert_eq!(m["indexers"][0]["api_key"], json!("secret"));
+        assert_eq!(m["indexers"][0]["name"], json!("prowlarr"));
+    }
+
+    #[test]
+    fn merge_bool_overwrites_when_present() {
+        let base = json!({"resolve_torrent_files": true});
+        let inc = json!({"resolve_torrent_files": false});
+        let m = merge(&base, &inc, &torznab_like());
+        assert_eq!(m["resolve_torrent_files"], json!(false));
+    }
+}

crates/meta-feeder-sdk/src/config_page.html

@@ -0,0 +1,245 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8" />
+<meta name="viewport" content="width=device-width, initial-scale=1" />
+<title>feeder config</title>
+<style>
+  :root { color-scheme: dark; }
+  * { box-sizing: border-box; }
+  body {
+    margin: 0; padding: 18px;
+    font: 14px/1.5 system-ui, -apple-system, Segoe UI, Roboto, sans-serif;
+    background: #0d0d12; color: #e6e6ea;
+  }
+  h1 { font-size: 16px; margin: 0 0 4px; }
+  .sub { color: #9a9aa6; font-size: 12px; margin: 0 0 16px; }
+  .field { margin: 0 0 16px; }
+  .field > label { display: block; font-weight: 600; margin: 0 0 4px; }
+  .help { color: #8a8a96; font-size: 12px; margin: 3px 0 0; }
+  input[type=text], input[type=password], input[type=number], textarea {
+    width: 100%; padding: 8px 10px; border-radius: 8px;
+    border: 1px solid #2a2a35; background: #15151c; color: #e6e6ea; font: inherit;
+  }
+  textarea { min-height: 56px; resize: vertical; }
+  input[type=checkbox] { width: 16px; height: 16px; vertical-align: middle; }
+  .pill { font-size: 10px; padding: 2px 7px; border-radius: 999px; margin-left: 8px; vertical-align: middle; }
+  .pill.set { background: #14361f; color: #6ee7a0; border: 1px solid #1f5132; }
+  .pill.unset { background: #2a2a35; color: #b9b9c4; border: 1px solid #3a3a45; }
+  .record { border: 1px solid #2a2a35; border-radius: 10px; padding: 12px; margin: 0 0 10px; background: #111118; }
+  .record .field { margin-bottom: 10px; }
+  .row-actions { text-align: right; }
+  button {
+    font: inherit; cursor: pointer; border-radius: 8px; padding: 7px 14px;
+    border: 1px solid #3a3a45; background: #1c1c26; color: #e6e6ea;
+  }
+  button.primary { background: #6d4aff; border-color: #6d4aff; color: #fff; font-weight: 600; }
+  button.danger { background: transparent; border-color: #5a2a2a; color: #e29; padding: 4px 10px; }
+  button.add { margin-top: 4px; }
+  .bar { position: sticky; bottom: 0; padding: 12px 0 2px; background: linear-gradient(#0d0d1200, #0d0d12 30%); }
+  .banner { padding: 9px 12px; border-radius: 8px; margin: 0 0 14px; font-size: 13px; display: none; }
+  .banner.ok { background: #14361f; color: #9ff0bf; border: 1px solid #1f5132; display: block; }
+  .banner.warn { background: #3a2f12; color: #f0d79f; border: 1px solid #5a4a1f; display: block; }
+  .banner.err { background: #3a1414; color: #f0a0a0; border: 1px solid #5a2020; display: block; }
+  .empty { color: #9a9aa6; }
+</style>
+</head>
+<body>
+  <h1 id="title">Plugin configuration</h1>
+  <p class="sub">Secrets are write-only — set values are shown as “set”, never echoed back. Saving requires a feeder restart to take effect.</p>
+  <div id="banner" class="banner"></div>
+  <form id="form"></form>
+  <div class="bar" id="bar" style="display:none">
+    <button type="button" class="primary" id="save">Save</button>
+  </div>
+
+<script>
+// All API calls are relative to this page's path, so the same page works when
+// served directly on the feeder (/config) or reverse-proxied through the
+// gateway (/api/gateway/plugins/<id>/ui/config).
+const BASE = location.pathname.replace(/\/+$/, "");
+let SCHEMA = { fields: [] };
+let VALUES = {};
+
+const $ = (id) => document.getElementById(id);
+function banner(kind, msg) {
+  const b = $("banner");
+  b.className = "banner " + kind;
+  b.textContent = msg;
+}
+
+function el(tag, attrs, ...kids) {
+  const e = document.createElement(tag);
+  for (const [k, v] of Object.entries(attrs || {})) {
+    if (k === "class") e.className = v;
+    else if (k === "text") e.textContent = v;
+    else if (k in e) e[k] = v;
+    else e.setAttribute(k, v);
+  }
+  for (const c of kids) if (c != null) e.append(c);
+  return e;
+}
+
+// Split help text into plain runs + clickable <a> nodes for any http(s) URL.
+// Built as DOM nodes (never innerHTML) so help stays XSS-safe; trailing
+// punctuation is kept outside the link so "see http://host/ui." reads right.
+function linkify(text) {
+  const out = [];
+  const re = /https?:\/\/[^\s]+/g;
+  let last = 0, m;
+  while ((m = re.exec(text)) !== null) {
+    if (m.index > last) out.push(text.slice(last, m.index));
+    let url = m[0], trail = "";
+    const t = url.match(/[.,;:)\]]+$/);
+    if (t) { trail = t[0]; url = url.slice(0, -trail.length); }
+    out.push(el("a", { href: url, target: "_blank", rel: "noopener noreferrer", text: url }));
+    if (trail) out.push(trail);
+    last = m.index + m[0].length;
+  }
+  if (last < text.length) out.push(text.slice(last));
+  return out;
+}
+
+function helpEl(f) { return f.help ? el("p", { class: "help" }, ...linkify(f.help)) : null; }
+
+// Render one scalar field. `vals` is the object holding the value/`_set` flag.
+function fieldInput(f, vals) {
+  if (f.type === "bool") {
+    const cb = el("input", { type: "checkbox" });
+    cb.checked = vals[f.key] === true;
+    cb._get = () => cb.checked;
+    return el("label", {}, cb, " ", el("span", { text: " " + f.label }));
+  }
+  const label = el("label", { text: f.label });
+  let input;
+  if (f.type === "secret") {
+    const isSet = vals[f.key + "_set"] === true;
+    input = el("input", { type: "password", placeholder: isSet ? "•••••••• (blank keeps current)" : "not set" });
+    label.append(el("span", { class: "pill " + (isSet ? "set" : "unset"), text: isSet ? "set" : "not set" }));
+    input._get = () => input.value; // blank => merge keeps stored secret
+  } else if (f.type === "number") {
+    input = el("input", { type: "number" });
+    if (vals[f.key] != null) input.value = vals[f.key];
+    input._get = () => (input.value === "" ? null : Number(input.value));
+  } else if (f.type === "list") {
+    input = el("textarea", { placeholder: "one per line or comma-separated" });
+    const arr = Array.isArray(vals[f.key]) ? vals[f.key] : [];
+    input.value = arr.join("\n");
+    input._get = () => input.value.split(/[\n,]/).map((s) => s.trim()).filter(Boolean);
+  } else {
+    input = el("input", { type: "text" });
+    if (vals[f.key] != null) input.value = vals[f.key];
+    input._get = () => input.value;
+  }
+  const wrap = el("div", { class: "field" }, label, input, helpEl(f));
+  wrap._collect = () => input._get();
+  wrap._key = f.key;
+  // bool path returns a <label>; normalize collect/key for it
+  if (f.type === "bool") {
+    const w = el("div", { class: "field" }, label, helpEl(f));
+    const cb = label.querySelector("input");
+    w._collect = () => cb.checked;
+    w._key = f.key;
+    return w;
+  }
+  return wrap;
+}
+
+// Render a record_array (repeatable group). Returns a wrapper exposing _collect.
+function recordArray(f, vals) {
+  const subs = f.fields || [];
+  const idKey = subs.length ? subs[0].key : null;
+  const list = el("div", {});
+  const records = Array.isArray(vals[f.key]) ? vals[f.key].slice() : [];
+
+  function addRecord(recVals) {
+    const card = el("div", { class: "record" });
+    const inputs = subs.map((sf) => {
+      const fi = fieldInput(sf, recVals || {});
+      card.append(fi);
+      return fi;
+    });
+    const rm = el("button", { type: "button", class: "danger", text: "remove" });
+    rm.onclick = () => card.remove();
+    card.append(el("div", { class: "row-actions" }, rm));
+    card._collect = () => {
+      const o = {};
+      for (const fi of inputs) o[fi._key] = fi._collect();
+      return o;
+    };
+    list.append(card);
+    return card;
+  }
+
+  records.forEach((r) => addRecord(r));
+  const add = el("button", { type: "button", class: "add", text: "+ add " + f.label.toLowerCase().replace(/s$/, "") });
+  add.onclick = () => addRecord({});
+
+  const wrap = el("div", { class: "field" },
+    el("label", { text: f.label }), helpEl(f), list, add);
+  wrap._key = f.key;
+  wrap._collect = () => Array.from(list.children).map((c) => c._collect());
+  return wrap;
+}
+
+function render() {
+  $("title").textContent = (SCHEMA.title || "Plugin") + " configuration";
+  const form = $("form");
+  form.innerHTML = "";
+  if (!SCHEMA.fields || !SCHEMA.fields.length) {
+    form.append(el("p", { class: "empty", text: "No configuration required for this plugin." }));
+    $("bar").style.display = "none";
+    return;
+  }
+  const widgets = [];
+  for (const f of SCHEMA.fields) {
+    const w = f.type === "record_array" ? recordArray(f, VALUES) : fieldInput(f, VALUES);
+    widgets.push(w);
+    form.append(w);
+  }
+  form._widgets = widgets;
+  $("bar").style.display = "block";
+}
+
+function collect() {
+  const out = {};
+  for (const w of $("form")._widgets || []) out[w._key] = w._collect();
+  return out;
+}
+
+async function save() {
+  $("save").disabled = true;
+  try {
+    const res = await fetch(BASE + "/values", {
+      method: "PUT",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(collect()),
+    });
+    if (!res.ok) throw new Error((await res.text()) || res.status);
+    const j = await res.json().catch(() => ({}));
+    if (j.restart_required) banner("warn", "Saved. Restart the feeder container for changes to take effect.");
+    else banner("ok", "Saved.");
+    // Re-fetch so secret pills reflect the new stored state.
+    await load();
+  } catch (e) {
+    banner("err", "Save failed: " + e.message);
+  } finally {
+    $("save").disabled = false;
+  }
+}
+
+async function load() {
+  const [s, v] = await Promise.all([
+    fetch(BASE + "/schema").then((r) => r.json()),
+    fetch(BASE + "/values").then((r) => r.json()),
+  ]);
+  SCHEMA = s || { fields: [] };
+  VALUES = v || {};
+  render();
+}
+
+$("save").onclick = save;
+load().catch((e) => banner("err", "Failed to load config: " + e.message));
+</script>
+</body>
+</html>

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

@@ -0,0 +1,590 @@
+//! Feeder-driven enrichment — the "smart feeder / dumb gateway" pipeline.
+//!
+//! A feeder that self-publishes (has a `META_CORE_URL`) writes its own base
+//! record, then drives the **reused meta-sort enrichment plugins**
+//! (`metamesh-plugin-filename-parser`, `metamesh-plugin-tmdb`) over HTTP. Each
+//! plugin is handed the `metaCoreUrl` and writes its own fields straight back to
+//! meta-core (`PATCH /meta/{cid}` merge), so the gateway never touches
+//! enrichment — full separation of concerns.
+//!
+//! ## Ordering — why we poll
+//!
+//! tmdb searches on `originalTitle`/`movieYear`/`videoType` (which it reads from
+//! the record), and those are produced by filename-parser. So the pipeline is
+//! **filename-parser → tmdb**. The plugin `/process` contract is *async*
+//! (returns `accepted`, then writes to meta-core and POSTs a callback), so after
+//! kicking filename-parser we **poll meta-core** until `originalTitle` appears
+//! (bounded), read the merged fields back, and hand them to tmdb. Polling keeps
+//! the feeder stateless — no callback-correlation bookkeeping. The plugins'
+//! callback POSTs still fire; they hit the feeder's no-op `/enrich/callback`.
+//!
+//! The whole pipeline runs in a spawned background task so it never adds latency
+//! to the libp2p resolve path; the meta-share ingester propagates the enriched
+//! record on its next poll (~eventually consistent).
+
+use std::collections::BTreeMap;
+use std::sync::Arc;
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
+
+use serde::Serialize;
+use tracing::{debug, warn};
+
+use crate::meta_core::{self, Provenance};
+use crate::types::DiscoveryRecord;
+
+/// Operator-supplied enrichment wiring, parsed once from the feeder env.
+#[derive(Clone, Debug)]
+pub struct EnrichmentConfig {
+    /// meta-core root the feeder self-publishes to + hands to the plugins.
+    pub meta_core_url: String,
+    /// `metamesh-plugin-filename-parser` base URL (optional).
+    pub filename_parser_url: Option<String>,
+    /// `metamesh-plugin-tmdb` base URL (optional).
+    pub tmdb_url: Option<String>,
+    /// TMDB v3 key / v4 bearer token, POSTed to tmdb `/configure`.
+    pub tmdb_token: Option<String>,
+    /// TMDB metadata language (default `en-US`).
+    pub tmdb_language: String,
+    /// Stable per-deployment id stamped into `provenance.gatewayPeer`.
+    pub gateway_peer: String,
+    /// URL the plugins POST their completion callback to (the feeder's own
+    /// no-op `/enrich/callback`). Polling is the real completion signal; this
+    /// only keeps the plugins from logging a failed-callback warning.
+    pub callback_url: String,
+    /// `metamesh-plugin-opensubtitles` base URL (optional). When set, the
+    /// pipeline drives subtitle fetching after tmdb resolves the ids — the
+    /// single reference implementation of OpenSubtitles fetching (the gateway
+    /// no longer carries its own).
+    pub opensubtitles_url: Option<String>,
+    /// OpenSubtitles consumer API key, POSTed to the plugin's `/configure`.
+    pub opensubtitles_api_key: Option<String>,
+    /// OpenSubtitles account username (required by the plugin for downloads).
+    pub opensubtitles_username: Option<String>,
+    /// OpenSubtitles account password.
+    pub opensubtitles_password: Option<String>,
+    /// Wanted subtitle languages (CSV of ISO 639-1 codes; default `en`).
+    pub opensubtitles_languages: String,
+}
+
+fn env_nonempty(key: &str) -> Option<String> {
+    std::env::var(key)
+        .ok()
+        .map(|s| s.trim().to_string())
+        .filter(|s| !s.is_empty())
+}
+
+impl EnrichmentConfig {
+    /// Parse from the feeder env. Returns `None` when `META_CORE_URL` is unset —
+    /// the feeder then falls back to the legacy "return record, gateway stores"
+    /// path (backward compatible; used by tests and gateway-stores deployments).
+    pub fn from_env() -> Option<Self> {
+        let meta_core_url = env_nonempty("META_CORE_URL")?;
+        Some(EnrichmentConfig {
+            meta_core_url,
+            filename_parser_url: env_nonempty("FILENAME_PARSER_URL"),
+            tmdb_url: env_nonempty("TMDB_PLUGIN_URL"),
+            tmdb_token: env_nonempty("TMDB_TOKEN").or_else(|| env_nonempty("TORZNAB_TMDB_TOKEN")),
+            tmdb_language: env_nonempty("TMDB_LANGUAGE").unwrap_or_else(|| "en-US".to_string()),
+            gateway_peer: env_nonempty("META_GATEWAY_PEER_ID")
+                .or_else(|| env_nonempty("HOSTNAME"))
+                .unwrap_or_else(|| "gateway-feeder".to_string()),
+            callback_url: env_nonempty("META_FEEDER_CALLBACK_URL")
+                .unwrap_or_else(|| "http://127.0.0.1:8080/enrich/callback".to_string()),
+            opensubtitles_url: env_nonempty("OPENSUBTITLES_PLUGIN_URL"),
+            opensubtitles_api_key: env_nonempty("OPENSUBTITLES_API_KEY"),
+            opensubtitles_username: env_nonempty("OPENSUBTITLES_USERNAME"),
+            opensubtitles_password: env_nonempty("OPENSUBTITLES_PASSWORD"),
+            opensubtitles_languages: env_nonempty("OPENSUBTITLES_LANGUAGES")
+                .unwrap_or_else(|| "en".to_string()),
+        })
+    }
+}
+
+/// How long to wait for filename-parser to merge `originalTitle` before handing
+/// off to tmdb. filename-parser is a pure, fast string parse; this is a generous
+/// ceiling, not an expected wait.
+const FILENAME_PARSE_POLL_TIMEOUT: Duration = Duration::from_secs(5);
+/// How long to wait for tmdb to merge `tmdbid` before fanning its anchor to the
+/// sibling files. tmdb hits the network, so a longer ceiling than filename.
+const TMDB_POLL_TIMEOUT: Duration = Duration::from_secs(20);
+const POLL_INTERVAL: Duration = Duration::from_millis(200);
+
+/// Show-level TMDB fields that are **fanned** from the primary file to every
+/// sibling file in a multi-file (season-pack) torrent — they're identical for
+/// every episode of a show, so we resolve tmdb **once** and copy these across
+/// (avoids N TMDB lookups; the tmdb plugin caches by `cid_midhash256`, which
+/// gateway BT records don't carry, so without this each file would re-hit TMDB).
+/// Per-file fields (`originalTitle`, `season`, `episode`, `fileName`, `sizeByte`,
+/// `quality`, …) are deliberately NOT fanned — filename-parser sets those per file.
+const ANCHOR_EXACT: &[&str] = &[
+    "tmdbid",
+    "imdbid",
+    "poster",
+    "backdrop",
+    "rating",
+    "releasedate",
+    "studio",
+    "genres",
+    "tags",
+];
+const ANCHOR_PREFIXES: &[&str] = &["plot/", "genres/", "studio/", "tags/"];
+
+/// One file (or pack) to enrich: its CID, the filename to parse, and the base
+/// fields already written to meta-core (handed to the plugins as `existingMeta`).
+#[derive(Clone, Debug)]
+pub struct EnrichTarget {
+    pub cid: String,
+    pub file_name: String,
+    pub base: BTreeMap<String, String>,
+}
+
+/// Drives meta-core writes + the enrichment plugins for one feeder. Cheap to
+/// clone (just a reqwest client + an `Arc<config>`), so it can be moved into a
+/// spawned per-resolve task.
+#[derive(Clone)]
+pub struct Enricher {
+    http: reqwest::Client,
+    cfg: Arc<EnrichmentConfig>,
+}
+
+#[derive(Serialize)]
+struct ProcessRequest<'a> {
+    #[serde(rename = "taskId")]
+    task_id: String,
+    cid: &'a str,
+    #[serde(rename = "filePath")]
+    file_path: &'a str,
+    #[serde(rename = "callbackUrl")]
+    callback_url: &'a str,
+    #[serde(rename = "metaCoreUrl")]
+    meta_core_url: &'a str,
+    #[serde(rename = "existingMeta")]
+    existing_meta: &'a BTreeMap<String, String>,
+}
+
+impl Enricher {
+    pub fn new(http: reqwest::Client, cfg: EnrichmentConfig) -> Self {
+        Enricher {
+            http,
+            cfg: Arc::new(cfg),
+        }
+    }
+
+    pub fn meta_core_url(&self) -> &str {
+        &self.cfg.meta_core_url
+    }
+
+    /// `true` if any enrichment plugin is wired (so there's enrichment to drive
+    /// beyond the base-record write).
+    pub fn has_plugins(&self) -> bool {
+        self.cfg.filename_parser_url.is_some() || self.cfg.tmdb_url.is_some()
+    }
+
+    /// Write the feeder's base record into meta-core (merge/upsert) under `cid`.
+    /// Identical shape to the gateway dispatcher's write (provenance + the
+    /// `cids/<cid>` key-set collapse).
+    pub async fn write_base_record(
+        &self,
+        cid: &str,
+        record: &DiscoveryRecord,
+    ) -> anyhow::Result<()> {
+        let prov = Provenance::now(
+            &record.upstream_id,
+            &record.record_id,
+            &self.cfg.gateway_peer,
+        );
+        let body = meta_core::build_metadata_body(record, &prov);
+        meta_core::put_record(&self.http, &self.cfg.meta_core_url, cid, &body).await
+    }
+
+    /// Run the video enrichment pipeline for one resolved CID: filename-parser
+    /// (clean title/year/season/episode) → tmdb (tmdbid/poster/overview). Each
+    /// plugin writes its own fields to meta-core. Best-effort: any plugin
+    /// failure is logged and the pipeline continues (the base record already
+    /// landed, so the item is at worst un-enriched, never lost).
+    pub async fn enrich_video(
+        &self,
+        cid: &str,
+        file_name: &str,
+        base_fields: BTreeMap<String, String>,
+    ) {
+        let mut meta = base_fields;
+
+        // 1. filename-parser → originalTitle / movieYear / videoType / season / episode.
+        if let Some(fp_url) = self.cfg.filename_parser_url.clone() {
+            match self.process(&fp_url, "filename-parser", cid, file_name, &meta).await {
+                Ok(()) => {
+                    // Poll for the merged fields so tmdb gets the clean title.
+                    if let Some(merged) = self
+                        .poll_until_field(cid, "originalTitle", FILENAME_PARSE_POLL_TIMEOUT)
+                        .await
+                    {
+                        meta = merged;
+                    } else {
+                        debug!(
+                            target: "meta-feeder::enrich",
+                            cid, "filename-parser produced no originalTitle within timeout; \
+                                  handing tmdb the base fields"
+                        );
+                    }
+                }
+                Err(e) => warn!(
+                    target: "meta-feeder::enrich",
+                    cid, error = %e, "filename-parser /process failed (non-fatal)"
+                ),
+            }
+        }
+
+        // 2. tmdb → tmdbid / poster / overview / genres (requires fileType=video +
+        //    a clean originalTitle, both now present on `meta`).
+        if let Some(tmdb_url) = self.cfg.tmdb_url.clone() {
+            // Configure the API key before each call — tmdb holds it in memory
+            // only, so this is robust to a tmdb-plugin restart. Cheap (local).
+            if let Err(e) = self.configure_tmdb(&tmdb_url).await {
+                warn!(
+                    target: "meta-feeder::enrich",
+                    cid, error = %e, "tmdb /configure failed (continuing; may soft-skip)"
+                );
+            }
+            if let Err(e) = self.process(&tmdb_url, "tmdb", cid, file_name, &meta).await {
+                warn!(
+                    target: "meta-feeder::enrich",
+                    cid, error = %e, "tmdb /process failed (non-fatal)"
+                );
+            }
+        }
+
+        // 3. opensubtitles → subtitles/<lang3>/<cid> (needs tmdbid/imdbid, which
+        //    tmdb writes async). Poll for tmdbid so the plugin has an id to
+        //    search by, then drive it like any other enrichment plugin.
+        if self.cfg.opensubtitles_url.is_some() {
+            let merged = self
+                .poll_until_field(cid, "tmdbid", TMDB_POLL_TIMEOUT)
+                .await
+                .unwrap_or(meta);
+            self.enrich_subtitles(cid, file_name, &merged).await;
+        }
+    }
+
+    /// Enrich a **bundle** from one torrent (metainfo fan-out): N video files +
+    /// an optional whole-torrent pack record. filename-parser runs on every
+    /// target; tmdb runs **once** on the primary video file and its show-level
+    /// anchor (`tmdbid`/poster/genres/…) is fanned to the sibling files. The
+    /// pack record gets filename-parse only (no tmdb — it's the non-playable
+    /// "release as a unit" search handle, D1). Best-effort throughout.
+    pub async fn enrich_bundle(&self, videos: Vec<EnrichTarget>, pack: Option<EnrichTarget>) {
+        // Pack: filename-parse only, no tmdb.
+        if let Some(p) = pack {
+            self.fire_filename(&p).await;
+        }
+
+        if videos.is_empty() {
+            return;
+        }
+        if videos.len() == 1 {
+            // Degenerate single-video bundle → the plain pipeline (filename → tmdb).
+            let v = &videos[0];
+            self.enrich_video(&v.cid, &v.file_name, v.base.clone()).await;
+            return;
+        }
+
+        // Multi-file: fire filename-parser for every file (fire-and-forget — only
+        // the primary is polled, the rest write async).
+        for v in &videos {
+            self.fire_filename(v).await;
+        }
+
+        // Primary = largest by `sizeByte` (ties → lowest index, i.e. first seen).
+        let primary = pick_primary(&videos);
+        let p = &videos[primary];
+
+        // Poll the primary for filename-parser's clean title, then tmdb it.
+        let merged = self
+            .poll_until_field(&p.cid, "originalTitle", FILENAME_PARSE_POLL_TIMEOUT)
+            .await
+            .unwrap_or_else(|| p.base.clone());
+
+        let Some(tmdb_url) = self.cfg.tmdb_url.clone() else {
+            return; // no tmdb wired → filename-only enrichment, done.
+        };
+        if let Err(e) = self.configure_tmdb(&tmdb_url).await {
+            warn!(target: "meta-feeder::enrich", cid = %p.cid, error = %e, "tmdb /configure failed");
+        }
+        if let Err(e) = self.process(&tmdb_url, "tmdb", &p.cid, &p.file_name, &merged).await {
+            warn!(target: "meta-feeder::enrich", cid = %p.cid, error = %e, "primary tmdb /process failed");
+            return;
+        }
+
+        // Read the resolved anchor back off the primary, then fan it to siblings.
+        let Some(anchor_rec) = self.poll_until_field(&p.cid, "tmdbid", TMDB_POLL_TIMEOUT).await
+        else {
+            debug!(target: "meta-feeder::enrich", cid = %p.cid, "primary produced no tmdbid; nothing to fan");
+            return;
+        };
+        let anchor = extract_anchor(&anchor_rec);
+        if anchor.is_empty() {
+            return;
+        }
+        for (i, v) in videos.iter().enumerate() {
+            if i == primary {
+                continue;
+            }
+            // PUT is merge/upsert — this layers the show-level anchor onto each
+            // file's record without disturbing its per-file fields.
+            if let Err(e) =
+                meta_core::put_record(&self.http, &self.cfg.meta_core_url, &v.cid, &anchor).await
+            {
+                warn!(target: "meta-feeder::enrich", cid = %v.cid, error = %e, "anchor fan PUT failed");
+            }
+        }
+
+        // Subtitles per video: each episode needs its own (its season/episode
+        // differs), but they share the fanned tmdbid. Poll each for tmdbid
+        // (present on the primary, fanned onto the siblings) and drive the
+        // opensubtitles plugin. Best-effort + quota-guarded inside the plugin.
+        if self.cfg.opensubtitles_url.is_some() {
+            for v in &videos {
+                let merged = self
+                    .poll_until_field(&v.cid, "tmdbid", TMDB_POLL_TIMEOUT)
+                    .await
+                    .unwrap_or_else(|| v.base.clone());
+                self.enrich_subtitles(&v.cid, &v.file_name, &merged).await;
+            }
+        }
+    }
+
+    /// Fire filename-parser `/process` for one target (fire-and-forget; the
+    /// plugin merges its fields to meta-core async).
+    async fn fire_filename(&self, t: &EnrichTarget) {
+        if let Some(fp_url) = self.cfg.filename_parser_url.clone() {
+            if let Err(e) = self
+                .process(&fp_url, "filename-parser", &t.cid, &t.file_name, &t.base)
+                .await
+            {
+                warn!(target: "meta-feeder::enrich", cid = %t.cid, error = %e, "filename-parser /process failed");
+            }
+        }
+    }
+
+    /// POST `/process` to a plugin. Returns once the plugin *accepts* the task
+    /// (the actual meta-core write is async on the plugin side; the caller polls
+    /// meta-core for the result when it needs it downstream).
+    async fn process(
+        &self,
+        plugin_url: &str,
+        plugin: &str,
+        cid: &str,
+        file_path: &str,
+        existing_meta: &BTreeMap<String, String>,
+    ) -> anyhow::Result<()> {
+        let url = format!("{}/process", plugin_url.trim_end_matches('/'));
+        let body = ProcessRequest {
+            task_id: format!("{plugin}-{cid}-{}", now_nanos()),
+            cid,
+            file_path,
+            callback_url: &self.cfg.callback_url,
+            meta_core_url: &self.cfg.meta_core_url,
+            existing_meta,
+        };
+        let resp = self.http.post(&url).json(&body).send().await?;
+        if !resp.status().is_success() {
+            let status = resp.status();
+            let text = resp.text().await.unwrap_or_default();
+            anyhow::bail!("{plugin} POST {url} returned {status}: {text}");
+        }
+        Ok(())
+    }
+
+    /// POST `/configure` to the tmdb plugin with the API key + language.
+    async fn configure_tmdb(&self, tmdb_url: &str) -> anyhow::Result<()> {
+        let Some(token) = self.cfg.tmdb_token.clone() else {
+            anyhow::bail!("no TMDB token configured");
+        };
+        let url = format!("{}/configure", tmdb_url.trim_end_matches('/'));
+        let body = serde_json::json!({
+            "config": { "apiKey": token, "language": self.cfg.tmdb_language }
+        });
+        let resp = self.http.post(&url).json(&body).send().await?;
+        if !resp.status().is_success() {
+            let status = resp.status();
+            let text = resp.text().await.unwrap_or_default();
+            anyhow::bail!("tmdb /configure returned {status}: {text}");
+        }
+        Ok(())
+    }
+
+    /// POST `/configure` to the opensubtitles plugin with the API key, account
+    /// credentials, and wanted languages.
+    async fn configure_opensubtitles(&self, url: &str) -> anyhow::Result<()> {
+        let Some(key) = self.cfg.opensubtitles_api_key.clone() else {
+            anyhow::bail!("no OpenSubtitles API key configured");
+        };
+        let cfg_url = format!("{}/configure", url.trim_end_matches('/'));
+        let body = serde_json::json!({
+            "config": {
+                "apiKey": key,
+                "username": self.cfg.opensubtitles_username,
+                "password": self.cfg.opensubtitles_password,
+                "languages": self.cfg.opensubtitles_languages,
+            }
+        });
+        let resp = self.http.post(&cfg_url).json(&body).send().await?;
+        if !resp.status().is_success() {
+            let status = resp.status();
+            let text = resp.text().await.unwrap_or_default();
+            anyhow::bail!("opensubtitles /configure returned {status}: {text}");
+        }
+        Ok(())
+    }
+
+    /// Drive the opensubtitles plugin for one video CID. Needs a `tmdbid` or
+    /// `imdbid` in `meta` to search by (the plugin reads them from `existingMeta`
+    /// and writes the `subtitles/<lang3>/<cid>` key-set straight to meta-core).
+    /// Best-effort: any failure is logged and swallowed.
+    async fn enrich_subtitles(&self, cid: &str, file_name: &str, meta: &BTreeMap<String, String>) {
+        let Some(url) = self.cfg.opensubtitles_url.clone() else {
+            return;
+        };
+        if !meta.contains_key("tmdbid") && !meta.contains_key("imdbid") {
+            debug!(
+                target: "meta-feeder::enrich",
+                cid, "no tmdbid/imdbid resolved; skipping opensubtitles"
+            );
+            return;
+        }
+        if let Err(e) = self.configure_opensubtitles(&url).await {
+            warn!(
+                target: "meta-feeder::enrich",
+                cid, error = %e, "opensubtitles /configure failed (continuing; may soft-skip)"
+            );
+        }
+        if let Err(e) = self.process(&url, "opensubtitles", cid, file_name, meta).await {
+            warn!(
+                target: "meta-feeder::enrich",
+                cid, error = %e, "opensubtitles /process failed (non-fatal)"
+            );
+        }
+    }
+
+    /// Poll `GET /meta/{cid}` until `field` is present (non-empty) or `timeout`
+    /// elapses; returns the merged field map on success.
+    async fn poll_until_field(
+        &self,
+        cid: &str,
+        field: &str,
+        timeout: Duration,
+    ) -> Option<BTreeMap<String, String>> {
+        let deadline = SystemTime::now() + timeout;
+        loop {
+            match meta_core::get_record(&self.http, &self.cfg.meta_core_url, cid).await {
+                Ok(Some(fields)) if fields.get(field).is_some_and(|v| !v.is_empty()) => {
+                    return Some(fields);
+                }
+                Ok(_) => {}
+                Err(e) => debug!(
+                    target: "meta-feeder::enrich",
+                    cid, error = %e, "poll get_record failed"
+                ),
+            }
+            if SystemTime::now() >= deadline {
+                return None;
+            }
+            tokio::time::sleep(POLL_INTERVAL).await;
+        }
+    }
+}
+
+fn now_nanos() -> u128 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_nanos())
+        .unwrap_or(0)
+}
+
+/// Pick the primary video file index: largest by `sizeByte`, ties broken toward
+/// the first (lowest index). Falls back to 0 for an all-unsized bundle.
+fn pick_primary(videos: &[EnrichTarget]) -> usize {
+    videos
+        .iter()
+        .enumerate()
+        .max_by_key(|(i, v)| {
+            let size = v.base.get("sizeByte").and_then(|s| s.parse::<u64>().ok()).unwrap_or(0);
+            // larger size wins; on a tie the smaller index wins (Reverse).
+            (size, std::cmp::Reverse(*i))
+        })
+        .map(|(i, _)| i)
+        .unwrap_or(0)
+}
+
+/// Project a resolved record down to the show-level anchor fields that are safe
+/// to copy to sibling files (see [`ANCHOR_EXACT`] / [`ANCHOR_PREFIXES`]).
+fn extract_anchor(fields: &BTreeMap<String, String>) -> BTreeMap<String, String> {
+    fields
+        .iter()
+        .filter(|(k, _)| {
+            ANCHOR_EXACT.contains(&k.as_str())
+                || ANCHOR_PREFIXES.iter().any(|p| k.starts_with(p))
+        })
+        .map(|(k, v)| (k.clone(), v.clone()))
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn from_env_requires_meta_core_url() {
+        // Not set in the test env → None (legacy gateway-stores path).
+        // (We don't mutate process env here to avoid cross-test races; this just
+        // documents the contract.)
+        assert!(EnrichmentConfig::from_env().is_none() || EnrichmentConfig::from_env().is_some());
+    }
+
+    fn tgt(cid: &str, size: u64) -> EnrichTarget {
+        EnrichTarget {
+            cid: cid.into(),
+            file_name: cid.into(),
+            base: BTreeMap::from_iter([("sizeByte".to_string(), size.to_string())]),
+        }
+    }
+
+    #[test]
+    fn pick_primary_is_largest_then_first() {
+        let v = vec![tgt("a", 100), tgt("b", 900), tgt("c", 900), tgt("d", 50)];
+        // 'b' and 'c' tie at 900 → lowest index ('b', idx 1) wins.
+        assert_eq!(pick_primary(&v), 1);
+    }
+
+    #[test]
+    fn pick_primary_all_unsized_defaults_first() {
+        let v = vec![
+            EnrichTarget { cid: "a".into(), file_name: "a".into(), base: BTreeMap::new() },
+            EnrichTarget { cid: "b".into(), file_name: "b".into(), base: BTreeMap::new() },
+        ];
+        assert_eq!(pick_primary(&v), 0);
+    }
+
+    #[test]
+    fn extract_anchor_keeps_show_level_drops_per_file() {
+        let rec = BTreeMap::from_iter([
+            ("tmdbid".to_string(), "10378".to_string()),
+            ("imdbid".to_string(), "tt1254207".to_string()),
+            ("poster".to_string(), "bafkPOSTER".to_string()),
+            ("genres".to_string(), "Animation|Comedy".to_string()),
+            ("plot/eng".to_string(), "a plot".to_string()),
+            ("tags".to_string(), "tmdb-verified".to_string()),
+            // per-file — must NOT be fanned:
+            ("originalTitle".to_string(), "Big Buck Bunny".to_string()),
+            ("season".to_string(), "1".to_string()),
+            ("episode".to_string(), "3".to_string()),
+            ("fileName".to_string(), "ep3.mkv".to_string()),
+        ]);
+        let a = extract_anchor(&rec);
+        assert!(a.contains_key("tmdbid") && a.contains_key("poster") && a.contains_key("plot/eng"));
+        assert!(a.contains_key("genres") && a.contains_key("tags") && a.contains_key("imdbid"));
+        assert!(!a.contains_key("originalTitle"), "originalTitle is per-file");
+        assert!(!a.contains_key("season") && !a.contains_key("episode"));
+        assert!(!a.contains_key("fileName"));
+    }
+}

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

@@ -0,0 +1,521 @@
+//! Content-addressed identifiers used by feeder plugins' `compute_outcomes`
+//! to hash upstream-fetched bytes.
+//!
+//! Two function families live here:
+//!
+//! - [`compute_midhash256`] / [`compute_midhash256_from_sample`] —
+//!   the size-prefix-plus-middle-1MB-sample identification primitive shared
+//!   with `meta-sort` / `meta-share` / meta-core. Custom multicodec `0x1000`.
+//!   Fast but not IPFS-interop. The `_from_sample` entrypoint takes the
+//!   middle slice directly so plugins that can only fetch a partial range
+//!   (BitTorrent BEP 9, sliced HTTP) don't have to buffer the whole file
+//!   just to throw most of it away. **Rust-only by construction.**
+//! - [`compute_ipfs_cid`] is the standard IPFS CIDv1 produced over the full
+//!   bytes — raw codec `0x55` for a single-chunk file, UnixFS dag-pb codec
+//!   `0x70` rooted on chunked raw-codec leaves for larger files. Output matches
+//!   `ipfs add --cid-version=1 --raw-leaves=true --chunker=size-262144
+//!   --hash=sha2-256 <file>`.
+//!
+//! Cross-implementation parity for `compute_midhash256` with the TypeScript
+//! `FastHash.ts` is load-bearing — keep the fixture tests pinned, a regression
+//! silently breaks dedupe across the platform.
+
+use bytes::Bytes;
+use sha2::{Digest, Sha256};
+
+/// Sample window in bytes (1 MiB). Files this size or smaller are hashed
+/// in full; larger files are hashed over the middle 1 MiB slice.
+const SAMPLE_SIZE: usize = 1024 * 1024;
+
+/// Per-block IPFS chunk size (256 KiB), matches kubo's
+/// `--chunker=size-262144` default.
+pub const IPFS_CHUNK_SIZE: usize = 256 * 1024;
+
+/// dag-pb fanout — children per internal node. Matches kubo's balanced
+/// builder default.
+pub const IPFS_FANOUT: usize = 174;
+
+/// Output of [`compute_ipfs_blocks`]. Carries the root CID **plus every
+/// intermediate block** (leaf chunks AND internal dag-pb nodes) keyed by
+/// their own CID. Pass `.blocks` to the gateway's blockstore so peers
+/// can fetch the file by CID via bitswap.
+#[derive(Debug, Clone)]
+pub struct IpfsBlocks {
+    /// Canonical "bafy…" / "bafk…" root cid string. Identical to what
+    /// [`compute_ipfs_cid`] returns for the same input.
+    pub root: String,
+    /// Every block — root included — keyed by its own cid. For a
+    /// single-leaf file, this is a one-entry vec `[(root, payload)]`
+    /// (raw codec, payload is the file bytes). For multi-leaf files,
+    /// it is `leaves ++ internal_nodes` in build order; the **last**
+    /// entry is the root.
+    pub blocks: Vec<(String, Bytes)>,
+}
+
+/// Compute a midhash256 CID matching the TypeScript implementation in
+/// `meta-hash/src/lib/file-id/FastHash.ts`:
+///
+/// 1. Build the hash input as `[size:u64-be][middle 1 MiB sample]` (or the
+///    whole file if it fits in 1 MiB).
+/// 2. SHA-256 it.
+/// 3. Wrap in CIDv1 with custom multicodec `0x1000` for both the codec field
+///    and the multihash hash-code field, length 32, then multibase-base32-lower
+///    with the `b` prefix.
+///
+/// **Endianness for the size prefix is big-endian** to match TS's
+/// `writeBigUInt64BE`. Cross-impl mismatch here would silently mangle every
+/// import — keep the test pinned.
+pub fn compute_midhash256(bytes: &[u8]) -> String {
+    let size = bytes.len();
+    let sample: &[u8] = if size <= SAMPLE_SIZE {
+        bytes
+    } else {
+        let start = (size - SAMPLE_SIZE) / 2;
+        &bytes[start..start + SAMPLE_SIZE]
+    };
+    compute_midhash256_from_sample(size as u64, sample)
+}
+
+/// Same midhash256 primitive as [`compute_midhash256`], but takes the
+/// sample bytes directly rather than slicing them out of a full file.
+/// For upstreams that can only deliver a middle range (BitTorrent BEP 9,
+/// sliced HTTP), this avoids forcing the caller to buffer the whole file
+/// in memory just to throw most of it away.
+///
+/// `total_size` is the **full file size in bytes** — used as the
+/// size prefix, NOT `middle_sample.len()`.
+pub fn compute_midhash256_from_sample(total_size: u64, middle_sample: &[u8]) -> String {
+    let size_be = total_size.to_be_bytes();
+    let mut hasher = Sha256::new();
+    hasher.update(size_be);
+    hasher.update(middle_sample);
+    let digest: [u8; 32] = hasher.finalize().into();
+
+    // Multicodec 0x1000 encoded as unsigned varint = [0x80, 0x20].
+    // CIDv1 layout: [version=0x01][codec varint][multihash code varint][len][digest]
+    let mut wire = Vec::with_capacity(38);
+    wire.push(0x01);
+    wire.extend_from_slice(&[0x80, 0x20]);
+    wire.extend_from_slice(&[0x80, 0x20]);
+    wire.push(0x20);
+    wire.extend_from_slice(&digest);
+
+    format!("b{}", base32_lower_no_padding(&wire))
+}
+
+/// Wrap a 20-byte BitTorrent v1 info-hash as a multibase CIDv1.
+///
+/// BT v1 info-hashes ARE SHA-1 digests of the bencoded info dict — so
+/// the natural CID form is `codec=raw (0x55)` + `multihash=sha1 (0x11),
+/// len=20, digest=<infohash bytes>`.
+pub fn compute_bt_info_cid(infohash_20: &[u8; 20]) -> String {
+    // CIDv1 layout: [version=0x01][codec varint][multihash code varint][len][digest]
+    let mut wire = Vec::with_capacity(4 + 20);
+    wire.push(0x01); // CIDv1 version
+    wire.push(0x55); // codec: raw
+    wire.push(0x11); // multihash: sha1
+    wire.push(0x14); // multihash length: 20
+    wire.extend_from_slice(infohash_20);
+    format!("b{}", base32_lower_no_padding(&wire))
+}
+
+/// Custom multicodec for a single file inside a **BitTorrent v1** torrent
+/// (`btih-v1-file`). MetaMesh-private, adjacent to midhash256's `0x1000`.
+pub const BTIH_V1_FILE_CODEC: u64 = 0x1001;
+
+/// Custom multicodec for a single file inside a **BitTorrent v2** torrent
+/// (`btih-v2-file`).
+pub const BTIH_V2_FILE_CODEC: u64 = 0x1002;
+
+/// Encode a `btih-v1-file` CID — a single file inside a BT v1 torrent,
+/// addressed by the 20-byte v1 infohash (SHA-1 of the bencoded `info`
+/// dict) plus the zero-based file index in `info.files`. Single-file
+/// torrents use index `0`.
+pub fn compute_bt_v1_file_cid(infohash_20: &[u8; 20], file_index: u64) -> String {
+    bt_file_cid(BTIH_V1_FILE_CODEC, infohash_20, file_index)
+}
+
+/// Encode a `btih-v2-file` CID — a single file inside a BT v2 torrent
+/// (BEP 52), addressed by the 32-byte v2 infohash plus the file's
+/// canonical traversal index in `file tree`.
+pub fn compute_bt_v2_file_cid(infohash_32: &[u8; 32], file_index: u64) -> String {
+    bt_file_cid(BTIH_V2_FILE_CODEC, infohash_32, file_index)
+}
+
+/// Shared encoder for the two torrent-file CID families. `infohash` is 20
+/// bytes for v1 / 32 for v2; `codec` selects the family.
+fn bt_file_cid(codec: u64, infohash: &[u8], file_index: u64) -> String {
+    // digest = infohash ‖ varint(file_index)
+    let mut digest = Vec::with_capacity(infohash.len() + 2);
+    digest.extend_from_slice(infohash);
+    write_pb_varint(file_index, &mut digest);
+
+    // CIDv1: [version][codec varint][multihash code varint][len varint][digest]
+    let mut wire = Vec::with_capacity(1 + 4 + digest.len());
+    wire.push(0x01);
+    write_pb_varint(codec, &mut wire);
+    write_pb_varint(codec, &mut wire);
+    write_pb_varint(digest.len() as u64, &mut wire);
+    wire.extend_from_slice(&digest);
+    format!("b{}", base32_lower_no_padding(&wire))
+}
+
+/// Compute a standard IPFS CIDv1 over `bytes`. Output matches kubo's
+/// `ipfs add --cid-version=1 --raw-leaves=true --chunker=size-262144
+/// --hash=sha2-256 <file>`.
+pub fn compute_ipfs_cid(bytes: &[u8]) -> String {
+    compute_ipfs_blocks(bytes).root
+}
+
+/// Same wire-format guarantee as [`compute_ipfs_cid`], but exposes every
+/// block produced along the way (leaf chunks + internal dag-pb nodes)
+/// so callers can populate a bitswap blockstore in one pass.
+pub fn compute_ipfs_blocks(bytes: &[u8]) -> IpfsBlocks {
+    // Leaves: chunk bytes into raw-codec blocks. Empty input still gets
+    // one (empty) leaf so the cid is well-defined.
+    let mut blocks: Vec<(String, Bytes)> = Vec::new();
+    let leaves: Vec<(Vec<u8>, u64)> = if bytes.is_empty() {
+        let cid_wire = ipfs_cid_wire(0x55, &sha2_256(b""));
+        blocks.push((cid_string(&cid_wire), Bytes::new()));
+        vec![(cid_wire, 0)]
+    } else {
+        bytes
+            .chunks(IPFS_CHUNK_SIZE)
+            .map(|chunk| {
+                let cid_wire = ipfs_cid_wire(0x55, &sha2_256(chunk));
+                blocks.push((cid_string(&cid_wire), Bytes::copy_from_slice(chunk)));
+                (cid_wire, chunk.len() as u64)
+            })
+            .collect()
+    };
+
+    // Single-leaf file: return the leaf cid as the file cid.
+    if leaves.len() == 1 {
+        let root = cid_string(&leaves[0].0);
+        return IpfsBlocks { root, blocks };
+    }
+
+    // Multi-leaf: build a balanced tree of UnixFS dag-pb internal nodes.
+    let mut level: Vec<(Vec<u8>, u64, u64)> = leaves
+        .into_iter()
+        .map(|(cid, size)| (cid, size, size))
+        .collect();
+
+    while level.len() > 1 {
+        let mut next: Vec<(Vec<u8>, u64, u64)> =
+            Vec::with_capacity(level.len().div_ceil(IPFS_FANOUT));
+        for batch in level.chunks(IPFS_FANOUT) {
+            let blocksizes: Vec<u64> = batch.iter().map(|(_, sz, _)| *sz).collect();
+            let filesize: u64 = blocksizes.iter().sum();
+            let unixfs_data = encode_unixfs_file(filesize, &blocksizes);
+            let node_bytes = encode_dagpb_node(batch, &unixfs_data);
+            let node_cid = ipfs_cid_wire(0x70, &sha2_256(&node_bytes));
+            blocks.push((cid_string(&node_cid), Bytes::from(node_bytes.clone())));
+            let own_tsize: u64 =
+                batch.iter().map(|(_, _, t)| *t).sum::<u64>() + node_bytes.len() as u64;
+            next.push((node_cid, filesize, own_tsize));
+        }
+        level = next;
+    }
+    let root = cid_string(&level[0].0);
+    IpfsBlocks { root, blocks }
+}
+
+fn cid_string(wire: &[u8]) -> String {
+    format!("b{}", base32_lower_no_padding(wire))
+}
+
+/// CIDv1 wire-form bytes: `[version=0x01][codec varint][multihash code
+/// varint=0x12 sha2-256][len=0x20][digest...]`.
+fn ipfs_cid_wire(codec: u8, digest: &[u8; 32]) -> Vec<u8> {
+    debug_assert!(
+        codec < 0x80,
+        "codec varint must fit in one byte for this helper"
+    );
+    let mut wire = Vec::with_capacity(36);
+    wire.push(0x01); // CIDv1
+    wire.push(codec);
+    wire.push(0x12); // multihash code: sha2-256
+    wire.push(0x20); // digest length: 32
+    wire.extend_from_slice(digest);
+    wire
+}
+
+fn sha2_256(bytes: &[u8]) -> [u8; 32] {
+    Sha256::digest(bytes).into()
+}
+
+/// LEB128 unsigned varint encoding into `out`.
+fn write_pb_varint(value: u64, out: &mut Vec<u8>) {
+    let mut v = value;
+    while v >= 0x80 {
+        out.push((v as u8) | 0x80);
+        v >>= 7;
+    }
+    out.push(v as u8);
+}
+
+fn write_pb_varint_field(field: u32, value: u64, out: &mut Vec<u8>) {
+    write_pb_varint((field as u64) << 3, out); // tag: (field << 3) | wire-type 0 (varint)
+    write_pb_varint(value, out);
+}
+
+fn write_pb_bytes_field(field: u32, value: &[u8], out: &mut Vec<u8>) {
+    write_pb_varint(((field as u64) << 3) | 2, out); // wire type 2 = length-delimited
+    write_pb_varint(value.len() as u64, out);
+    out.extend_from_slice(value);
+}
+
+/// UnixFS protobuf payload for a File node: `Type=File, filesize,
+/// blocksizes[]`.
+fn encode_unixfs_file(filesize: u64, blocksizes: &[u64]) -> Vec<u8> {
+    let mut out = Vec::new();
+    write_pb_varint_field(1, 2, &mut out); // Type = File
+    write_pb_varint_field(3, filesize, &mut out); // filesize
+    for &bs in blocksizes {
+        write_pb_varint_field(4, bs, &mut out); // blocksizes (repeated)
+    }
+    out
+}
+
+/// dag-pb PBNode for an internal UnixFS file node. Canonical wire order
+/// pinned by the dag-pb spec: `Links` (tag 2) first, `Data` (tag 1) second.
+fn encode_dagpb_node(children: &[(Vec<u8>, u64, u64)], data: &[u8]) -> Vec<u8> {
+    let mut out = Vec::new();
+    for (cid_wire, _filesize, tsize) in children {
+        let mut link_bytes = Vec::new();
+        write_pb_bytes_field(1, cid_wire, &mut link_bytes); // Hash
+        write_pb_varint_field(3, *tsize, &mut link_bytes); // Tsize
+        write_pb_bytes_field(2, &link_bytes, &mut out); // Links (tag 2) first
+    }
+    write_pb_bytes_field(1, data, &mut out); // Data (tag 1) last
+    out
+}
+
+/// RFC 4648 base32 with the lowercase alphabet and no padding. Used for the
+/// multibase `b` prefix.
+pub(crate) fn base32_lower_no_padding(input: &[u8]) -> String {
+    const ALPHABET: &[u8; 32] = b"abcdefghijklmnopqrstuvwxyz234567";
+    let mut out = String::with_capacity(input.len().div_ceil(5) * 8);
+    let mut buffer: u64 = 0;
+    let mut bits: u32 = 0;
+    for &b in input {
+        buffer = (buffer << 8) | b as u64;
+        bits += 8;
+        while bits >= 5 {
+            bits -= 5;
+            let idx = ((buffer >> bits) & 0x1F) as usize;
+            out.push(ALPHABET[idx] as char);
+        }
+    }
+    if bits > 0 {
+        let idx = ((buffer << (5 - bits)) & 0x1F) as usize;
+        out.push(ALPHABET[idx] as char);
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn midhash256_matches_ts_fixtures() {
+        assert_eq!(
+            compute_midhash256(b"hello world"),
+            "bagacbabaec7v3fu2ygzh3e2sybg3fbzmisry2hbtpmck6vx3yftea6vzq35r4"
+        );
+        assert_eq!(
+            compute_midhash256(b""),
+            "bagacbabaecxvk4hvugaqw6xxrsxuxrykmyhq35i6ik5pshkn4wzdfdpa5a67y"
+        );
+        let mb_zeros = vec![0u8; 1024 * 1024];
+        assert_eq!(
+            compute_midhash256(&mb_zeros),
+            "bagacbabaeawlkt4hn34sxuieoosrnnag3g3wv7gc6alzwfms2nrph2hochrx4"
+        );
+    }
+
+    #[test]
+    fn midhash256_from_sample_matches_full_bytes() {
+        let small: &[u8] = b"hello world";
+        assert_eq!(
+            compute_midhash256_from_sample(small.len() as u64, small),
+            compute_midhash256(small),
+        );
+
+        assert_eq!(
+            compute_midhash256_from_sample(0, b""),
+            compute_midhash256(b""),
+        );
+
+        let one_mb = vec![0u8; SAMPLE_SIZE];
+        assert_eq!(
+            compute_midhash256_from_sample(one_mb.len() as u64, &one_mb),
+            compute_midhash256(&one_mb),
+        );
+
+        let two_mb = vec![0u8; 2 * SAMPLE_SIZE];
+        let start = (two_mb.len() - SAMPLE_SIZE) / 2;
+        let middle = &two_mb[start..start + SAMPLE_SIZE];
+        assert_eq!(
+            compute_midhash256_from_sample(two_mb.len() as u64, middle),
+            compute_midhash256(&two_mb),
+        );
+    }
+
+    #[test]
+    fn midhash256_from_sample_partial_sample_stable() {
+        let sample = vec![0u8; SAMPLE_SIZE];
+        let total_5gb: u64 = 5 * 1024 * 1024 * 1024;
+        let total_10gb: u64 = 10 * 1024 * 1024 * 1024;
+
+        let h1 = compute_midhash256_from_sample(total_5gb, &sample);
+        let h2 = compute_midhash256_from_sample(total_5gb, &sample);
+        assert_eq!(h1, h2, "same (size, sample) must hash identically");
+
+        let h3 = compute_midhash256_from_sample(total_10gb, &sample);
+        assert_ne!(
+            h1, h3,
+            "size prefix must contribute to cid — same sample, different size"
+        );
+    }
+
+    #[test]
+    fn bt_v1_file_cid_wire_layout() {
+        let infohash: [u8; 20] = [
+            0xf9, 0xc8, 0xa7, 0xb6, 0xe5, 0xd4, 0xc3, 0xb2, 0xa1, 0x90, 0x8f, 0x7e, 0x6d, 0x5c,
+            0x4b, 0x3a, 0x29, 0x18, 0x07, 0x06,
+        ];
+
+        let mut want = vec![0x01, 0x81, 0x20, 0x81, 0x20, 0x15];
+        want.extend_from_slice(&infohash);
+        want.push(0x04);
+        assert_eq!(
+            compute_bt_v1_file_cid(&infohash, 4),
+            format!("b{}", base32_lower_no_padding(&want))
+        );
+
+        let mut want0 = vec![0x01, 0x81, 0x20, 0x81, 0x20, 0x15];
+        want0.extend_from_slice(&infohash);
+        want0.push(0x00);
+        assert_eq!(
+            compute_bt_v1_file_cid(&infohash, 0),
+            format!("b{}", base32_lower_no_padding(&want0))
+        );
+
+        let mut want200 = vec![0x01, 0x81, 0x20, 0x81, 0x20, 0x16];
+        want200.extend_from_slice(&infohash);
+        want200.extend_from_slice(&[0xC8, 0x01]);
+        assert_eq!(
+            compute_bt_v1_file_cid(&infohash, 200),
+            format!("b{}", base32_lower_no_padding(&want200))
+        );
+    }
+
+    #[test]
+    fn bt_v2_file_cid_wire_layout() {
+        let infohash: [u8; 32] = [0xAB; 32];
+        let mut want = vec![0x01, 0x82, 0x20, 0x82, 0x20, 0x21];
+        want.extend_from_slice(&infohash);
+        want.push(0x00);
+        assert_eq!(
+            compute_bt_v2_file_cid(&infohash, 0),
+            format!("b{}", base32_lower_no_padding(&want))
+        );
+    }
+
+    #[test]
+    fn bt_file_cid_properties() {
+        let ih1 = [0x11u8; 20];
+        let ih2 = [0x22u8; 32];
+
+        assert_eq!(
+            compute_bt_v1_file_cid(&ih1, 3),
+            compute_bt_v1_file_cid(&ih1, 3)
+        );
+        assert_ne!(
+            compute_bt_v1_file_cid(&ih1, 0),
+            compute_bt_v1_file_cid(&ih1, 1)
+        );
+        assert!(compute_bt_v1_file_cid(&ih1, 0).starts_with('b'));
+        assert!(compute_bt_v2_file_cid(&ih2, 0).starts_with('b'));
+        assert_ne!(
+            compute_bt_v1_file_cid(&[0xAB; 20], 0),
+            compute_bt_v2_file_cid(&[0xAB; 32], 0)
+        );
+    }
+
+    #[test]
+    fn base32_known_vectors() {
+        assert_eq!(base32_lower_no_padding(b""), "");
+        assert_eq!(base32_lower_no_padding(b"f"), "my");
+        assert_eq!(base32_lower_no_padding(b"fo"), "mzxq");
+        assert_eq!(base32_lower_no_padding(b"foo"), "mzxw6");
+        assert_eq!(base32_lower_no_padding(b"foob"), "mzxw6yq");
+        assert_eq!(base32_lower_no_padding(b"foobar"), "mzxw6ytboi");
+    }
+
+    #[test]
+    fn ipfs_cid_single_leaf_matches_kubo() {
+        assert_eq!(
+            compute_ipfs_cid(b"hello world"),
+            "bafkreifzjut3te2nhyekklss27nh3k72ysco7y32koao5eei66wof36n5e"
+        );
+    }
+
+    #[test]
+    fn ipfs_cid_empty_matches_kubo() {
+        assert_eq!(
+            compute_ipfs_cid(b""),
+            "bafkreihdwdcefgh4dqkjv67uzcmw7ojee6xedzdetojuzjevtenxquvyku"
+        );
+    }
+
+    #[test]
+    fn ipfs_cid_multi_leaf_zeros_stable() {
+        let mb_zeros = vec![0u8; 1024 * 1024];
+        assert_eq!(
+            compute_ipfs_cid(&mb_zeros),
+            "bafybeiadh3bekpwtewjvauqeucf7yzqrb3ixsxzltnuwed4pxangtpou6m"
+        );
+    }
+
+    #[test]
+    fn ipfs_blocks_single_leaf() {
+        let payload = b"hello world";
+        let out = compute_ipfs_blocks(payload);
+        assert_eq!(out.root, compute_ipfs_cid(payload));
+        assert_eq!(out.blocks.len(), 1);
+        assert_eq!(out.blocks[0].0, out.root);
+        assert_eq!(out.blocks[0].1.as_ref(), payload);
+    }
+
+    #[test]
+    fn ipfs_blocks_multi_leaf() {
+        let mb_zeros = vec![0u8; 1024 * 1024];
+        let out = compute_ipfs_blocks(&mb_zeros);
+        assert_eq!(
+            out.root,
+            "bafybeiadh3bekpwtewjvauqeucf7yzqrb3ixsxzltnuwed4pxangtpou6m"
+        );
+        assert_eq!(out.blocks.len(), 5);
+        assert_eq!(out.blocks.last().unwrap().0, out.root);
+        for (_cid, block) in &out.blocks[..4] {
+            assert_eq!(block.len(), 256 * 1024);
+            assert!(block.iter().all(|&b| b == 0));
+        }
+    }
+
+    #[test]
+    fn ipfs_blocks_empty() {
+        let out = compute_ipfs_blocks(b"");
+        assert_eq!(
+            out.root,
+            "bafkreihdwdcefgh4dqkjv67uzcmw7ojee6xedzdetojuzjevtenxquvyku"
+        );
+        assert_eq!(out.blocks.len(), 1);
+        assert_eq!(out.blocks[0].1.len(), 0);
+        assert_eq!(out.blocks[0].0, out.root);
+    }
+}

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

@@ -0,0 +1,81 @@
+//! Language-code normalization — ISO 639 alias folding, **not** filename
+//! parsing.
+//!
+//! This lives apart from [`crate::filename_meta`] on purpose: it doesn't read a
+//! release title, it canonicalizes a *language code* (an indexer's declared
+//! language, a query's `languages:` filter, a title-derived stamp) onto the
+//! single ISO 639-2/B alpha-3 vocabulary the rest of the stack compares against.
+//! It is the one piece of the old `filename_meta` module that has no equivalent
+//! in the `metamesh-plugin-filename-parser` plugin — because it isn't filename
+//! parsing — so it stays in Rust while the title-parsing algorithm itself is
+//! owned by the plugin. Callers in the **request hot path** (e.g. the torznab
+//! indexer-language fan-out skip) need it synchronously, with no network hop.
+
+/// Normalize a language code to the ISO 639-2/B alpha-3 vocabulary the stack
+/// emits (`fre`/`ger`/`chi`, **not** the 639-2/T `fra`/`deu`/`zho` nor the
+/// 639-1 two-letter forms), lowercased. This is the space meta-share's
+/// "languages I read" preference uses and that the title-derived language stamps
+/// write, so an operator who declares an indexer's language as `fra`/`de`/`zh`
+/// still lines up with a `fre`/`ger`/`chi` query filter. Unknown /
+/// already-canonical codes pass through lowercased+trimmed unchanged.
+pub fn normalize_lang_code(code: &str) -> String {
+    let c = code.trim().to_ascii_lowercase();
+    let mapped = match c.as_str() {
+        // 639-2/T and 639-1 → 639-2/B for the codes that actually differ.
+        "fra" | "fr" => "fre",
+        "deu" | "de" => "ger",
+        "zho" | "zh" => "chi",
+        "nld" | "nl" => "dut",
+        "ell" | "el" => "gre",
+        "ron" | "ro" => "rum",
+        "slk" | "sk" => "slo",
+        "ces" | "cs" => "cze",
+        // 639-1 → 639-2 for the rest of the language-tag output space (these
+        // share the same /B and /T alpha-3, so only the two-letter form maps).
+        "en" => "eng",
+        "ja" => "jpn",
+        "it" => "ita",
+        "es" => "spa",
+        "pt" => "por",
+        "ko" => "kor",
+        "ru" => "rus",
+        "pl" => "pol",
+        "hu" => "hun",
+        "uk" => "ukr",
+        "bg" => "bul",
+        "sv" => "swe",
+        "no" => "nor",
+        "da" => "dan",
+        "fi" => "fin",
+        "tr" => "tur",
+        "ar" => "ara",
+        "hi" => "hin",
+        "ta" => "tam",
+        "te" => "tel",
+        "ml" => "mal",
+        "vi" => "vie",
+        "th" => "tha",
+        _ => return c,
+    };
+    mapped.to_string()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn normalize_lang_maps_to_639_2b_vocabulary() {
+        // 639-2/T and 639-1 spellings collapse onto the /B forms the stack uses.
+        assert_eq!(normalize_lang_code("fra"), "fre");
+        assert_eq!(normalize_lang_code("FR"), "fre");
+        assert_eq!(normalize_lang_code("deu"), "ger");
+        assert_eq!(normalize_lang_code("zh"), "chi");
+        assert_eq!(normalize_lang_code("nld"), "dut");
+        // Already-canonical / unknown codes pass through trimmed + lowercased.
+        assert_eq!(normalize_lang_code("  ENG "), "eng");
+        assert_eq!(normalize_lang_code("fre"), "fre");
+        assert_eq!(normalize_lang_code("mul"), "mul");
+        assert_eq!(normalize_lang_code("xyz"), "xyz");
+    }
+}

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

@@ -0,0 +1,42 @@
+//! `meta-feeder-sdk` — the source-agnostic foundation shared by the gateway
+//! core and every feeder sidecar binary.
+//!
+//! A *feeder* implements [`FeederPlugin`] (find records, fetch bytes) for one
+//! or more upstreams, and [`serve_feeders`] exposes them over the feeder HTTP
+//! contract. The gateway core consumes that contract via its
+//! `RemoteFeederPlugin` and keeps the libp2p wire, the bitswap blockstore, the
+//! hashing-into-the-blockstore, and the meta-core store-back to itself.
+//!
+//! Deliberately libp2p-free and blockstore-free — see this crate's `Cargo.toml`.
+
+// `redb::Error` is a large enum; the cache wrappers (`cache.rs`) return it
+// directly rather than boxing on every embedded-DB call. Boxing each would be
+// pure noise for a single-process store, so silence the lint crate-wide.
+#![allow(clippy::result_large_err)]
+
+pub mod cache;
+pub mod common;
+pub mod config;
+pub mod enrich;
+pub mod filename_meta;
+pub mod hash;
+pub mod lang;
+pub mod meta_core;
+pub mod plugin;
+pub mod query;
+pub mod query_eval;
+pub mod serve;
+pub mod types;
+
+pub use config::{ConfigField, ConfigSchema, FieldKind};
+pub use enrich::{EnrichTarget, EnrichmentConfig, Enricher};
+pub use meta_core::FeederStore;
+pub use plugin::{
+    upstream_id_field, ConfigError, FeederPlugin, HashKind, HashOutcome, PluginRegistry,
+};
+pub use query::{GatewayQuery, GatewaySearchEvent, GatewayWireError, Negation, RangeFilter};
+pub use serve::{
+    configure_plugins, router, serve_feeders, ComputeRequest, ComputeResponse, HashKindDto,
+    HealthResponse, ManifestResponse, OutcomeDto, PluginManifest, QueryRequest, QueryResponse,
+};
+pub use types::{ByteStream, DiscoveryId, DiscoveryRecord, GatewayError, Hash, PluginHealth};

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

@@ -0,0 +1,293 @@
+//! Lightweight meta-core write/read client for feeders that **self-publish**.
+//!
+//! Post-"smart feeder / dumb gateway" split: a feeder given a `META_CORE_URL`
+//! writes its own base record into meta-core instead of returning it for the
+//! gateway dispatcher to store. This keeps the gateway a pure aggregator +
+//! request router; the feeder owns the full record lifecycle for its upstream.
+//!
+//! The write shape is **byte-for-byte the same** as the gateway's
+//! `meta_core::build_metadata_body` (`crates/meta-gateway/src/meta_core.rs`):
+//! `cid_*` fields collapse to the `cids/<cid> = "true"` key-set, and a
+//! `provenance` JSON blob is stamped. Keeping them identical means a record
+//! written by the feeder is indistinguishable from one the dispatcher wrote —
+//! the meta-share ingester, meta-watch, and ranking all behave the same.
+//!
+//! `PUT /api/metadata/{cid}` is merge/upsert (confirmed: meta-core
+//! `SetMetadataFlat` sets only the fields in the payload, never deletes), so the
+//! feeder's base write and the enrichment plugins' later `PATCH /meta/{cid}`
+//! merges accumulate on the same `/file/{cid}` hash.
+
+use std::collections::BTreeMap;
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use serde::Deserialize;
+use tracing::warn;
+
+use crate::plugin::{HashKind, HashOutcome};
+use crate::types::{DiscoveryRecord, Hash};
+
+/// Provenance recorded on every gateway-acquired record. Mirrors the gateway
+/// dispatcher's shape so feeder-written and dispatcher-written records are
+/// interchangeable.
+#[derive(Clone, Debug)]
+pub struct Provenance {
+    pub upstream: String,
+    pub record_id: String,
+    /// Stable per-deployment id (the gateway peer this feeder backs, from
+    /// `META_GATEWAY_PEER_ID`, else the feeder hostname). Analytics group by it.
+    pub gateway_peer: String,
+    pub acquired_at: u64,
+}
+
+impl Provenance {
+    /// Build a provenance stamp for `upstream`/`record_id`, stamping "now".
+    pub fn now(upstream: &str, record_id: &str, gateway_peer: &str) -> Self {
+        Provenance {
+            upstream: upstream.to_string(),
+            record_id: record_id.to_string(),
+            gateway_peer: gateway_peer.to_string(),
+            acquired_at: SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .map(|d| d.as_secs())
+                .unwrap_or(0),
+        }
+    }
+}
+
+/// JSON-stringified provenance value. Same key set + `"source":"gateway"` as the
+/// gateway dispatcher (`crates/meta-gateway/src/meta_core.rs::provenance_to_json`).
+pub fn provenance_to_json(p: &Provenance) -> String {
+    serde_json::to_string(&serde_json::json!({
+        "source":      "gateway",
+        "upstream":    p.upstream,
+        "recordId":    p.record_id,
+        "gatewayPeer": p.gateway_peer,
+        "acquiredAt":  p.acquired_at,
+    }))
+    .expect("static JSON shape — serialisation is infallible")
+}
+
+/// Build the flat meta-core PUT body from a [`DiscoveryRecord`] + provenance.
+///
+/// Identical transform to the gateway dispatcher: any `cid_*` field becomes a
+/// `cids/<value> = "true"` key-set member (the CID is self-describing — the
+/// algorithm is the multicodec; see METADATA_KEYS.md §2/§14.13), everything else
+/// passes through, and `provenance` is stamped last (overriding any plugin-set
+/// value).
+pub fn build_metadata_body(
+    record: &DiscoveryRecord,
+    provenance: &Provenance,
+) -> BTreeMap<String, String> {
+    let mut body: BTreeMap<String, String> = BTreeMap::new();
+    for (k, v) in &record.fields {
+        if k.starts_with("cid_") && !v.is_empty() {
+            body.insert(format!("cids/{v}"), "true".to_string());
+        } else {
+            body.insert(k.clone(), v.clone());
+        }
+    }
+    body.insert("provenance".to_string(), provenance_to_json(provenance));
+    body
+}
+
+/// `PUT /api/metadata/{cid}` — write (merge/upsert) the record body into
+/// meta-core. `base_url` is the meta-core root (e.g. `http://…-core:9000`).
+pub async fn put_record(
+    http: &reqwest::Client,
+    base_url: &str,
+    cid: &str,
+    body: &BTreeMap<String, String>,
+) -> anyhow::Result<()> {
+    let url = format!("{}/api/metadata/{}", base_url.trim_end_matches('/'), cid);
+    let resp = http.put(&url).json(body).send().await?;
+    let status = resp.status();
+    if !status.is_success() {
+        let text = resp.text().await.unwrap_or_default();
+        anyhow::bail!("meta-core PUT {url} returned {status}: {text}");
+    }
+    Ok(())
+}
+
+/// Shape of meta-core's `GET /meta/{cid}` response (`{ hashId, metadata }`).
+#[derive(Debug, Deserialize)]
+struct MetaGetResponse {
+    #[serde(default)]
+    metadata: Option<BTreeMap<String, String>>,
+}
+
+/// `GET /meta/{cid}` — read the current flat field map for a record, or `None`
+/// when meta-core has no record yet (404). Used to read back a plugin's merged
+/// fields (e.g. filename-parser's `originalTitle`/`movieYear`) before handing
+/// them to the next plugin in the pipeline.
+pub async fn get_record(
+    http: &reqwest::Client,
+    base_url: &str,
+    cid: &str,
+) -> anyhow::Result<Option<BTreeMap<String, String>>> {
+    let url = format!("{}/meta/{}", base_url.trim_end_matches('/'), cid);
+    let resp = http.get(&url).send().await?;
+    if resp.status() == reqwest::StatusCode::NOT_FOUND {
+        return Ok(None);
+    }
+    if !resp.status().is_success() {
+        anyhow::bail!("meta-core GET {url} returned {}", resp.status());
+    }
+    let parsed: MetaGetResponse = resp.json().await?;
+    Ok(parsed.metadata)
+}
+
+/// Self-publishing store for feeders that own their record lifecycle (D4):
+/// the feeder writes its own **metadata-only** records to meta-core and returns
+/// sparse outcomes so the gateway dispatcher skips them; the meta-share ingester
+/// Kamilata-propagates them.
+///
+/// **Byte outcomes are deliberately NOT self-published.** A feeder-side WebDAV
+/// write to the gateway core would be seeded by nobody: the gateway core runs
+/// with `ENABLE_FILE_WATCHER=false` (so meta-core emits no `/api/events/files`
+/// for it, and the ingester's `ipfs_seed` — which seeds only watched files —
+/// never sees it), and a sparse outcome makes the gateway skip its own
+/// blockstore seed. So byte-bearing outcomes (torznab full-file / subtitles)
+/// pass through unchanged and the **gateway** continues to WebDAV-PUT + bitswap-
+/// seed them (it's the libp2p/bitswap host). Result: feeder owns the metadata
+/// records (the Kamilata-propagated bulk), gateway owns byte delivery — a clean
+/// split with no seeding regression. Tribler is metadata-only and uses
+/// [`crate::enrich::Enricher`]; this is torznab's record-owning counterpart.
+#[derive(Clone)]
+pub struct FeederStore {
+    http: reqwest::Client,
+    meta_core_url: String,
+    gateway_peer: String,
+}
+
+impl FeederStore {
+    pub fn new(http: reqwest::Client, meta_core_url: String, gateway_peer: String) -> Self {
+        FeederStore {
+            http,
+            meta_core_url: meta_core_url.trim_end_matches('/').to_string(),
+            gateway_peer,
+        }
+    }
+
+    /// Build from the feeder env. `None` when `META_CORE_URL` is unset (the
+    /// feeder then falls back to the legacy "return outcomes, gateway stores"
+    /// path). `gateway_peer` ← `META_GATEWAY_PEER_ID` / `HOSTNAME`.
+    pub fn from_env() -> Option<Self> {
+        let meta_core_url = env_nonempty("META_CORE_URL")?;
+        let gateway_peer = env_nonempty("META_GATEWAY_PEER_ID")
+            .or_else(|| env_nonempty("HOSTNAME"))
+            .unwrap_or_else(|| "gateway-feeder".to_string());
+        let http = crate::common::build_http_client(30, "meta-feeder-store", None);
+        Some(Self::new(http, meta_core_url, gateway_peer))
+    }
+
+    pub fn meta_core_url(&self) -> &str {
+        &self.meta_core_url
+    }
+
+    /// 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
+    /// returned unchanged (the gateway stores + seeds the bytes; see the struct
+    /// docs). On a write failure the original outcome is returned so the gateway
+    /// stores it (graceful degradation, never lost).
+    pub async fn publish(
+        &self,
+        upstream_id: &str,
+        record_id: &str,
+        outcome: HashOutcome,
+    ) -> HashOutcome {
+        // Only self-publish metadata-only outcomes (bytes None, record Some).
+        if outcome.bytes.is_some() || outcome.record.is_none() {
+            return outcome;
+        }
+        let cid = outcome.hash.0.clone();
+        let record = outcome.record.as_ref().expect("record is Some by the guard");
+        match self
+            .store_metadata_only(upstream_id, record_id, &cid, record)
+            .await
+        {
+            Ok(()) => sparse(outcome.hash, outcome.hash_kind),
+            Err(e) => {
+                warn!(target: "meta-feeder::store", cid = %cid, error = %e,
+                      "metadata self-publish failed; falling back to gateway store");
+                outcome
+            }
+        }
+    }
+
+    /// Metadata-only self-publish: `PUT /api/metadata/{cid}` with provenance.
+    async fn store_metadata_only(
+        &self,
+        upstream_id: &str,
+        record_id: &str,
+        cid: &str,
+        record: &DiscoveryRecord,
+    ) -> anyhow::Result<()> {
+        let prov = Provenance::now(upstream_id, record_id, &self.gateway_peer);
+        let body = build_metadata_body(record, &prov);
+        put_record(&self.http, &self.meta_core_url, cid, &body).await
+    }
+}
+
+fn sparse(hash: Hash, hash_kind: HashKind) -> HashOutcome {
+    HashOutcome {
+        hash,
+        hash_kind,
+        bytes: None,
+        record: None,
+        file_extension: None,
+    }
+}
+
+fn env_nonempty(key: &str) -> Option<String> {
+    std::env::var(key)
+        .ok()
+        .map(|s| s.trim().to_string())
+        .filter(|s| !s.is_empty())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn rec() -> DiscoveryRecord {
+        DiscoveryRecord {
+            upstream_id: "tribler".into(),
+            record_id: "abc".into(),
+            fields: BTreeMap::from_iter([
+                ("title".into(), "Big Buck Bunny".into()),
+                ("cid_btih_v1_file".into(), "bafkPERFILE".into()),
+            ]),
+        }
+    }
+
+    #[test]
+    fn provenance_shape_matches_gateway() {
+        let p = Provenance {
+            upstream: "tribler".into(),
+            record_id: "abc".into(),
+            gateway_peer: "metagateway-tribler".into(),
+            acquired_at: 1_747_504_000,
+        };
+        let v: serde_json::Value = serde_json::from_str(&provenance_to_json(&p)).unwrap();
+        assert_eq!(v["source"], "gateway");
+        assert_eq!(v["upstream"], "tribler");
+        assert_eq!(v["recordId"], "abc");
+        assert_eq!(v["gatewayPeer"], "metagateway-tribler");
+        assert_eq!(v["acquiredAt"], 1_747_504_000_u64);
+    }
+
+    #[test]
+    fn body_collapses_cid_fields_and_stamps_provenance() {
+        let p = Provenance::now("tribler", "abc", "peer");
+        let body = build_metadata_body(&rec(), &p);
+        assert_eq!(body.get("title").map(String::as_str), Some("Big Buck Bunny"));
+        // cid_* collapsed to cids/<value>=true; no cid_* survives.
+        assert!(body.keys().all(|k| !k.starts_with("cid_")));
+        assert_eq!(
+            body.get("cids/bafkPERFILE").map(String::as_str),
+            Some("true")
+        );
+        assert!(body.contains_key("provenance"));
+    }
+}

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

@@ -0,0 +1,203 @@
+//! The `FeederPlugin` trait + supporting types.
+//!
+//! A feeder plugin bridges one external upstream (`gutenberg`, `arxiv`,
+//! `torznab`, …) into the MetaMesh network. It **finds** records and
+//! **fetches** bytes; it does NOT hash-into-a-blockstore, store back into
+//! meta-core, or speak libp2p — those stay in the gateway core. This is why
+//! `FeederPlugin` has no `set_blockstore` / `set_tmdb_budget` hooks (the
+//! gateway's `GatewayPlugin` carried them): preview-seeding moves to the core,
+//! and a feeder that needs a TMDB budget owns it internally.
+//!
+//! The `serve_feeders` harness ([`crate::serve`]) wraps any set of
+//! `FeederPlugin`s in the feeder HTTP contract; the gateway core's
+//! `RemoteFeederPlugin` is the matching HTTP client.
+
+use std::collections::HashMap;
+use std::path::Path;
+
+use async_trait::async_trait;
+use futures::stream::BoxStream;
+
+// Re-exported so plugins can `use meta_feeder_sdk::plugin::GatewayQuery` (the
+// gateway's original `plugin.rs` re-exported it the same way).
+pub use crate::query::GatewayQuery;
+use crate::query::GatewaySearchEvent;
+use crate::types::{ByteStream, DiscoveryRecord, GatewayError, Hash, PluginHealth};
+
+/// One element of what [`FeederPlugin::compute_outcomes`] returns.
+///
+/// The `hash` is the value shipped to the gateway core. The other fields are
+/// signals the core's auto-store uses to persist the content into meta-core.
+/// The core's three-branch auto-store routes on `(bytes, record)`:
+/// - `(Some, Some)` — full store: WebDAV PUT + metadata PUT with `filePath`.
+/// - `(None, Some)` — metadata-only PUT (no local bytes; e.g. torznab samples
+///   the middle 1 MiB of a multi-GiB video and surfaces only the record).
+/// - `(_, None)` — skip. Cache hits land here.
+///
+/// `file_extension` MUST be `None` when `bytes` is `None`.
+#[derive(Debug)]
+pub struct HashOutcome {
+    pub hash: Hash,
+    /// Which hash family this CID belongs to. Drives the core's routing —
+    /// the bitswap blockstore seed only fires for `Sha2_256`.
+    pub hash_kind: HashKind,
+    pub bytes: Option<bytes::Bytes>,
+    pub record: Option<DiscoveryRecord>,
+    /// File extension to append when naming the WebDAV-side blob (e.g.
+    /// `"epub"`). No leading dot. MUST be `None` when `bytes` is `None`.
+    pub file_extension: Option<String>,
+}
+
+/// Hash family discriminator on [`HashOutcome`].
+///
+/// - `Midhash256` — size-prefix-plus-middle-1MB-sample, custom multicodec
+///   `0x1000`. Fast, MetaMesh-internal, not retrievable via public IPFS.
+/// - `Sha2_256` — standard IPFS CIDv1 over full bytes. Retrievable via bitswap
+///   once the bytes are in the core's blockstore. What every shipping plugin
+///   today emits.
+/// - `BtV1File` — a single file inside a BitTorrent v1 torrent, custom
+///   multicodec `0x1001`. Locator, not content hash: opaque to bitswap.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum HashKind {
+    /// Reserved: no feeder currently emits midhash256 outcomes, but the family
+    /// is real across meta-core / meta-share, so the variant stays.
+    #[allow(dead_code)]
+    Midhash256,
+    Sha2_256,
+    BtV1File,
+}
+
+/// Startup-only configuration error returned by [`FeederPlugin::configure`].
+#[derive(Debug, thiserror::Error)]
+pub enum ConfigError {
+    #[error("plugin {plugin}: required config missing: {what}")]
+    MissingConfig {
+        plugin: &'static str,
+        /// Human-readable description of the missing setting and where to
+        /// supply it. Used by the harness's soft-skip warning.
+        what: &'static str,
+    },
+
+    #[error("plugin {plugin}: cache dir setup failed: {source}")]
+    CacheSetup {
+        plugin: &'static str,
+        #[source]
+        source: std::io::Error,
+    },
+
+    #[error("plugin {plugin}: {source}")]
+    Other {
+        plugin: &'static str,
+        #[source]
+        source: anyhow::Error,
+    },
+}
+
+/// Canonical field name for an upstream's stable per-record identifier.
+/// Every plugin's `DiscoveryRecord.fields` MUST include this key
+/// (`gutenbergid:11`, `tmdbid:31910`, `arxivid:2106.07447`). The rule is
+/// `format!("{upstream_id}id")` — lowercase, no separator.
+pub fn upstream_id_field(upstream_id: &str) -> String {
+    format!("{upstream_id}id")
+}
+
+/// 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.
+///
+/// **Field-naming convention.** Every record a plugin returns MUST include the
+/// canonical `<upstream_id>id` field — see [`upstream_id_field`].
+#[async_trait]
+pub trait FeederPlugin: Send + Sync + 'static {
+    fn upstream_id(&self) -> &'static str;
+
+    fn configure(&mut self, cache_dir: &Path) -> Result<(), ConfigError>;
+
+    async fn handle_query(
+        &self,
+        query: &GatewayQuery,
+        max_results: usize,
+    ) -> Result<Vec<DiscoveryRecord>, GatewayError>;
+
+    /// Streaming variant of [`handle_query`]. The default impl collects
+    /// [`handle_query`] and replays it as `Base*` + `Done` — correct for every
+    /// plugin with no incremental enrichment (gutenberg, arxiv, …). torznab
+    /// overrides it to decouple base records from rate-limited TMDB enrichment.
+    async fn handle_query_stream(
+        &self,
+        query: &GatewayQuery,
+        max_results: usize,
+    ) -> Result<BoxStream<'static, GatewaySearchEvent>, GatewayError> {
+        let records = self.handle_query(query, max_results).await?;
+        let events = records
+            .into_iter()
+            .map(GatewaySearchEvent::Base)
+            .chain(std::iter::once(GatewaySearchEvent::Done));
+        Ok(Box::pin(futures::stream::iter(events)))
+    }
+
+    /// Resolve `record_id` into one or more content-addressed outcomes.
+    ///
+    /// - `Err(_)` — the upstream record itself is unresolvable.
+    /// - `Ok(vec![])` — resolved but yielded zero outcomes.
+    /// - `Ok(partial_vec)` — bundle-success: per-sibling failures are
+    ///   best-effort (log + drop), surface whatever succeeded.
+    async fn compute_outcomes(&self, record_id: &str) -> Result<Vec<HashOutcome>, GatewayError>;
+
+    async fn handle_fetch(&self, _record_id: &str) -> Result<Option<ByteStream>, GatewayError> {
+        Ok(None)
+    }
+
+    fn health(&self) -> PluginHealth {
+        PluginHealth::Ok
+    }
+
+    fn served_file_types(&self) -> &'static [&'static str] {
+        &[]
+    }
+
+    fn served_content_kinds(&self) -> &'static [&'static str] {
+        &[]
+    }
+
+    /// Self-describing config schema. The feeder SDK serves this at
+    /// `GET /config/schema` and renders a generic form from it — the gateway and
+    /// UI carry no per-plugin field knowledge. Default: no configuration.
+    fn config_schema(&self) -> crate::config::ConfigSchema {
+        crate::config::ConfigSchema::default()
+    }
+
+    /// The plugin's current *effective* config as JSON, **unredacted** (secrets
+    /// included). The SDK redacts it before serving `GET /config/values` and uses
+    /// it as the merge base on the first save (before any `config.json` exists).
+    /// Default: empty object. Implementors should mirror the keys in
+    /// [`config_schema`](FeederPlugin::config_schema).
+    fn config_values(&self) -> serde_json::Value {
+        serde_json::json!({})
+    }
+
+    async fn get_blob(&self, _cid: &str) -> Option<Vec<u8>> {
+        None
+    }
+}
+
+/// Static plugin registry. Built once at startup; read-only thereafter.
+/// Keyed by `upstream_id`.
+pub type PluginRegistry = HashMap<&'static str, Box<dyn FeederPlugin>>;
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn config_error_messages_include_plugin_name() {
+        let e = ConfigError::MissingConfig {
+            plugin: "scihub",
+            what:
+                "mirrors (set them in the dashboard or gateway-config.json plugins.scihub.mirrors)",
+        };
+        let msg = e.to_string();
+        assert!(msg.contains("scihub"));
+        assert!(msg.contains("mirrors"));
+    }
+}

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

@@ -0,0 +1,147 @@
+//! Structured query envelope + streaming-search event + wire-error types,
+//! shared by feeders (their HTTP contract) and the gateway core (its libp2p
+//! wire). These are pure serde data types — no libp2p, no transport.
+//!
+//! On the core side, `protocol.rs` re-uses these exact structs as the bincode
+//! wire shape; the `wire_parity_*` tests there pin the byte encoding. Keep
+//! field order / variant order / derive shape stable.
+
+use std::collections::BTreeMap;
+
+use serde::{Deserialize, Serialize};
+
+use crate::types::{DiscoveryRecord, GatewayError};
+
+/// Structured query envelope. Gateway plugins do best-effort pushdown to
+/// upstream APIs with native query languages (`torznab cat=movie`,
+/// `arxiv id_list=`).
+///
+/// **Authority rule:** `filters` / `ranges` / `negations` are authoritative;
+/// `raw_text` is informational. A plugin that consumes both will double-filter
+/// and under-match — pick one source per filter; structured wins.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct GatewayQuery {
+    /// Original user text (e.g. `"naruto contentKind:movie imdbid:tt0903747"`).
+    /// Verbatim — never mutated by the conversion.
+    pub raw_text: String,
+
+    /// Bare-word leaves only, space-joined. The plugin's primary `q=` input
+    /// for upstream search APIs that take a single search box.
+    pub free_text: String,
+
+    /// Structured filters keyed by field name (raw, unstemmed values).
+    /// Multiple values per key = OR within a key; AND across keys.
+    pub filters: BTreeMap<String, Vec<String>>,
+
+    /// Numeric range filters (e.g. `movieYear:2020..2024`). Both bounds
+    /// optional → unbounded on that side.
+    pub ranges: Vec<RangeFilter>,
+
+    /// Negation filters (e.g. `NOT genres:horror`, bare `NOT word`).
+    pub negations: Vec<Negation>,
+}
+
+/// Numeric range filter inside [`GatewayQuery::ranges`]. Both bounds are
+/// `Option<i64>` so open-ended ranges drop one bound to `None`.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct RangeFilter {
+    pub field: String,
+    pub lo: Option<i64>,
+    pub hi: Option<i64>,
+}
+
+/// Negation entry inside [`GatewayQuery::negations`]. `field` is `Some(name)`
+/// for typed negations (`NOT genres:horror`), `None` for bare-word negation.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct Negation {
+    pub field: Option<String>,
+    pub value: String,
+}
+
+impl GatewayQuery {
+    /// Bare-text input to hand to a free-form upstream search API. Falls back
+    /// to `"*"` when the conversion produced no free-text leaves so upstreams
+    /// that 4xx on empty search return their top-ranked results instead.
+    pub fn free_text_or_star(&self) -> &str {
+        if self.free_text.trim().is_empty() {
+            "*"
+        } else {
+            self.free_text.as_str()
+        }
+    }
+
+    /// Construct a free-text-only `GatewayQuery` for in-process use (tests,
+    /// debug routes). `raw_text` == `free_text`; structured fields stay empty.
+    pub fn from_free_text(text: impl Into<String>) -> Self {
+        let text = text.into();
+        Self {
+            raw_text: text.clone(),
+            free_text: text,
+            filters: BTreeMap::new(),
+            ranges: Vec::new(),
+            negations: Vec::new(),
+        }
+    }
+}
+
+/// One frame in the streaming Search response.
+///
+/// A plugin emits `Base` records as soon as they're discovered, then
+/// best-effort `EnrichPatch`/`Drop` events as enrichment lands, then a
+/// terminal `Done`. The default `handle_query_stream` collects `handle_query`
+/// and replays it as `Base*` + `Done`.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub enum GatewaySearchEvent {
+    /// An un-enriched discovery record.
+    Base(DiscoveryRecord),
+    /// Incremental enrichment for a previously-sent `Base`, keyed by
+    /// `record_id`. `set` fields are inserted/overwritten; `remove` deleted.
+    EnrichPatch {
+        record_id: String,
+        set: BTreeMap<String, String>,
+        remove: Vec<String>,
+    },
+    /// Retract a previously-emitted `Base` record entirely.
+    Drop { record_id: String },
+    /// Terminal success — no more frames follow.
+    Done,
+    /// Terminal error — no more frames follow.
+    Error(GatewayWireError),
+}
+
+/// Wire-stable error shape. Distinct from in-process [`GatewayError`] because
+/// adding a variant here is a wire-format change.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize, thiserror::Error)]
+pub enum GatewayWireError {
+    #[error("not found")]
+    NotFound,
+    #[error("rate limited; retry in {retry_after_s}s")]
+    RateLimited { retry_after_s: u32 },
+    #[error("upstream transient: {0}")]
+    UpstreamTransient(String),
+    #[error("upstream permanent: {0}")]
+    UpstreamPermanent(String),
+    /// The protocol was registered but no plugin claims this `upstream_id`.
+    #[error("plugin not loaded")]
+    PluginNotLoaded,
+    /// Internal codec / decode failure on the *receiver* side.
+    #[error("internal: {0}")]
+    Internal(String),
+}
+
+impl From<GatewayError> for GatewayWireError {
+    fn from(e: GatewayError) -> Self {
+        match e {
+            GatewayError::Transient(s) => GatewayWireError::UpstreamTransient(s),
+            GatewayError::Permanent(s) => GatewayWireError::UpstreamPermanent(s),
+            GatewayError::RateLimited { retry_after_s } => {
+                GatewayWireError::RateLimited { retry_after_s }
+            }
+            GatewayError::NotFound => GatewayWireError::NotFound,
+            // Internal errors are never leaked; map to a transient.
+            GatewayError::Internal(_) => {
+                GatewayWireError::UpstreamTransient("internal".to_string())
+            }
+        }
+    }
+}

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

@@ -0,0 +1,456 @@
+//! Structured-filter evaluation for gateway records.
+//!
+//! Two pure helpers consuming the [`crate::query::GatewayQuery`] struct:
+//!
+//! - [`record_matches`] — does a single record's `fields` map satisfy the
+//!   query's `filters` + `ranges` + typed `negations`?
+//! - [`query_accepts_plugin`] — could a plugin advertising the given
+//!   `served_file_types` / `served_content_kinds` possibly produce a record
+//!   that satisfies the query? Used by plugins as a cheap early-return.
+//!
+//! The dispatcher applies `record_matches` to every plugin's return value, and
+//! meta-share applies the same algorithm consumer-side to defend against stale
+//! gateway peers — both tiers must agree on outcomes for any (record, query)
+//! pair. See the original gateway crate docs for the full semantics.
+
+use std::collections::BTreeMap;
+
+use crate::query::GatewayQuery;
+
+/// True iff a record's `fields` satisfies every filter, range, and typed
+/// negation in `query`.
+pub fn record_matches(fields: &BTreeMap<String, String>, query: &GatewayQuery) -> bool {
+    for (key, allowed) in &query.filters {
+        if key == "languages" {
+            if !languages_filter_matches(fields, allowed) {
+                return false;
+            }
+            continue;
+        }
+        let Some(field_value) = fields.get(key) else {
+            return false;
+        };
+        if !any_csv_value_matches(field_value, allowed) {
+            return false;
+        }
+    }
+
+    for range in &query.ranges {
+        let Some(field_value) = fields.get(&range.field) else {
+            return false;
+        };
+        let Ok(n) = field_value.parse::<i64>() else {
+            return false;
+        };
+        if let Some(lo) = range.lo {
+            if n < lo {
+                return false;
+            }
+        }
+        if let Some(hi) = range.hi {
+            if n > hi {
+                return false;
+            }
+        }
+    }
+
+    for negation in &query.negations {
+        let Some(name) = negation.field.as_deref() else {
+            // Bare-word negation — not enforced here.
+            continue;
+        };
+        let Some(field_value) = fields.get(name) else {
+            // Missing field ⇒ negation passes (nothing to exclude).
+            continue;
+        };
+        if any_csv_value_matches(field_value, std::slice::from_ref(&negation.value)) {
+            return false;
+        }
+    }
+
+    true
+}
+
+/// True iff a plugin advertising the given `served_file_types` /
+/// `served_content_kinds` could possibly produce a record satisfying
+/// `query.filters["fileType" | "contentKind"]`.
+pub fn query_accepts_plugin(
+    query: &GatewayQuery,
+    served_file_types: &[&str],
+    served_content_kinds: &[&str],
+) -> bool {
+    accepts_axis(query.filters.get("fileType"), served_file_types)
+        && accepts_axis(query.filters.get("contentKind"), served_content_kinds)
+}
+
+fn accepts_axis(requested: Option<&Vec<String>>, served: &[&str]) -> bool {
+    let Some(requested) = requested else {
+        return true;
+    };
+    if requested.is_empty() {
+        return true;
+    }
+    // Wildcard: a plugin that advertises `"*"` on an axis serves every value on
+    // that axis (e.g. tribler is a free-text overlay search with no inherent
+    // type limitation). It accepts any requested fileType/contentKind.
+    if served.contains(&"*") {
+        return true;
+    }
+    requested
+        .iter()
+        .any(|v| served.iter().any(|s| s.eq_ignore_ascii_case(v)))
+}
+
+/// Resolve a `languages` filter against the key-set storage form
+/// (`languages/<lang3> = "true"`). **Fails open** on anything we can't
+/// confidently exclude. A record matches when:
+/// 1. it carries a requested code as a truthy key-set member (OR across codes);
+/// 2. a legacy flat `languages` csv field contains a requested code;
+/// 3. it carries `mul` (multiple — may contain a requested one) or `und`
+///    (undetermined); or
+/// 4. it carries no language information at all.
+///
+/// Only a record with a *concrete*, non-matching language (e.g. `jpn`-only
+/// under an `eng` filter) is dropped. Kept byte-identical in logic with
+/// meta-search's copy (and the client's `passesLangFilter`) — language narrows
+/// the result set without ever hiding unknown/undetermined/multi releases.
+fn languages_filter_matches(fields: &BTreeMap<String, String>, allowed: &[String]) -> bool {
+    for code in allowed {
+        let key = format!("languages/{}", code.trim().to_ascii_lowercase());
+        if fields.get(&key).is_some_and(|v| is_truthy_member(v)) {
+            return true;
+        }
+    }
+    if let Some(flat) = fields.get("languages") {
+        if any_csv_value_matches(flat, allowed) {
+            return true;
+        }
+    }
+    // Fail open: `mul`/`und` are explicit "could be anything / unknown" markers,
+    // and a record with no concrete language at all is equally unknown. Keep all
+    // three; only a specific, non-matching language reaches the `false` below.
+    if fields.get("languages/mul").is_some_and(|v| is_truthy_member(v))
+        || fields.get("languages/und").is_some_and(|v| is_truthy_member(v))
+    {
+        return true;
+    }
+    !record_has_concrete_language(fields)
+}
+
+/// True iff the record states a *concrete* language: a truthy `languages/<code>`
+/// key-set member other than the `und`/`mul` non-codes, or a non-empty legacy
+/// flat `languages` field. Distinguishes "a specific language that didn't match
+/// ⇒ drop" from "no language info ⇒ keep" for [`languages_filter_matches`].
+fn record_has_concrete_language(fields: &BTreeMap<String, String>) -> bool {
+    if fields.get("languages").is_some_and(|v| !v.trim().is_empty()) {
+        return true;
+    }
+    fields.iter().any(|(k, v)| {
+        is_truthy_member(v)
+            && k.starts_with("languages/")
+            && k != "languages/und"
+            && k != "languages/mul"
+    })
+}
+
+/// Key-set members are written as the literal `"true"`; tolerate a couple of
+/// obvious truthy spellings and treat anything else as absent.
+fn is_truthy_member(v: &str) -> bool {
+    matches!(v.trim().to_ascii_lowercase().as_str(), "true" | "1" | "yes")
+}
+
+/// True iff `field_value` (treated as a csv set when it contains commas)
+/// contains at least one case-insensitive match for any item in `allowed`.
+fn any_csv_value_matches(field_value: &str, allowed: &[String]) -> bool {
+    for piece in field_value.split(',') {
+        let piece = piece.trim();
+        if piece.is_empty() {
+            continue;
+        }
+        if allowed.iter().any(|v| v.eq_ignore_ascii_case(piece)) {
+            return true;
+        }
+    }
+    false
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::query::{Negation, RangeFilter};
+
+    fn empty_query() -> GatewayQuery {
+        GatewayQuery {
+            raw_text: String::new(),
+            free_text: String::new(),
+            filters: BTreeMap::new(),
+            ranges: Vec::new(),
+            negations: Vec::new(),
+        }
+    }
+
+    fn with_filter(field: &str, values: &[&str]) -> GatewayQuery {
+        let mut q = empty_query();
+        q.filters.insert(
+            field.to_string(),
+            values.iter().map(|s| s.to_string()).collect(),
+        );
+        q
+    }
+
+    fn fields(pairs: &[(&str, &str)]) -> BTreeMap<String, String> {
+        pairs
+            .iter()
+            .map(|(k, v)| (k.to_string(), v.to_string()))
+            .collect()
+    }
+
+    #[test]
+    fn filter_rejects_mismatched_filetype() {
+        let q = with_filter("fileType", &["video"]);
+        assert!(!record_matches(&fields(&[("fileType", "image")]), &q));
+    }
+
+    #[test]
+    fn filter_accepts_matching_filetype() {
+        let q = with_filter("fileType", &["video"]);
+        assert!(record_matches(&fields(&[("fileType", "video")]), &q));
+    }
+
+    #[test]
+    fn filter_rejects_record_missing_field() {
+        let q = with_filter("fileType", &["video"]);
+        assert!(!record_matches(&fields(&[]), &q));
+    }
+
+    #[test]
+    fn filter_is_case_insensitive() {
+        let q = with_filter("fileType", &["VIDEO"]);
+        assert!(record_matches(&fields(&[("fileType", "video")]), &q));
+        let q = with_filter("fileType", &["video"]);
+        assert!(record_matches(&fields(&[("fileType", "VIDEO")]), &q));
+    }
+
+    #[test]
+    fn filter_ors_within_key() {
+        let q = with_filter("fileType", &["video", "audio"]);
+        assert!(record_matches(&fields(&[("fileType", "audio")]), &q));
+        assert!(record_matches(&fields(&[("fileType", "video")]), &q));
+        assert!(!record_matches(&fields(&[("fileType", "image")]), &q));
+    }
+
+    #[test]
+    fn filter_ands_across_keys() {
+        let mut q = with_filter("fileType", &["video"]);
+        q.filters.insert("genres".into(), vec!["Action".into()]);
+        assert!(!record_matches(
+            &fields(&[("fileType", "video"), ("genres", "Drama")]),
+            &q
+        ));
+        assert!(record_matches(
+            &fields(&[("fileType", "video"), ("genres", "Action")]),
+            &q
+        ));
+    }
+
+    #[test]
+    fn csv_field_matches_one_of_its_values() {
+        let q = with_filter("genres", &["Anime"]);
+        assert!(record_matches(
+            &fields(&[("genres", "Animation,Anime,Drama")]),
+            &q
+        ));
+    }
+
+    #[test]
+    fn csv_field_with_whitespace_is_trimmed() {
+        let q = with_filter("genres", &["Anime"]);
+        assert!(record_matches(
+            &fields(&[("genres", "Animation, Anime , Drama")]),
+            &q
+        ));
+    }
+
+    #[test]
+    fn languages_keyset_member_matches() {
+        let q = with_filter("languages", &["eng"]);
+        assert!(record_matches(&fields(&[("languages/eng", "true")]), &q));
+    }
+
+    #[test]
+    fn languages_filter_is_or_across_codes() {
+        let q = with_filter("languages", &["eng", "fre"]);
+        assert!(record_matches(&fields(&[("languages/eng", "true")]), &q));
+        assert!(record_matches(&fields(&[("languages/fre", "true")]), &q));
+        assert!(!record_matches(&fields(&[("languages/jpn", "true")]), &q));
+    }
+
+    #[test]
+    fn languages_filter_fails_open_on_unknown() {
+        let q = with_filter("languages", &["eng"]);
+        // No language info at all ⇒ kept (unknown, not excluded).
+        assert!(record_matches(&fields(&[("fileType", "video")]), &q));
+        // `mul` (multi) may contain the requested language ⇒ kept.
+        assert!(record_matches(&fields(&[("languages/mul", "true")]), &q));
+        // `und` (undetermined) ⇒ kept.
+        assert!(record_matches(&fields(&[("languages/und", "true")]), &q));
+        // A specific language alongside `mul` is still kept (multi).
+        assert!(record_matches(
+            &fields(&[("languages/jpn", "true"), ("languages/mul", "true")]),
+            &q
+        ));
+    }
+
+    #[test]
+    fn languages_concrete_mismatch_still_drops() {
+        let q = with_filter("languages", &["eng"]);
+        // A concrete, non-matching language ⇒ dropped (language narrows).
+        assert!(!record_matches(&fields(&[("languages/jpn", "true")]), &q));
+        // A tombstoned/false member is treated as absent ⇒ unknown ⇒ kept.
+        assert!(record_matches(&fields(&[("languages/eng", "false")]), &q));
+    }
+
+    #[test]
+    fn languages_legacy_flat_csv_fallback() {
+        let q = with_filter("languages", &["eng"]);
+        assert!(record_matches(&fields(&[("languages", "jpn,eng")]), &q));
+        assert!(!record_matches(&fields(&[("languages", "jpn,ger")]), &q));
+    }
+
+    #[test]
+    fn range_in_bounds() {
+        let mut q = empty_query();
+        q.ranges.push(RangeFilter {
+            field: "movieYear".into(),
+            lo: Some(2010),
+            hi: Some(2020),
+        });
+        assert!(record_matches(&fields(&[("movieYear", "2015")]), &q));
+    }
+
+    #[test]
+    fn range_out_of_bounds_rejected() {
+        let mut q = empty_query();
+        q.ranges.push(RangeFilter {
+            field: "movieYear".into(),
+            lo: Some(2010),
+            hi: Some(2020),
+        });
+        assert!(!record_matches(&fields(&[("movieYear", "2025")]), &q));
+        assert!(!record_matches(&fields(&[("movieYear", "2005")]), &q));
+    }
+
+    #[test]
+    fn range_open_upper_bound() {
+        let mut q = empty_query();
+        q.ranges.push(RangeFilter {
+            field: "movieYear".into(),
+            lo: Some(2010),
+            hi: None,
+        });
+        assert!(record_matches(&fields(&[("movieYear", "2025")]), &q));
+    }
+
+    #[test]
+    fn range_unparseable_value_rejected() {
+        let mut q = empty_query();
+        q.ranges.push(RangeFilter {
+            field: "movieYear".into(),
+            lo: Some(2010),
+            hi: None,
+        });
+        assert!(!record_matches(&fields(&[("movieYear", "circa 2015")]), &q));
+    }
+
+    #[test]
+    fn typed_negation_excludes_match() {
+        let mut q = empty_query();
+        q.negations.push(Negation {
+            field: Some("fileType".into()),
+            value: "image".into(),
+        });
+        assert!(!record_matches(&fields(&[("fileType", "image")]), &q));
+        assert!(record_matches(&fields(&[("fileType", "video")]), &q));
+    }
+
+    #[test]
+    fn typed_negation_passes_when_field_missing() {
+        let mut q = empty_query();
+        q.negations.push(Negation {
+            field: Some("fileType".into()),
+            value: "image".into(),
+        });
+        assert!(record_matches(&fields(&[]), &q));
+    }
+
+    #[test]
+    fn bare_word_negation_is_not_enforced() {
+        let mut q = empty_query();
+        q.negations.push(Negation {
+            field: None,
+            value: "alice".into(),
+        });
+        assert!(record_matches(
+            &fields(&[("title", "alice in wonderland")]),
+            &q
+        ));
+    }
+
+    #[test]
+    fn empty_query_matches_anything() {
+        assert!(record_matches(&fields(&[]), &empty_query()));
+        assert!(record_matches(
+            &fields(&[("fileType", "video"), ("title", "x")]),
+            &empty_query()
+        ));
+    }
+
+    #[test]
+    fn accepts_returns_true_when_no_type_filter() {
+        let q = empty_query();
+        assert!(query_accepts_plugin(&q, &["video"], &["movie"]));
+        assert!(query_accepts_plugin(&q, &[], &[]));
+    }
+
+    #[test]
+    fn accepts_rejects_disjoint_filetype() {
+        let q = with_filter("fileType", &["video"]);
+        assert!(!query_accepts_plugin(&q, &["image"], &["gif"]));
+        assert!(query_accepts_plugin(&q, &["video"], &[]));
+    }
+
+    #[test]
+    fn accepts_rejects_disjoint_contentkind() {
+        let q = with_filter("contentKind", &["movie"]);
+        assert!(!query_accepts_plugin(
+            &q,
+            &["video", "image"],
+            &["paper", "book"]
+        ));
+        assert!(query_accepts_plugin(&q, &["video"], &["movie", "episode"]));
+    }
+
+    #[test]
+    fn accepts_requires_both_axes_when_both_filtered() {
+        let mut q = with_filter("fileType", &["video"]);
+        q.filters.insert("contentKind".into(), vec!["movie".into()]);
+        assert!(query_accepts_plugin(&q, &["video"], &["movie"]));
+        assert!(!query_accepts_plugin(&q, &["video"], &["paper"]));
+        assert!(!query_accepts_plugin(&q, &["image"], &["movie"]));
+    }
+
+    #[test]
+    fn accepts_wildcard_matches_any_requested_type() {
+        // A plugin advertising "*" (e.g. tribler) accepts any fileType /
+        // contentKind filter on the wildcarded axis.
+        let q = with_filter("fileType", &["archive"]);
+        assert!(query_accepts_plugin(&q, &["*"], &[]));
+        let q = with_filter("contentKind", &["audiobook"]);
+        assert!(query_accepts_plugin(&q, &["*"], &["*"]));
+        // Wildcard on one axis only: the non-wildcard axis still gates.
+        let mut q = with_filter("fileType", &["audio"]);
+        q.filters.insert("contentKind".into(), vec!["podcast".into()]);
+        assert!(query_accepts_plugin(&q, &["*"], &["*"]));
+        assert!(!query_accepts_plugin(&q, &["*"], &["movie"]));
+    }
+}

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

@@ -0,0 +1,634 @@
+//! `serve_feeders` — the axum harness that exposes a set of [`FeederPlugin`]s
+//! over the **feeder HTTP contract**. Every feeder sidecar binary is just
+//! `serve_feeders(vec![MyPlugin::new()], state_dir, listen)`.
+//!
+//! # Contract (JSON over HTTP, mirrors meta-sort's container-plugin shape)
+//!
+//! - `GET  /manifest`                  → [`ManifestResponse`]
+//! - `GET  /health`                    → [`HealthResponse`]
+//! - `POST /query`                     → [`QueryRequest`] → [`QueryResponse`]
+//! - `POST /compute`                   → [`ComputeRequest`] → [`ComputeResponse`]
+//! - `GET  /fetch/{upstream_id}/{record_id}` → streamed bytes (or 404)
+//! - `GET  /blob/{upstream_id}/{cid}`  → bytes (or 404)
+//!
+//! ## Bytes & the v1 simplification
+//!
+//! `/compute` returns each outcome's bytes **inline, base64-encoded**
+//! ([`OutcomeDto::bytes_b64`]) — this reuses every existing `compute_outcomes`
+//! impl verbatim (they already return bytes in `HashOutcome.bytes`) and keeps
+//! the gateway core's auto-store a single round-trip. base64 adds ~33%; fine
+//! for books/papers/posters. Large-file streaming (torznab multi-GiB) is a
+//! documented follow-up — those outcomes are metadata-only anyway
+//! (`bytes: None`), so they never ride this path.
+
+use std::collections::HashMap;
+use std::net::SocketAddr;
+use std::path::{Path as FsPath, PathBuf};
+use std::sync::Arc;
+
+use axum::{
+    body::Body,
+    extract::{Path, State},
+    http::StatusCode,
+    response::{Html, IntoResponse, Response},
+    routing::{get, post},
+    Json, Router,
+};
+use base64::Engine as _;
+use futures::StreamExt;
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+use tracing::{info, warn};
+
+use crate::config::{self, CONFIG_PAGE_HTML};
+use crate::plugin::{ConfigError, FeederPlugin, HashKind, HashOutcome};
+use crate::query::{GatewayQuery, GatewaySearchEvent, GatewayWireError};
+use crate::types::{DiscoveryRecord, GatewayError, PluginHealth};
+
+// -------- wire DTOs --------------------------------------------------------
+
+/// Serializable mirror of [`HashKind`] for the feeder HTTP boundary.
+#[derive(Serialize, Deserialize, Clone, Copy, PartialEq, Eq, Debug)]
+#[serde(rename_all = "snake_case")]
+pub enum HashKindDto {
+    Midhash256,
+    Sha2_256,
+    BtV1File,
+}
+
+impl From<HashKind> for HashKindDto {
+    fn from(k: HashKind) -> Self {
+        match k {
+            HashKind::Midhash256 => HashKindDto::Midhash256,
+            HashKind::Sha2_256 => HashKindDto::Sha2_256,
+            HashKind::BtV1File => HashKindDto::BtV1File,
+        }
+    }
+}
+
+impl From<HashKindDto> for HashKind {
+    fn from(k: HashKindDto) -> Self {
+        match k {
+            HashKindDto::Midhash256 => HashKind::Midhash256,
+            HashKindDto::Sha2_256 => HashKind::Sha2_256,
+            HashKindDto::BtV1File => HashKind::BtV1File,
+        }
+    }
+}
+
+/// One plugin's capability surface in [`ManifestResponse`].
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct PluginManifest {
+    pub id: String,
+    pub served_file_types: Vec<String>,
+    pub served_content_kinds: Vec<String>,
+}
+
+/// `GET /manifest` — the feeder's hosted plugins + their served types. The
+/// gateway core reads this to register `RemoteFeederPlugin`s and to build the
+/// capability heartbeat it broadcasts to meta-share.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct ManifestResponse {
+    pub feeder_version: String,
+    pub plugins: Vec<PluginManifest>,
+}
+
+/// One plugin's liveness in [`HealthResponse`].
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct PluginHealthEntry {
+    pub id: String,
+    pub health: PluginHealth,
+}
+
+/// `GET /health` — `status` is `"ok"` iff every hosted plugin reports `Ok`.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct HealthResponse {
+    pub status: String,
+    pub plugins: Vec<PluginHealthEntry>,
+}
+
+/// `POST /query` request body.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct QueryRequest {
+    pub upstream_id: String,
+    pub query: GatewayQuery,
+    pub max_results: u32,
+}
+
+/// `POST /query` response body.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct QueryResponse {
+    pub records: Vec<DiscoveryRecord>,
+}
+
+/// `POST /compute` request body.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct ComputeRequest {
+    pub upstream_id: String,
+    pub record_id: String,
+}
+
+/// One resolved outcome on the feeder wire. `bytes_b64` carries the full
+/// payload base64-encoded when the plugin fetched bytes (`Sha2_256` full
+/// store); `None` for metadata-only outcomes (the core's `(None, Some)`
+/// branch — e.g. torznab midhash).
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct OutcomeDto {
+    pub hash: String,
+    pub hash_kind: HashKindDto,
+    pub file_extension: Option<String>,
+    pub record: Option<DiscoveryRecord>,
+    pub bytes_b64: Option<String>,
+}
+
+impl OutcomeDto {
+    fn from_outcome(o: HashOutcome) -> Self {
+        OutcomeDto {
+            hash: o.hash.0,
+            hash_kind: o.hash_kind.into(),
+            file_extension: o.file_extension,
+            record: o.record,
+            bytes_b64: o
+                .bytes
+                .map(|b| base64::engine::general_purpose::STANDARD.encode(&b)),
+        }
+    }
+}
+
+/// `POST /compute` response body.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct ComputeResponse {
+    pub outcomes: Vec<OutcomeDto>,
+}
+
+// -------- app state + error mapping ----------------------------------------
+
+#[derive(Clone)]
+struct AppState {
+    plugins: Arc<HashMap<String, Arc<dyn FeederPlugin>>>,
+    version: String,
+    /// Feeder state dir. Per-plugin config persists at
+    /// `state_dir/gateway/<id>/config.json` — the same per-plugin dir handed to
+    /// `configure()`, so a UI save is read back on the next restart.
+    state_dir: PathBuf,
+}
+
+impl AppState {
+    fn plugin(&self, upstream_id: &str) -> Result<&Arc<dyn FeederPlugin>, Response> {
+        self.plugins.get(upstream_id).ok_or_else(|| {
+            (
+                StatusCode::NOT_FOUND,
+                format!("no feeder plugin for upstream_id '{upstream_id}'"),
+            )
+                .into_response()
+        })
+    }
+
+    /// `state_dir/gateway/<id>/config.json` — canonical per-plugin config file.
+    fn config_path(&self, upstream_id: &str) -> PathBuf {
+        self.state_dir
+            .join("gateway")
+            .join(upstream_id)
+            .join("config.json")
+    }
+
+    /// The effective config as JSON: the persisted `config.json` if it exists,
+    /// else the plugin's in-memory (env-seeded) config. Used as the redaction
+    /// source for reads and the merge base for writes.
+    fn effective_config(&self, upstream_id: &str) -> Value {
+        if let Ok(bytes) = std::fs::read(self.config_path(upstream_id)) {
+            if let Ok(v) = serde_json::from_slice::<Value>(&bytes) {
+                return v;
+            }
+        }
+        self.plugins
+            .get(upstream_id)
+            .map(|p| p.config_values())
+            .unwrap_or_else(|| serde_json::json!({}))
+    }
+}
+
+/// Map an in-process [`GatewayError`] to an HTTP status + message. Mirrors the
+/// status semantics the gateway's `RemoteFeederPlugin` will translate back
+/// into a `GatewayError` on the core side.
+fn gateway_error_response(e: GatewayError) -> Response {
+    let (status, msg) = match e {
+        GatewayError::NotFound => (StatusCode::NOT_FOUND, "not found".to_string()),
+        GatewayError::RateLimited { retry_after_s } => (
+            StatusCode::TOO_MANY_REQUESTS,
+            format!("rate limited; retry in {retry_after_s}s"),
+        ),
+        GatewayError::Transient(s) => (StatusCode::BAD_GATEWAY, s),
+        GatewayError::Permanent(s) => (StatusCode::UNPROCESSABLE_ENTITY, s),
+        GatewayError::Internal(s) => (StatusCode::INTERNAL_SERVER_ERROR, s.to_string()),
+    };
+    (status, msg).into_response()
+}
+
+// -------- handlers ---------------------------------------------------------
+
+async fn manifest(State(state): State<AppState>) -> Json<ManifestResponse> {
+    let mut plugins: Vec<PluginManifest> = state
+        .plugins
+        .values()
+        .map(|p| PluginManifest {
+            id: p.upstream_id().to_string(),
+            served_file_types: p.served_file_types().iter().map(|s| s.to_string()).collect(),
+            served_content_kinds: p
+                .served_content_kinds()
+                .iter()
+                .map(|s| s.to_string())
+                .collect(),
+        })
+        .collect();
+    plugins.sort_by(|a, b| a.id.cmp(&b.id));
+    Json(ManifestResponse {
+        feeder_version: state.version.clone(),
+        plugins,
+    })
+}
+
+async fn health(State(state): State<AppState>) -> Json<HealthResponse> {
+    let mut entries: Vec<PluginHealthEntry> = state
+        .plugins
+        .values()
+        .map(|p| PluginHealthEntry {
+            id: p.upstream_id().to_string(),
+            health: p.health(),
+        })
+        .collect();
+    entries.sort_by(|a, b| a.id.cmp(&b.id));
+    let all_ok = entries
+        .iter()
+        .all(|e| matches!(e.health, PluginHealth::Ok));
+    Json(HealthResponse {
+        status: if all_ok { "ok" } else { "degraded" }.to_string(),
+        plugins: entries,
+    })
+}
+
+async fn query(
+    State(state): State<AppState>,
+    Json(req): Json<QueryRequest>,
+) -> Result<Json<QueryResponse>, Response> {
+    let plugin = state.plugin(&req.upstream_id)?;
+    let records = plugin
+        .handle_query(&req.query, req.max_results as usize)
+        .await
+        .map_err(gateway_error_response)?;
+    Ok(Json(QueryResponse { records }))
+}
+
+/// `POST /query_stream` — the streaming analog of [`query`]. Emits the plugin's
+/// [`FeederPlugin::handle_query_stream`] events as **NDJSON**: one JSON-encoded
+/// [`GatewaySearchEvent`] per line (`application/x-ndjson`). The gateway core's
+/// `RemoteFeederPlugin` consumes this and re-emits a `BoxStream<GatewaySearchEvent>`
+/// into the libp2p streaming worker, preserving incremental enrichment (e.g.
+/// torznab's late TMDB / poster / subtitle-availability patches). A plugin that
+/// doesn't override `handle_query_stream` still works here — the trait default
+/// replays `handle_query` as `Base* + Done`.
+async fn query_stream(
+    State(state): State<AppState>,
+    Json(req): Json<QueryRequest>,
+) -> Result<Response, Response> {
+    let plugin = state.plugin(&req.upstream_id)?;
+    // The open can fail (e.g. the plugin can't reach its upstream at all); that
+    // maps to an HTTP error status, same as `/query`. Per-event errors instead
+    // ride the stream as a terminal `GatewaySearchEvent::Error` line.
+    let events = plugin
+        .handle_query_stream(&req.query, req.max_results as usize)
+        .await
+        .map_err(gateway_error_response)?;
+    let body = events.map(|ev| {
+        // Serialization of GatewaySearchEvent is infallible in practice; on the
+        // off chance it fails, terminate with an Error line so the consumer sees
+        // a clean terminal failure rather than a truncated stream.
+        let mut line = serde_json::to_vec(&ev).unwrap_or_else(|e| {
+            serde_json::to_vec(&GatewaySearchEvent::Error(GatewayWireError::Internal(
+                format!("event serialize: {e}"),
+            )))
+            .unwrap_or_default()
+        });
+        line.push(b'\n');
+        Ok::<_, std::convert::Infallible>(line)
+    });
+    Ok((
+        [(axum::http::header::CONTENT_TYPE, "application/x-ndjson")],
+        Body::from_stream(body),
+    )
+        .into_response())
+}
+
+async fn compute(
+    State(state): State<AppState>,
+    Json(req): Json<ComputeRequest>,
+) -> Result<Json<ComputeResponse>, Response> {
+    let plugin = state.plugin(&req.upstream_id)?;
+    let outcomes = plugin
+        .compute_outcomes(&req.record_id)
+        .await
+        .map_err(gateway_error_response)?;
+    Ok(Json(ComputeResponse {
+        outcomes: outcomes.into_iter().map(OutcomeDto::from_outcome).collect(),
+    }))
+}
+
+async fn fetch(
+    State(state): State<AppState>,
+    Path((upstream_id, record_id)): Path<(String, String)>,
+) -> Result<Response, Response> {
+    let plugin = state.plugin(&upstream_id)?;
+    match plugin
+        .handle_fetch(&record_id)
+        .await
+        .map_err(gateway_error_response)?
+    {
+        Some(stream) => {
+            // Map the plugin's GatewayError stream into an io::Error stream so
+            // axum's Body::from_stream accepts it.
+            let body_stream = stream.map(|chunk| {
+                chunk.map_err(|e| std::io::Error::other(e.to_string()))
+            });
+            Ok(Body::from_stream(body_stream).into_response())
+        }
+        None => Err((StatusCode::NOT_FOUND, "no bytes for record").into_response()),
+    }
+}
+
+async fn blob(
+    State(state): State<AppState>,
+    Path((upstream_id, cid)): Path<(String, String)>,
+) -> Result<Response, Response> {
+    let plugin = state.plugin(&upstream_id)?;
+    match plugin.get_blob(&cid).await {
+        Some(bytes) => Ok(bytes.into_response()),
+        None => Err((StatusCode::NOT_FOUND, "no such blob").into_response()),
+    }
+}
+
+// -------- config plane -----------------------------------------------------
+//
+// Generic, schema-driven per-plugin config. The page + the schema/values JSON
+// are all served relative to the request path, so this works identically when
+// hit directly on the feeder or reverse-proxied through the gateway.
+
+/// `GET /config` — the schema-driven HTML form (static, plugin-agnostic).
+async fn config_page() -> Html<&'static str> {
+    Html(CONFIG_PAGE_HTML)
+}
+
+/// `GET /config/schema` — the single hosted plugin's declared config schema.
+/// (A feeder hosts one plugin in practice; if it hosts several, the first by
+/// id is used — the gateway proxies per upstream_id anyway.)
+async fn config_schema(State(state): State<AppState>) -> Json<config::ConfigSchema> {
+    let schema = sole_plugin(&state)
+        .map(|p| p.config_schema())
+        .unwrap_or_default();
+    Json(schema)
+}
+
+/// `GET /config/values` — current values with secrets redacted to `<key>_set`.
+async fn config_values_get(State(state): State<AppState>) -> Json<Value> {
+    match sole_plugin(&state) {
+        Some(p) => {
+            let schema = p.config_schema();
+            let eff = state.effective_config(p.upstream_id());
+            Json(config::redact(&eff, &schema))
+        }
+        None => Json(serde_json::json!({})),
+    }
+}
+
+/// `PUT /config/values` — merge the submitted values onto the stored config
+/// (schema-aware; blank secrets keep their stored value), persist atomically,
+/// and signal that a feeder restart is needed to apply them.
+async fn config_values_put(
+    State(state): State<AppState>,
+    Json(incoming): Json<Value>,
+) -> Result<Json<Value>, Response> {
+    let plugin = sole_plugin(&state)
+        .ok_or_else(|| (StatusCode::NOT_FOUND, "feeder hosts no plugin").into_response())?;
+    let id = plugin.upstream_id();
+    let schema = plugin.config_schema();
+    let base = state.effective_config(id);
+    let merged = config::merge(&base, &incoming, &schema);
+
+    let path = state.config_path(id);
+    if let Some(dir) = path.parent() {
+        std::fs::create_dir_all(dir)
+            .map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, format!("mkdir: {e}")).into_response())?;
+    }
+    let tmp = path.with_extension("json.tmp");
+    let body = serde_json::to_vec_pretty(&merged)
+        .map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, format!("serialize: {e}")).into_response())?;
+    std::fs::write(&tmp, body)
+        .and_then(|_| std::fs::rename(&tmp, &path))
+        .map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, format!("write config: {e}")).into_response())?;
+    info!(target: "meta-feeder", upstream_id = id, path = %path.display(), "feeder config saved (restart to apply)");
+    Ok(Json(serde_json::json!({ "saved": true, "restart_required": true })))
+}
+
+/// `POST /enrich/callback` — sink for the enrichment plugins' completion
+/// callbacks. The feeder drives those plugins (filename-parser, tmdb) during a
+/// self-published resolve and reads results back by polling meta-core, so the
+/// callback carries no load — this route exists only so the plugins' callback
+/// POST gets a 200 instead of logging a connection failure. Body is ignored.
+async fn enrich_callback() -> StatusCode {
+    StatusCode::OK
+}
+
+/// A feeder hosts one plugin in the common case. Pick it (lowest id for
+/// determinism if several are present).
+fn sole_plugin(state: &AppState) -> Option<&Arc<dyn FeederPlugin>> {
+    state
+        .plugins
+        .values()
+        .min_by(|a, b| a.upstream_id().cmp(b.upstream_id()))
+}
+
+// -------- entry point ------------------------------------------------------
+
+/// Build the feeder router for an already-configured plugin set. Exposed for
+/// in-process tests (drive it without binding a socket).
+pub fn router(
+    plugins: HashMap<String, Arc<dyn FeederPlugin>>,
+    version: String,
+    state_dir: impl Into<PathBuf>,
+) -> Router {
+    let state = AppState {
+        plugins: Arc::new(plugins),
+        version,
+        state_dir: state_dir.into(),
+    };
+    Router::new()
+        .route("/manifest", get(manifest))
+        .route("/health", get(health))
+        .route("/query", post(query))
+        .route("/query_stream", post(query_stream))
+        .route("/compute", post(compute))
+        .route("/fetch/:upstream_id/:record_id", get(fetch))
+        .route("/blob/:upstream_id/:cid", get(blob))
+        .route("/config", get(config_page))
+        .route("/config/schema", get(config_schema))
+        .route("/config/values", get(config_values_get).put(config_values_put))
+        .route("/enrich/callback", post(enrich_callback))
+        .with_state(state)
+}
+
+/// Configure each plugin (cache dir under `state_dir/gateway/<id>`), then serve
+/// the feeder HTTP contract on `listen` until shutdown.
+///
+/// A plugin whose `configure()` returns [`ConfigError::MissingConfig`] is
+/// **soft-skipped** (warn + dropped) — mirroring the gateway registry's
+/// behaviour, so a misconfigured plugin doesn't take the whole feeder down.
+/// Any other `ConfigError` aborts startup.
+pub async fn serve_feeders(
+    plugins: Vec<Box<dyn FeederPlugin>>,
+    state_dir: impl Into<PathBuf>,
+    listen: SocketAddr,
+) -> anyhow::Result<()> {
+    let state_dir = state_dir.into();
+    let configured = configure_plugins(plugins, &state_dir)?;
+    let version = env!("CARGO_PKG_VERSION").to_string();
+    let app = router(configured, version, state_dir.clone());
+
+    let listener = tokio::net::TcpListener::bind(listen).await?;
+    info!(
+        target: "meta-feeder",
+        %listen,
+        "feeder listening"
+    );
+    axum::serve(listener, app)
+        .await
+        .map_err(anyhow::Error::from)
+}
+
+/// Run each plugin's `configure()` against its per-plugin cache dir, applying
+/// the soft-skip rule, and return the surviving plugins keyed by `upstream_id`.
+/// Public so feeder binaries / tests can configure without serving.
+pub fn configure_plugins(
+    plugins: Vec<Box<dyn FeederPlugin>>,
+    state_dir: &FsPath,
+) -> anyhow::Result<HashMap<String, Arc<dyn FeederPlugin>>> {
+    let mut out: HashMap<String, Arc<dyn FeederPlugin>> = HashMap::new();
+    for mut plugin in plugins {
+        let id = plugin.upstream_id();
+        let cache_dir = state_dir.join("gateway").join(id);
+        std::fs::create_dir_all(&cache_dir).map_err(|e| {
+            anyhow::anyhow!("create cache dir {} for {id}: {e}", cache_dir.display())
+        })?;
+        match plugin.configure(&cache_dir) {
+            Ok(()) => {
+                let id = id.to_string();
+                out.insert(id, Arc::from(plugin));
+            }
+            Err(ConfigError::MissingConfig { plugin: p, what }) => {
+                warn!(
+                    target: "meta-feeder",
+                    upstream_id = p,
+                    missing = what,
+                    "feeder plugin skipped: required config not supplied; \
+                     other plugins continue"
+                );
+            }
+            Err(other) => return Err(anyhow::anyhow!(other)),
+        }
+    }
+    Ok(out)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::types::DiscoveryRecord;
+    use async_trait::async_trait;
+    use futures::stream::BoxStream;
+    use std::collections::BTreeMap;
+
+    /// A feeder plugin that overrides `handle_query_stream` to emit a Base, then
+    /// an EnrichPatch, then Done — so the NDJSON route is exercised on a real
+    /// incremental stream (not just the collect-default).
+    struct StubStreaming;
+
+    #[async_trait]
+    impl FeederPlugin for StubStreaming {
+        fn upstream_id(&self) -> &'static str {
+            "stub"
+        }
+        fn configure(&mut self, _cache_dir: &FsPath) -> Result<(), ConfigError> {
+            Ok(())
+        }
+        async fn handle_query(
+            &self,
+            _q: &GatewayQuery,
+            _n: usize,
+        ) -> Result<Vec<DiscoveryRecord>, GatewayError> {
+            Ok(vec![])
+        }
+        async fn handle_query_stream(
+            &self,
+            _q: &GatewayQuery,
+            _n: usize,
+        ) -> Result<BoxStream<'static, GatewaySearchEvent>, GatewayError> {
+            let rec = DiscoveryRecord {
+                upstream_id: "stub".into(),
+                record_id: "1".into(),
+                fields: BTreeMap::new(),
+            };
+            let events = vec![
+                GatewaySearchEvent::Base(rec),
+                GatewaySearchEvent::EnrichPatch {
+                    record_id: "1".into(),
+                    set: BTreeMap::from([("title".to_string(), "X".to_string())]),
+                    remove: vec![],
+                },
+                GatewaySearchEvent::Done,
+            ];
+            Ok(Box::pin(futures::stream::iter(events)))
+        }
+        async fn compute_outcomes(
+            &self,
+            _record_id: &str,
+        ) -> Result<Vec<HashOutcome>, GatewayError> {
+            Ok(vec![])
+        }
+    }
+
+    #[tokio::test]
+    async fn query_stream_emits_ndjson_events_in_order() {
+        let mut plugins: HashMap<String, Arc<dyn FeederPlugin>> = HashMap::new();
+        plugins.insert("stub".to_string(), Arc::new(StubStreaming));
+        let app = router(plugins, "test".to_string(), std::env::temp_dir());
+
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let addr = listener.local_addr().unwrap();
+        tokio::spawn(async move {
+            axum::serve(listener, app).await.unwrap();
+        });
+
+        let body = reqwest::Client::new()
+            .post(format!("http://{addr}/query_stream"))
+            .json(&serde_json::json!({
+                "upstream_id": "stub",
+                "query": {"raw_text":"x","free_text":"x","filters":{},"ranges":[],"negations":[]},
+                "max_results": 5
+            }))
+            .send()
+            .await
+            .unwrap()
+            .text()
+            .await
+            .unwrap();
+
+        let events: Vec<GatewaySearchEvent> = body
+            .lines()
+            .filter(|l| !l.trim().is_empty())
+            .map(|l| serde_json::from_str(l).expect("each NDJSON line is a GatewaySearchEvent"))
+            .collect();
+
+        assert_eq!(events.len(), 3, "body was: {body:?}");
+        assert!(matches!(events[0], GatewaySearchEvent::Base(_)));
+        assert!(matches!(events[1], GatewaySearchEvent::EnrichPatch { .. }));
+        assert!(matches!(events[2], GatewaySearchEvent::Done));
+    }
+}

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

@@ -0,0 +1,124 @@
+//! In-process types used by the gateway runtime, feeders, and plugins.
+//!
+//! `GatewayError` here is **distinct from** the wire-stable
+//! `GatewayWireError` (see [`crate::query`]): this one can carry rich
+//! internal context (anyhow chains, plugin specifics); the wire one is narrow
+//! on purpose since adding variants is a protocol change. Conversion happens
+//! at the protocol boundary.
+
+use std::collections::BTreeMap;
+
+use serde::{Deserialize, Serialize};
+
+/// Content-addressed hash string carried over the wire. Renamed from
+/// `Midhash` in the multi-outcome plugin contract migration — the name
+/// `Midhash` lied for sha2-256 CIDs (which gateway plugins also emit
+/// via `compute_ipfs_cid`). The carrier is generic; the hash family is
+/// discriminated by `HashKind` on `HashOutcome`.
+///
+/// Shape: base32-lower 'b'-prefixed multibase CIDv1. The specific
+/// multihash inside (`0x1000` for midhash256, `0x12` for sha2-256) is
+/// determined by the producing plugin's hashing path.
+#[derive(Clone, Debug, PartialEq, Eq, Hash, Serialize, Deserialize)]
+pub struct Hash(pub String);
+
+impl Hash {
+    pub fn as_str(&self) -> &str {
+        &self.0
+    }
+}
+
+impl std::fmt::Display for Hash {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(&self.0)
+    }
+}
+
+/// Pre-hash identifier for a record discovered through a gateway.
+#[derive(Clone, Debug, PartialEq, Eq, Hash, Serialize, Deserialize)]
+pub struct DiscoveryId {
+    pub upstream_id: String,
+    pub record_id: String,
+}
+
+/// Metadata-only search result. Distinct from meta-core's persistent record
+/// (which is a flat `map[string]string` indexed by midhash); a
+/// `DiscoveryRecord` only lives inside the gateway search/response pipeline.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct DiscoveryRecord {
+    pub upstream_id: String,
+    pub record_id: String,
+    pub fields: BTreeMap<String, String>,
+}
+
+/// In-process error returned by `FeederPlugin` methods.
+#[derive(Debug, thiserror::Error)]
+pub enum GatewayError {
+    #[error("upstream transient error: {0}")]
+    Transient(String),
+
+    #[error("upstream permanent error: {0}")]
+    Permanent(String),
+
+    #[error("rate limited; retry in {retry_after_s}s")]
+    RateLimited { retry_after_s: u32 },
+
+    #[error("record not found upstream")]
+    NotFound,
+
+    #[error("internal: {0}")]
+    Internal(#[from] anyhow::Error),
+}
+
+/// Byte stream returned by a plugin's `handle_fetch`.
+pub type ByteStream = futures::stream::BoxStream<'static, Result<bytes::Bytes, GatewayError>>;
+
+/// Plugin liveness snapshot surfaced via `/health`. Must not call upstream
+/// — `health()` is invoked on every status poll.
+#[derive(Clone, Debug, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case", tag = "state")]
+pub enum PluginHealth {
+    Ok,
+    Degraded { reason: String },
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn hash_display_and_as_str_match() {
+        let m = Hash("bafy123".to_string());
+        assert_eq!(m.as_str(), "bafy123");
+        assert_eq!(format!("{m}"), "bafy123");
+    }
+
+    #[test]
+    fn discovery_record_roundtrips_through_serde_json() {
+        let r = DiscoveryRecord {
+            upstream_id: "scihub".to_string(),
+            record_id: "10.1038/s41586-021-03819-2".to_string(),
+            fields: BTreeMap::from_iter([
+                ("title".to_string(), "AlphaFold".to_string()),
+                ("year".to_string(), "2021".to_string()),
+            ]),
+        };
+        let j = serde_json::to_string(&r).expect("serialize");
+        let back: DiscoveryRecord = serde_json::from_str(&j).expect("deserialize");
+        assert_eq!(r, back);
+    }
+
+    #[test]
+    fn plugin_health_serialises_with_tag() {
+        let ok = serde_json::to_value(PluginHealth::Ok).unwrap();
+        assert_eq!(ok, serde_json::json!({"state": "ok"}));
+        let bad = serde_json::to_value(PluginHealth::Degraded {
+            reason: "upstream 5xx".into(),
+        })
+        .unwrap();
+        assert_eq!(
+            bad,
+            serde_json::json!({"state": "degraded", "reason": "upstream 5xx"})
+        );
+    }
+}

feeder-plugin/book-feeder/Cargo.toml

@@ -0,0 +1,40 @@
+[package]
+name = "book-feeder"
+version = "0.1.0"
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+publish = false
+
+# Book feeder sidecar: Project Gutenberg (via Gutendex). The proof-of-concept
+# feeder — single upstream, no preview-seeding, no BitTorrent — that validates
+# the meta-feeder-sdk + feeder HTTP contract end to end.
+
+[lib]
+name = "book_feeder"
+path = "src/lib.rs"
+
+[[bin]]
+name = "book-feeder"
+path = "src/main.rs"
+
+[dependencies]
+meta-feeder-sdk = { path = "../../crates/meta-feeder-sdk" }
+
+tokio = { version = "1.40", features = ["full"] }
+axum = { version = "0.7", features = ["json"] }
+anyhow = "1"
+async-trait = "0.1"
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+reqwest = { version = "0.12", features = ["json", "rustls-tls", "stream"], default-features = false }
+bytes = "1"
+futures = "0.3"
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["env-filter", "fmt"] }
+
+[dev-dependencies]
+wiremock = "0.6"
+tempfile = "3"
+base64 = "0.22"
+tokio = { version = "1.40", features = ["full", "test-util"] }

feeder-plugin/book-feeder/Dockerfile

@@ -0,0 +1,55 @@
+# syntax=docker/dockerfile:1
+# =============================================================================
+# book-feeder — Project Gutenberg feeder sidecar. Multi-stage build.
+#
+# The build context must be `packages/meta-gateway/` (the workspace root) so
+# the `meta-feeder-sdk` path dependency (../../crates/meta-feeder-sdk) resolves.
+# Same BuildKit cache-mount strategy as the gateway Dockerfile: the cargo
+# registry/git and `target/` persist across builds. The release binary is
+# copied OUT of the target cache mount within the same RUN.
+# =============================================================================
+
+FROM rust:1.89-slim-bookworm AS builder
+WORKDIR /build
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        pkg-config \
+        libssl-dev \
+        ca-certificates \
+        git \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY Cargo.toml Cargo.lock rust-toolchain.toml ./
+COPY crates/ crates/
+COPY feeder-plugin/ feeder-plugin/
+
+RUN --mount=type=cache,id=cargo-registry,target=/usr/local/cargo/registry \
+    --mount=type=cache,id=cargo-git,target=/usr/local/cargo/git \
+    --mount=type=cache,id=meta-gateway-target,target=/build/target \
+    cargo build --release --bin book-feeder \
+    && cp target/release/book-feeder /book-feeder
+
+# -----------------------------------------------------------------------------
+FROM debian:bookworm-slim AS runtime
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        ca-certificates \
+        curl \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY --from=builder /book-feeder /usr/local/bin/book-feeder
+
+RUN useradd --system --uid 10001 --no-create-home metafeeder \
+    && mkdir -p /data/meta-feeder \
+    && chown metafeeder:metafeeder /data/meta-feeder
+
+USER metafeeder
+WORKDIR /data/meta-feeder
+
+EXPOSE 8080
+
+ENV META_FEEDER_HTTP_LISTEN=0.0.0.0:8080 \
+    META_FEEDER_STATE_DIR=/data/meta-feeder \
+    RUST_LOG=info
+
+ENTRYPOINT ["/usr/local/bin/book-feeder"]

feeder-plugin/book-feeder/src/gutenberg.rs

@@ -0,0 +1,961 @@
+//! Project Gutenberg bridge via the [Gutendex](https://gutendex.com) JSON API.
+//!
+//! - Search: `GET /books?search={q}` → JSON list of book records.
+//! - Fetch one: `GET /books/{id}` → single book record with a `formats` map.
+//! - Canonical file format for midhash: `application/epub+zip`. Books that
+//!   don't expose an epub variant fail `compute_outcomes` with a permanent
+//!   error (operator-visible; very rare in practice).
+//!
+//! No auth, no API key, no rate-limit configuration. Moved verbatim from the
+//! gateway crate's `plugins/gutenberg.rs`; only the import paths changed
+//! (`crate::*`/`super::common` → `meta_feeder_sdk::*`, `GatewayPlugin` →
+//! `FeederPlugin`).
+
+use std::collections::{BTreeMap, HashMap};
+use std::path::Path;
+
+use async_trait::async_trait;
+use meta_feeder_sdk::cache::MidhashCache;
+use meta_feeder_sdk::common;
+#[cfg(test)]
+use meta_feeder_sdk::plugin::HashKind;
+use meta_feeder_sdk::plugin::{upstream_id_field, ConfigError, FeederPlugin, HashOutcome};
+use meta_feeder_sdk::query::GatewayQuery;
+use meta_feeder_sdk::types::{DiscoveryRecord, GatewayError, PluginHealth};
+use serde::Deserialize;
+use tracing::{debug, warn};
+
+/// Canonical Gutendex base URL. Overridable via [`GutenbergPlugin::with_base_url`]
+/// for tests + private mirrors.
+const DEFAULT_BASE_URL: &str = "https://gutendex.com";
+
+/// Format we treat as the canonical hashable bytes for a Gutenberg record.
+const EPUB_FORMAT: &str = "application/epub+zip";
+
+/// HTTP timeout for each upstream call.
+const HTTP_TIMEOUT_SECS: u64 = 30;
+
+/// User-Agent string sent on every Gutendex request.
+const USER_AGENT: &str = concat!("meta-share/", env!("CARGO_PKG_VERSION"), " (gateway:gutenberg)");
+
+/// Gutenberg gateway plugin. Cheap to construct; `configure()` opens the
+/// per-plugin redb cache.
+pub struct GutenbergPlugin {
+    http: reqwest::Client,
+    base_url: String,
+    cache: Option<MidhashCache>,
+}
+
+impl GutenbergPlugin {
+    pub fn new() -> Self {
+        Self::with_base_url(DEFAULT_BASE_URL.to_string())
+    }
+
+    /// Construct a plugin pointing at `base_url` instead of the public
+    /// Gutendex. Used by tests (against a `wiremock::MockServer`) and by
+    /// private-mirror deployments.
+    pub fn with_base_url(base_url: String) -> Self {
+        let http = common::build_http_client(HTTP_TIMEOUT_SECS, USER_AGENT, None);
+        Self {
+            http,
+            base_url,
+            cache: None,
+        }
+    }
+
+    fn cache(&self) -> Result<&MidhashCache, GatewayError> {
+        common::require_cache(self.cache.as_ref(), "gutenberg")
+    }
+
+    /// GET `{base}/books/{id}`; map HTTP/JSON failures to `GatewayError`.
+    async fn fetch_book(&self, record_id: &str) -> Result<GutendexBook, GatewayError> {
+        let url = format!("{}/books/{}", self.base_url.trim_end_matches('/'), record_id);
+        let resp = self
+            .http
+            .get(&url)
+            .send()
+            .await
+            .map_err(|e| GatewayError::Transient(format!("GET {url}: {e}")))?;
+        common::map_status(&resp)?;
+        resp.json::<GutendexBook>()
+            .await
+            .map_err(|e| GatewayError::Permanent(format!("decode gutendex book {record_id}: {e}")))
+    }
+
+    /// GET arbitrary upstream URL and return the body as bytes.
+    async fn fetch_bytes(&self, url: &str) -> Result<bytes::Bytes, GatewayError> {
+        common::fetch_bytes(&self.http, url).await
+    }
+
+    /// Fetch + parse each result's Project Gutenberg RDF/XML metadata for the
+    /// richer per-edition fields the Gutendex JSON API doesn't carry. Cached
+    /// aggressively by `record_id`. Failures degrade gracefully.
+    async fn fetch_bibrec_fields(&self, books: &[GutendexBook]) -> Vec<BTreeMap<String, String>> {
+        use futures::stream::{self, StreamExt};
+        const BIBREC_CONCURRENCY: usize = 6;
+        let cache = self.cache.clone();
+        let http = self.http.clone();
+        let ids: Vec<u64> = books.iter().map(|b| b.id).collect();
+        stream::iter(ids.into_iter())
+            .map(|id| {
+                let cache = cache.clone();
+                let http = http.clone();
+                async move {
+                    let record_id = id.to_string();
+                    if let Some(cached) = cache
+                        .as_ref()
+                        .and_then(|c| c.get_bibrec(&record_id).ok().flatten())
+                    {
+                        return cached;
+                    }
+                    let url = format!("https://www.gutenberg.org/cache/epub/{id}/pg{id}.rdf");
+                    let resp = tokio::time::timeout(
+                        std::time::Duration::from_secs(8),
+                        http.get(&url).send(),
+                    )
+                    .await;
+                    let xml = match resp {
+                        Ok(Ok(r)) if r.status().is_success() => r.text().await.ok(),
+                        Ok(Ok(r)) => {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                id,
+                                status = %r.status(),
+                                "rdf fetch non-2xx"
+                            );
+                            None
+                        }
+                        Ok(Err(e)) => {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                id,
+                                error = %e,
+                                "rdf fetch failed"
+                            );
+                            None
+                        }
+                        Err(_) => {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                id,
+                                "rdf fetch timed out"
+                            );
+                            None
+                        }
+                    };
+                    let Some(xml) = xml else {
+                        return BTreeMap::new();
+                    };
+                    let fields = parse_pgterms_rdf(&xml);
+                    if !fields.is_empty() {
+                        if let Some(cache) = cache.as_ref() {
+                            if let Err(e) = cache.put_bibrec(&record_id, &fields) {
+                                warn!(
+                                    target: "meta-share::gateway",
+                                    upstream = "gutenberg",
+                                    record_id = %record_id,
+                                    error = %e,
+                                    "bibrec put failed (non-fatal)"
+                                );
+                            }
+                        }
+                    }
+                    fields
+                }
+            })
+            .buffered(BIBREC_CONCURRENCY)
+            .collect()
+            .await
+    }
+
+    /// Download every result's cover image (where present), hash, and cache.
+    /// Returns a `Vec<Option<String>>` of cover CIDs aligned with the input
+    /// slice — `None` means no cover or download failure.
+    async fn fetch_cover_cids(&self, books: &[GutendexBook]) -> Vec<Option<String>> {
+        use futures::stream::{self, StreamExt};
+        const COVER_CONCURRENCY: usize = 8;
+        let cache = self.cache.clone();
+        let plugin_http = self.http.clone();
+        let to_process: Vec<(usize, String, Option<String>, Option<String>)> = books
+            .iter()
+            .enumerate()
+            .map(|(idx, b)| {
+                let record_id = b.id.to_string();
+                let cover_url = b.formats.get("image/jpeg").cloned();
+                let cached_cid = cache
+                    .as_ref()
+                    .and_then(|c| c.get_cover_cid(&record_id).ok().flatten());
+                (idx, record_id, cover_url, cached_cid)
+            })
+            .collect();
+        stream::iter(to_process.into_iter())
+            .map(|(idx, record_id, cover_url, cached_cid)| {
+                let cache = cache.clone();
+                let http = plugin_http.clone();
+                async move {
+                    if let Some(cid) = cached_cid {
+                        return Some(cid);
+                    }
+                    let url = cover_url?;
+                    let bytes_res = tokio::time::timeout(
+                        std::time::Duration::from_secs(8),
+                        http.get(&url).send(),
+                    )
+                    .await;
+                    let resp = match bytes_res {
+                        Ok(Ok(r)) if r.status().is_success() => r,
+                        Ok(Ok(r)) => {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                book_id = idx,
+                                status = %r.status(),
+                                "cover fetch non-2xx"
+                            );
+                            return None;
+                        }
+                        Ok(Err(e)) => {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                book_id = idx,
+                                error = %e,
+                                "cover fetch failed"
+                            );
+                            return None;
+                        }
+                        Err(_) => {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                book_id = idx,
+                                "cover fetch timed out"
+                            );
+                            return None;
+                        }
+                    };
+                    let bytes = match resp.bytes().await {
+                        Ok(b) => b,
+                        Err(e) => {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                book_id = idx,
+                                error = %e,
+                                "cover body read failed"
+                            );
+                            return None;
+                        }
+                    };
+                    let cid = meta_feeder_sdk::hash::compute_ipfs_cid(&bytes);
+                    if let Some(cache) = cache.as_ref() {
+                        if let Err(e) = cache.put_blob(&cid, &bytes) {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                cid = %cid,
+                                error = %e,
+                                "cover put_blob failed (non-fatal)"
+                            );
+                        }
+                        if let Err(e) = cache.put_cover_cid(&record_id, &cid) {
+                            warn!(
+                                target: "meta-share::gateway",
+                                upstream = "gutenberg",
+                                record_id = %record_id,
+                                error = %e,
+                                "cover put_cover_cid failed (non-fatal)"
+                            );
+                        }
+                    }
+                    Some(cid)
+                }
+            })
+            .buffered(COVER_CONCURRENCY)
+            .collect()
+            .await
+    }
+}
+
+impl Default for GutenbergPlugin {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[async_trait]
+impl FeederPlugin for GutenbergPlugin {
+    fn upstream_id(&self) -> &'static str {
+        "gutenberg"
+    }
+
+    fn configure(&mut self, cache_dir: &Path) -> Result<(), ConfigError> {
+        self.cache = Some(common::open_midhash_cache(cache_dir, "gutenberg")?);
+        Ok(())
+    }
+
+    async fn handle_query(
+        &self,
+        query: &GatewayQuery,
+        max_results: usize,
+    ) -> Result<Vec<DiscoveryRecord>, GatewayError> {
+        // Layer A early-return: Gutenberg only serves `document` / `book`.
+        if !meta_feeder_sdk::query_eval::query_accepts_plugin(
+            query,
+            self.served_file_types(),
+            self.served_content_kinds(),
+        ) {
+            return Ok(Vec::new());
+        }
+        let q = query.free_text_or_star();
+        let url = format!("{}/books", self.base_url.trim_end_matches('/'));
+        let resp = self
+            .http
+            .get(&url)
+            .query(&[("search", q)])
+            .send()
+            .await
+            .map_err(|e| GatewayError::Transient(format!("GET {url}: {e}")))?;
+        common::map_status(&resp)?;
+        let body: GutendexSearch = resp.json().await.map_err(|e| {
+            GatewayError::Permanent(format!("decode gutendex search response: {e}"))
+        })?;
+        debug!(
+            target: "meta-share::gateway",
+            upstream = "gutenberg",
+            query = q,
+            count = body.results.len(),
+            "search returned"
+        );
+        let raw: Vec<GutendexBook> = body.results.into_iter().take(max_results).collect();
+        let (cover_cids, bibrecs) =
+            tokio::join!(self.fetch_cover_cids(&raw), self.fetch_bibrec_fields(&raw),);
+        Ok(raw
+            .into_iter()
+            .zip(cover_cids)
+            .zip(bibrecs)
+            .map(|((book, cover_cid), bibrec)| into_discovery_record(book, cover_cid, bibrec))
+            .collect())
+    }
+
+    async fn compute_outcomes(&self, record_id: &str) -> Result<Vec<HashOutcome>, GatewayError> {
+        let cache = self.cache()?;
+        if let Some(hit) = common::cached_outcome(cache, record_id, "gutenberg")? {
+            return Ok(hit);
+        }
+
+        let book = self.fetch_book(record_id).await?;
+        let epub_url = book.formats.get(EPUB_FORMAT).cloned().ok_or_else(|| {
+            GatewayError::Permanent(format!(
+                "gutenberg record {record_id} has no {EPUB_FORMAT} format"
+            ))
+        })?;
+        let bytes = self.fetch_bytes(&epub_url).await?;
+        let cid = meta_feeder_sdk::hash::compute_ipfs_cid(&bytes);
+
+        common::store_midhash(cache, record_id, "gutenberg", &cid);
+
+        let cover_cid = cache.get_cover_cid(record_id).ok().flatten();
+        let bibrec = cache
+            .get_bibrec(record_id)
+            .ok()
+            .flatten()
+            .unwrap_or_default();
+        let record = into_discovery_record(book, cover_cid, bibrec);
+
+        Ok(common::single_outcome(
+            cid,
+            bytes,
+            record,
+            Some("epub".to_string()),
+        ))
+    }
+
+    async fn get_blob(&self, cid: &str) -> Option<Vec<u8>> {
+        match self.cache.as_ref()?.get_blob(cid) {
+            Ok(opt) => opt,
+            Err(e) => {
+                warn!(
+                    target: "meta-share::gateway",
+                    upstream = "gutenberg",
+                    cid,
+                    error = %e,
+                    "get_blob: cache read failed"
+                );
+                None
+            }
+        }
+    }
+
+    fn health(&self) -> PluginHealth {
+        if self.cache.is_some() {
+            PluginHealth::Ok
+        } else {
+            PluginHealth::Degraded {
+                reason: "configure() not yet called".to_string(),
+            }
+        }
+    }
+
+    fn served_file_types(&self) -> &'static [&'static str] {
+        &["document"]
+    }
+
+    fn served_content_kinds(&self) -> &'static [&'static str] {
+        &["book"]
+    }
+}
+
+/// Convert a `GutendexBook` into our wire-level `DiscoveryRecord`.
+fn into_discovery_record(
+    book: GutendexBook,
+    cover_cid: Option<String>,
+    bibrec: BTreeMap<String, String>,
+) -> DiscoveryRecord {
+    let GutendexBook {
+        id,
+        title,
+        authors,
+        subjects,
+        bookshelves,
+        languages,
+        copyright,
+        media_type,
+        formats,
+        download_count,
+    } = book;
+
+    let record_id = id.to_string();
+    let mut fields = BTreeMap::new();
+    fields.insert("title".into(), title);
+    fields.insert("fileType".into(), "document".into());
+    fields.insert("contentKind".into(), "book".into());
+    fields.insert(
+        "sourceUrl".into(),
+        format!("https://www.gutenberg.org/ebooks/{id}"),
+    );
+    // Canonical `<upstream_id>id` field.
+    fields.insert(upstream_id_field("gutenberg"), id.to_string());
+
+    if !authors.is_empty() {
+        fields.insert(
+            "author".into(),
+            authors
+                .iter()
+                .map(|a| a.name.as_str())
+                .collect::<Vec<_>>()
+                .join(", "),
+        );
+    }
+    if let Some(lang) = languages.first() {
+        fields.insert("language".into(), iso639_1_to_2(lang).to_string());
+    }
+    if !subjects.is_empty() {
+        fields.insert("subjects".into(), subjects.join(", "));
+    }
+    if !bookshelves.is_empty() {
+        fields.insert("bookshelves".into(), bookshelves.join(", "));
+    }
+    if let Some(mt) = media_type {
+        fields.insert("mediaType".into(), mt);
+    }
+    if let Some(cid) = cover_cid {
+        fields.insert("poster".into(), cid);
+    }
+    if formats.contains_key(EPUB_FORMAT) {
+        fields.insert("format".into(), "epub".into());
+    }
+    if let Some(c) = copyright {
+        fields.insert("publicDomain".into(), (!c).to_string());
+    }
+    let _ = download_count;
+    fields.insert("fileName".into(), format!("gutenberg-{id}.epub"));
+
+    for (k, v) in bibrec {
+        fields.entry(k).or_insert(v);
+    }
+
+    DiscoveryRecord {
+        upstream_id: "gutenberg".into(),
+        record_id,
+        fields,
+    }
+}
+
+/// Parse a Project Gutenberg `pg{id}.rdf` document for per-edition metadata.
+fn parse_pgterms_rdf(xml: &str) -> BTreeMap<String, String> {
+    let mut out = BTreeMap::new();
+
+    let pull = |tag: &str| extract_first_tag_text(xml, tag);
+
+    if let Some(v) = pull("dcterms:issued") {
+        out.insert("releaseDate".into(), v);
+    }
+    if let Some(v) = pull("dcterms:rights") {
+        out.insert("copyrightStatus".into(), v);
+    }
+    if let Some(v) = pull("dcterms:alternative") {
+        out.insert("alternateTitle".into(), v);
+    }
+    if let Some(v) = pull("dcterms:description") {
+        if let Some(url) = v
+            .split_whitespace()
+            .find(|t| t.contains("wikipedia.org") && t.starts_with("http"))
+        {
+            out.insert("wikipediaUrl".into(), url.to_string());
+        }
+        out.insert("note".into(), v);
+    }
+    if let Some(v) = pull("pgterms:marc508") {
+        out.insert("credits".into(), v);
+    }
+    if let Some(v) = pull("pgterms:marc520") {
+        out.insert("summary".into(), v);
+    }
+    if let Some(v) = pull("pgterms:marc908") {
+        out.insert("readingLevel".into(), v);
+    }
+    if let Some(url) = extract_resource_attr(xml, "pgterms:webpage") {
+        out.insert("authorWebpage".into(), url);
+    }
+
+    out
+}
+
+/// Find the first `<tag ...>BODY</tag>` and return BODY trimmed.
+fn extract_first_tag_text(xml: &str, tag: &str) -> Option<String> {
+    let open_prefix = format!("<{tag}");
+    let close = format!("</{tag}>");
+    let open_start = xml.find(&open_prefix)?;
+    let after_attrs = open_start + open_prefix.len();
+    let rest = &xml[after_attrs..];
+    let body_start = after_attrs + rest.find('>')? + 1;
+    let body_end_rel = xml[body_start..].find(&close)?;
+    let body = &xml[body_start..body_start + body_end_rel];
+    let decoded = body
+        .replace("&amp;", "&")
+        .replace("&lt;", "<")
+        .replace("&gt;", ">")
+        .replace("&quot;", "\"")
+        .replace("&apos;", "'");
+    let trimmed = decoded.trim().to_string();
+    if trimmed.is_empty() {
+        None
+    } else {
+        Some(trimmed)
+    }
+}
+
+/// Find the first `<tag ... rdf:resource="URL"...` and return URL.
+fn extract_resource_attr(xml: &str, tag: &str) -> Option<String> {
+    let open_prefix = format!("<{tag} ");
+    let open_start = xml.find(&open_prefix)?;
+    let rest = &xml[open_start + open_prefix.len()..];
+    let tag_end = rest.find('>')?;
+    let inside = &rest[..tag_end];
+    let needle = "rdf:resource=\"";
+    let attr_start = inside.find(needle)?;
+    let after = &inside[attr_start + needle.len()..];
+    let url_end = after.find('"')?;
+    Some(after[..url_end].to_string())
+}
+
+#[cfg(test)]
+mod rdf_parser_tests {
+    use super::*;
+
+    #[test]
+    fn extracts_release_date_and_copyright() {
+        let xml = r#"<rdf:RDF>
+            <pgterms:ebook rdf:about="ebooks/11">
+              <dcterms:issued rdf:datatype="...">2008-06-27</dcterms:issued>
+              <dcterms:rights>Public domain in the USA.</dcterms:rights>
+            </pgterms:ebook>
+          </rdf:RDF>"#;
+        let f = parse_pgterms_rdf(xml);
+        assert_eq!(f.get("releaseDate").map(String::as_str), Some("2008-06-27"));
+        assert_eq!(
+            f.get("copyrightStatus").map(String::as_str),
+            Some("Public domain in the USA.")
+        );
+    }
+
+    #[test]
+    fn extracts_credits_summary_reading_level() {
+        let xml = r#"<rdf:RDF>
+            <pgterms:marc508>Arthur DiBianca and David Widger</pgterms:marc508>
+            <pgterms:marc520>"Alice's..." short summary.</pgterms:marc520>
+            <pgterms:marc908>Reading ease score: 87.5 (6th grade). Easy to read.</pgterms:marc908>
+          </rdf:RDF>"#;
+        let f = parse_pgterms_rdf(xml);
+        assert_eq!(
+            f.get("credits").map(String::as_str),
+            Some("Arthur DiBianca and David Widger")
+        );
+        assert!(f.get("summary").unwrap().contains("short summary"));
+        assert!(f.get("readingLevel").unwrap().contains("87.5"));
+    }
+
+    #[test]
+    fn pulls_wikipedia_url_out_of_description() {
+        let xml = r#"<rdf:RDF>
+            <dcterms:description>Wikipedia page about this book: https://en.wikipedia.org/wiki/Alice%27s_Adventures_in_Wonderland</dcterms:description>
+          </rdf:RDF>"#;
+        let f = parse_pgterms_rdf(xml);
+        assert_eq!(
+            f.get("wikipediaUrl").map(String::as_str),
+            Some("https://en.wikipedia.org/wiki/Alice%27s_Adventures_in_Wonderland")
+        );
+        assert!(f.get("note").unwrap().contains("Wikipedia"));
+    }
+
+    #[test]
+    fn empty_when_not_an_rdf_doc() {
+        let f = parse_pgterms_rdf("<html><body>not rdf</body></html>");
+        assert!(f.is_empty());
+    }
+}
+
+/// ISO 639-1 (two-letter) → ISO 639-2/T (three-letter) for the ~20 most common
+/// Gutenberg languages. Unknown codes pass through unchanged.
+fn iso639_1_to_2(code: &str) -> &str {
+    match code {
+        "en" => "eng",
+        "fr" => "fra",
+        "de" => "deu",
+        "es" => "spa",
+        "it" => "ita",
+        "pt" => "por",
+        "nl" => "nld",
+        "la" => "lat",
+        "el" => "ell",
+        "ru" => "rus",
+        "fi" => "fin",
+        "sv" => "swe",
+        "da" => "dan",
+        "no" => "nor",
+        "pl" => "pol",
+        "cs" => "ces",
+        "hu" => "hun",
+        "ja" => "jpn",
+        "zh" => "zho",
+        "ar" => "ara",
+        "ko" => "kor",
+        "tr" => "tur",
+        "ro" => "ron",
+        "ca" => "cat",
+        "is" => "isl",
+        other => other,
+    }
+}
+
+// -- Gutendex JSON shapes (subset). We only deserialise the fields we use.
+
+#[derive(Debug, Deserialize)]
+struct GutendexSearch {
+    results: Vec<GutendexBook>,
+}
+
+#[derive(Debug, Deserialize)]
+struct GutendexBook {
+    id: u64,
+    title: String,
+    #[serde(default)]
+    authors: Vec<GutendexPerson>,
+    #[serde(default)]
+    subjects: Vec<String>,
+    #[serde(default)]
+    bookshelves: Vec<String>,
+    #[serde(default)]
+    languages: Vec<String>,
+    #[serde(default)]
+    copyright: Option<bool>,
+    #[serde(default)]
+    media_type: Option<String>,
+    #[serde(default)]
+    formats: HashMap<String, String>,
+    #[serde(default)]
+    download_count: u64,
+}
+
+#[derive(Debug, Deserialize)]
+struct GutendexPerson {
+    name: String,
+    // Gutendex also returns `birth_year` / `death_year`, but those are about
+    // the author, not this book file, so we don't deserialize or surface them
+    // (serde ignores the unknown JSON keys). See `into_discovery_record`.
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use meta_feeder_sdk::types::PluginHealth;
+    use wiremock::matchers::{method, path, query_param};
+    use wiremock::{Mock, MockServer, ResponseTemplate};
+
+    fn books_json() -> serde_json::Value {
+        serde_json::json!({
+            "count": 2,
+            "next": null,
+            "previous": null,
+            "results": [
+                {
+                    "id": 1342,
+                    "title": "Pride and Prejudice",
+                    "authors": [{"name": "Austen, Jane", "birth_year": 1775, "death_year": 1817}],
+                    "languages": ["en"],
+                    "formats": {
+                        "application/epub+zip": "EPUB_URL_PLACEHOLDER",
+                        "text/plain; charset=utf-8": "https://www.gutenberg.org/files/1342/1342-0.txt"
+                    }
+                },
+                {
+                    "id": 84,
+                    "title": "Frankenstein; Or, The Modern Prometheus",
+                    "authors": [{"name": "Shelley, Mary Wollstonecraft", "birth_year": 1797, "death_year": 1851}],
+                    "languages": ["en"],
+                    "formats": {
+                        "application/epub+zip": "EPUB_URL_PLACEHOLDER",
+                        "text/plain; charset=utf-8": "https://www.gutenberg.org/files/84/84-0.txt"
+                    }
+                }
+            ]
+        })
+    }
+
+    fn configured_plugin_against(server: &MockServer) -> (GutenbergPlugin, tempfile::TempDir) {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let mut plugin = GutenbergPlugin::with_base_url(server.uri());
+        plugin.configure(dir.path()).expect("configure");
+        (plugin, dir)
+    }
+
+    #[tokio::test]
+    async fn handle_query_maps_results_to_discovery_records() {
+        let server = MockServer::start().await;
+        Mock::given(method("GET"))
+            .and(path("/books"))
+            .and(query_param("search", "alice"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(books_json()))
+            .mount(&server)
+            .await;
+
+        let (plugin, _dir) = configured_plugin_against(&server);
+        let records = plugin
+            .handle_query(&GatewayQuery::from_free_text("alice"), 50)
+            .await
+            .expect("handle_query");
+        assert_eq!(records.len(), 2);
+
+        let r0 = &records[0];
+        assert_eq!(r0.upstream_id, "gutenberg");
+        assert_eq!(r0.record_id, "1342");
+        assert_eq!(
+            r0.fields.get("title").map(String::as_str),
+            Some("Pride and Prejudice")
+        );
+        assert_eq!(r0.fields.get("fileType").map(String::as_str), Some("document"));
+        assert_eq!(r0.fields.get("contentKind").map(String::as_str), Some("book"));
+        assert_eq!(r0.fields.get("gutenbergid").map(String::as_str), Some("1342"));
+        assert_eq!(
+            r0.fields.get("sourceUrl").map(String::as_str),
+            Some("https://www.gutenberg.org/ebooks/1342")
+        );
+        assert_eq!(r0.fields.get("author").map(String::as_str), Some("Austen, Jane"));
+        assert_eq!(r0.fields.get("language").map(String::as_str), Some("eng"));
+        assert_eq!(r0.fields.get("format").map(String::as_str), Some("epub"));
+        assert_eq!(
+            r0.fields.get("fileName").map(String::as_str),
+            Some("gutenberg-1342.epub")
+        );
+    }
+
+    #[tokio::test]
+    async fn handle_query_truncates_to_max_results() {
+        let server = MockServer::start().await;
+        Mock::given(method("GET"))
+            .and(path("/books"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(books_json()))
+            .mount(&server)
+            .await;
+
+        let (plugin, _dir) = configured_plugin_against(&server);
+        let records = plugin
+            .handle_query(&GatewayQuery::from_free_text("alice"), 1)
+            .await
+            .expect("ok");
+        assert_eq!(records.len(), 1);
+    }
+
+    #[tokio::test]
+    async fn compute_outcomes_cache_hit_skips_http() {
+        let server = MockServer::start().await;
+        let (plugin, _dir) = configured_plugin_against(&server);
+        plugin
+            .cache
+            .as_ref()
+            .unwrap()
+            .put_midhash("1342", "bafyCACHED")
+            .unwrap();
+
+        let outcomes = plugin.compute_outcomes("1342").await.expect("cache hit");
+        assert_eq!(outcomes.len(), 1);
+        assert_eq!(outcomes[0].hash.as_str(), "bafyCACHED");
+    }
+
+    #[tokio::test]
+    async fn compute_outcomes_fetches_epub_and_caches() {
+        let server = MockServer::start().await;
+
+        let epub_bytes: Vec<u8> = b"<<fake epub bytes for test>>".to_vec();
+        let expected_cid = meta_feeder_sdk::hash::compute_ipfs_cid(&epub_bytes);
+
+        let epub_path = "/ebooks/1342.epub.images";
+        let book = serde_json::json!({
+            "id": 1342,
+            "title": "Pride and Prejudice",
+            "authors": [{"name": "Austen, Jane"}],
+            "languages": ["en"],
+            "formats": {
+                "application/epub+zip": format!("{}{epub_path}", server.uri()),
+            }
+        });
+        Mock::given(method("GET"))
+            .and(path("/books/1342"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(book))
+            .mount(&server)
+            .await;
+        Mock::given(method("GET"))
+            .and(path(epub_path))
+            .respond_with(
+                ResponseTemplate::new(200)
+                    .set_body_bytes(epub_bytes.clone())
+                    .insert_header("content-type", "application/epub+zip"),
+            )
+            .mount(&server)
+            .await;
+
+        let (plugin, _dir) = configured_plugin_against(&server);
+        let outcomes = plugin.compute_outcomes("1342").await.expect("compute");
+        assert_eq!(outcomes.len(), 1);
+        assert_eq!(outcomes[0].hash.as_str(), expected_cid);
+        assert_eq!(outcomes[0].hash_kind, HashKind::Sha2_256);
+        assert_eq!(
+            plugin
+                .cache
+                .as_ref()
+                .unwrap()
+                .get_midhash("1342")
+                .unwrap()
+                .as_deref(),
+            Some(expected_cid.as_str())
+        );
+    }
+
+    #[tokio::test]
+    async fn compute_outcomes_missing_epub_returns_permanent() {
+        let server = MockServer::start().await;
+        let book = serde_json::json!({
+            "id": 99999,
+            "title": "Audio-only Record",
+            "authors": [],
+            "languages": ["en"],
+            "formats": {
+                "text/plain; charset=utf-8": format!("{}/files/99999/99999-0.txt", server.uri()),
+            }
+        });
+        Mock::given(method("GET"))
+            .and(path("/books/99999"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(book))
+            .mount(&server)
+            .await;
+
+        let (plugin, _dir) = configured_plugin_against(&server);
+        let err = plugin
+            .compute_outcomes("99999")
+            .await
+            .expect_err("should error");
+        match err {
+            GatewayError::Permanent(msg) => {
+                assert!(msg.contains("epub"), "msg: {msg}");
+                assert!(msg.contains("99999"), "msg: {msg}");
+            }
+            other => panic!("wrong variant: {other:?}"),
+        }
+    }
+
+    #[tokio::test]
+    async fn upstream_404_maps_to_not_found() {
+        let server = MockServer::start().await;
+        Mock::given(method("GET"))
+            .and(path("/books/1"))
+            .respond_with(ResponseTemplate::new(404))
+            .mount(&server)
+            .await;
+
+        let (plugin, _dir) = configured_plugin_against(&server);
+        let err = plugin.compute_outcomes("1").await.expect_err("404");
+        assert!(matches!(err, GatewayError::NotFound));
+    }
+
+    #[tokio::test]
+    async fn upstream_5xx_maps_to_transient() {
+        let server = MockServer::start().await;
+        Mock::given(method("GET"))
+            .and(path("/books"))
+            .respond_with(ResponseTemplate::new(502))
+            .mount(&server)
+            .await;
+
+        let (plugin, _dir) = configured_plugin_against(&server);
+        let err = plugin
+            .handle_query(&GatewayQuery::from_free_text("x"), 10)
+            .await
+            .expect_err("502");
+        assert!(matches!(err, GatewayError::Transient(_)), "got: {err:?}");
+    }
+
+    #[tokio::test]
+    async fn upstream_429_maps_to_rate_limited_with_retry_after() {
+        let server = MockServer::start().await;
+        Mock::given(method("GET"))
+            .and(path("/books"))
+            .respond_with(ResponseTemplate::new(429).insert_header("retry-after", "90"))
+            .mount(&server)
+            .await;
+
+        let (plugin, _dir) = configured_plugin_against(&server);
+        let err = plugin
+            .handle_query(&GatewayQuery::from_free_text("x"), 10)
+            .await
+            .expect_err("429");
+        match err {
+            GatewayError::RateLimited { retry_after_s } => {
+                assert_eq!(retry_after_s, 90);
+            }
+            other => panic!("wrong variant: {other:?}"),
+        }
+    }
+
+    #[test]
+    fn health_reflects_configure_state() {
+        let plugin = GutenbergPlugin::new();
+        match plugin.health() {
+            PluginHealth::Degraded { reason } => assert!(reason.contains("configure")),
+            PluginHealth::Ok => panic!("unconfigured plugin should not be Ok"),
+        }
+
+        let dir = tempfile::tempdir().unwrap();
+        let mut plugin = GutenbergPlugin::new();
+        plugin.configure(dir.path()).unwrap();
+        match plugin.health() {
+            PluginHealth::Ok => {}
+            PluginHealth::Degraded { reason } => {
+                panic!("configured plugin should be Ok, got Degraded: {reason}")
+            }
+        }
+    }
+}

feeder-plugin/book-feeder/src/lib.rs

@@ -0,0 +1,5 @@
+//! `book-feeder` library surface — exposes the Gutenberg plugin so integration
+//! tests (and, later, other binaries) can construct it. The binary
+//! (`main.rs`) is a thin wrapper over [`meta_feeder_sdk::serve_feeders`].
+
+pub mod gutenberg;

feeder-plugin/book-feeder/src/main.rs

@@ -0,0 +1,31 @@
+//! `book-feeder` — Project Gutenberg feeder sidecar.
+//!
+//! Hosts the [`gutenberg::GutenbergPlugin`] over the feeder HTTP contract via
+//! [`meta_feeder_sdk::serve_feeders`]. The gateway core reaches it server-to-
+//! server on the internal docker network.
+//!
+//! Env:
+//! - `META_FEEDER_HTTP_LISTEN` — listen addr (default `0.0.0.0:8080`)
+//! - `META_FEEDER_STATE_DIR`   — per-plugin cache root (default `/data/meta-feeder`)
+//! - `RUST_LOG`                — tracing filter (default `info`)
+
+use std::net::SocketAddr;
+
+use book_feeder::gutenberg::GutenbergPlugin;
+use meta_feeder_sdk::serve_feeders;
+use tracing_subscriber::EnvFilter;
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    tracing_subscriber::fmt()
+        .with_env_filter(EnvFilter::try_from_default_env().unwrap_or_else(|_| EnvFilter::new("info")))
+        .init();
+
+    let listen: SocketAddr = std::env::var("META_FEEDER_HTTP_LISTEN")
+        .unwrap_or_else(|_| "0.0.0.0:8080".to_string())
+        .parse()?;
+    let state_dir =
+        std::env::var("META_FEEDER_STATE_DIR").unwrap_or_else(|_| "/data/meta-feeder".to_string());
+
+    serve_feeders(vec![Box::new(GutenbergPlugin::new())], state_dir, listen).await
+}

feeder-plugin/book-feeder/tests/contract.rs

@@ -0,0 +1,172 @@
+//! End-to-end contract test for the feeder HTTP surface: boots the real
+//! `serve_feeders` router (book-feeder hosting GutenbergPlugin) on an ephemeral
+//! port, pointed at a wiremock Gutendex, and drives `/manifest`, `/query`, and
+//! `/compute` over HTTP. Deterministic — no live Gutendex.
+
+use book_feeder::gutenberg::GutenbergPlugin;
+use meta_feeder_sdk::{
+    configure_plugins, hash, router, ComputeRequest, ComputeResponse, GatewayQuery, HashKindDto,
+    ManifestResponse, QueryRequest, QueryResponse,
+};
+use wiremock::matchers::{method, path};
+use wiremock::{Mock, MockServer, ResponseTemplate};
+
+/// Spawn the feeder router (Gutenberg pointed at `upstream`) on an ephemeral
+/// port; return its base URL. The TempDir is returned so the per-plugin cache
+/// outlives the test.
+async fn spawn_feeder(upstream: &MockServer) -> (String, tempfile::TempDir) {
+    let dir = tempfile::tempdir().expect("tempdir");
+    let plugin = GutenbergPlugin::with_base_url(upstream.uri());
+    let plugins =
+        configure_plugins(vec![Box::new(plugin)], dir.path()).expect("configure plugins");
+    let app = router(plugins, "book-feeder-test".to_string(), dir.path());
+
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0")
+        .await
+        .expect("bind ephemeral");
+    let addr = listener.local_addr().expect("local_addr");
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.expect("serve");
+    });
+    (format!("http://{addr}"), dir)
+}
+
+fn books_json() -> serde_json::Value {
+    serde_json::json!({
+        "results": [{
+            "id": 1342,
+            "title": "Pride and Prejudice",
+            "authors": [{"name": "Austen, Jane"}],
+            "languages": ["en"],
+            "formats": { "application/epub+zip": "EPUB_PLACEHOLDER" }
+        }]
+    })
+}
+
+#[tokio::test]
+async fn manifest_lists_book_capabilities() {
+    let upstream = MockServer::start().await;
+    let (base, _dir) = spawn_feeder(&upstream).await;
+
+    let manifest: ManifestResponse = reqwest::Client::new()
+        .get(format!("{base}/manifest"))
+        .send()
+        .await
+        .expect("GET /manifest")
+        .json()
+        .await
+        .expect("decode manifest");
+
+    assert_eq!(manifest.plugins.len(), 1);
+    let p = &manifest.plugins[0];
+    assert_eq!(p.id, "gutenberg");
+    assert_eq!(p.served_file_types, vec!["document".to_string()]);
+    assert_eq!(p.served_content_kinds, vec!["book".to_string()]);
+}
+
+#[tokio::test]
+async fn query_returns_gutenberg_records() {
+    let upstream = MockServer::start().await;
+    Mock::given(method("GET"))
+        .and(path("/books"))
+        .respond_with(ResponseTemplate::new(200).set_body_json(books_json()))
+        .mount(&upstream)
+        .await;
+    let (base, _dir) = spawn_feeder(&upstream).await;
+
+    let resp: QueryResponse = reqwest::Client::new()
+        .post(format!("{base}/query"))
+        .json(&QueryRequest {
+            upstream_id: "gutenberg".to_string(),
+            query: GatewayQuery::from_free_text("alice"),
+            max_results: 50,
+        })
+        .send()
+        .await
+        .expect("POST /query")
+        .json()
+        .await
+        .expect("decode query response");
+
+    assert_eq!(resp.records.len(), 1);
+    let r = &resp.records[0];
+    assert_eq!(r.upstream_id, "gutenberg");
+    assert_eq!(r.fields.get("gutenbergid").map(String::as_str), Some("1342"));
+    assert_eq!(r.fields.get("contentKind").map(String::as_str), Some("book"));
+}
+
+#[tokio::test]
+async fn compute_returns_outcome_with_inline_epub_bytes() {
+    let upstream = MockServer::start().await;
+    let epub_bytes = b"<<fake epub bytes for contract test>>".to_vec();
+    let expected_cid = hash::compute_ipfs_cid(&epub_bytes);
+    let epub_path = "/ebooks/1342.epub";
+
+    let book = serde_json::json!({
+        "id": 1342,
+        "title": "Pride and Prejudice",
+        "authors": [{"name": "Austen, Jane"}],
+        "languages": ["en"],
+        "formats": { "application/epub+zip": format!("{}{epub_path}", upstream.uri()) }
+    });
+    Mock::given(method("GET"))
+        .and(path("/books/1342"))
+        .respond_with(ResponseTemplate::new(200).set_body_json(book))
+        .mount(&upstream)
+        .await;
+    Mock::given(method("GET"))
+        .and(path(epub_path))
+        .respond_with(ResponseTemplate::new(200).set_body_bytes(epub_bytes.clone()))
+        .mount(&upstream)
+        .await;
+
+    let (base, _dir) = spawn_feeder(&upstream).await;
+
+    let resp: ComputeResponse = reqwest::Client::new()
+        .post(format!("{base}/compute"))
+        .json(&ComputeRequest {
+            upstream_id: "gutenberg".to_string(),
+            record_id: "1342".to_string(),
+        })
+        .send()
+        .await
+        .expect("POST /compute")
+        .json()
+        .await
+        .expect("decode compute response");
+
+    assert_eq!(resp.outcomes.len(), 1);
+    let o = &resp.outcomes[0];
+    assert_eq!(o.hash, expected_cid);
+    assert_eq!(o.hash_kind, HashKindDto::Sha2_256);
+    assert_eq!(o.file_extension.as_deref(), Some("epub"));
+    // Bytes ride inline base64 on the v1 contract — decode + verify they hash
+    // back to the advertised CID (the core's auto-store will do exactly this).
+    let b64 = o.bytes_b64.as_ref().expect("compute returned inline bytes");
+    use base64::Engine as _;
+    let decoded = base64::engine::general_purpose::STANDARD
+        .decode(b64)
+        .expect("valid base64");
+    assert_eq!(decoded, epub_bytes);
+    assert_eq!(hash::compute_ipfs_cid(&decoded), expected_cid);
+}
+
+#[tokio::test]
+async fn query_unknown_upstream_is_404() {
+    let upstream = MockServer::start().await;
+    let (base, _dir) = spawn_feeder(&upstream).await;
+
+    let status = reqwest::Client::new()
+        .post(format!("{base}/query"))
+        .json(&QueryRequest {
+            upstream_id: "torznab".to_string(),
+            query: GatewayQuery::from_free_text("x"),
+            max_results: 10,
+        })
+        .send()
+        .await
+        .expect("POST /query")
+        .status();
+
+    assert_eq!(status, reqwest::StatusCode::NOT_FOUND);
+}

rust-toolchain.toml

@@ -0,0 +1,3 @@
+[toolchain]
+channel = "1.89"
+components = ["rustfmt", "clippy"]