initial commit

.gitattributes

@@ -0,0 +1,6 @@
+# FIX CRLF always when developer has not set
+# Linux/Mac: git config --global core.autocrlf input
+# Windows: git config --global core.autocrlf true
+# Auto detect text files and perform LF normalization
+* text=auto
+* eol=lf

.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/card-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/card-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-card"

README.md

@@ -1 +1,98 @@
 # meta-feeder-card
+
+The **card tier** feeder for the MetaMesh gateway — the *discovery* half of
+gateway search.
+
+A **card** is a metadata-only record identifying a *work* (a series, a film),
+not a file: title, poster, description, and an **id bag** of every cross-source
+identifier the bridge could resolve. It has no bytes, and by design never will.
+
+```
+Phase 1 — CARD (this feeder)          Phase 2 — RETRIEVAL (content feeders)
+  <free text> fileType:card              fileType:video + the card's id bag
+  one cached TMDB search/multi           the full indexer budget, ONE work
+  → a poster grid of works               → the episode / version list
+                        ── user clicks a card ──
+```
+
+Splitting the two halves is the point. Previously one query resolved ten
+speculative candidate works *and* searched every indexer for all of them, against
+a shared 1 req/s bucket — so the work the user actually wanted got a tenth of
+the budget. A card query touches TMDB only; the indexer budget is spent later,
+entirely on the card that was clicked.
+
+## Upstreams
+
+| upstream_id | source | status |
+|---|---|---|
+| `tmdb` | themoviedb.org | shipping |
+| `myanimelist` | myanimelist.net | planned — a sibling `FeederPlugin` in this same binary |
+
+Both declare `served_file_types = ["card"]`, which is the whole routing story:
+meta-search fans a `fileType:card` query out to every gateway advertising it, so
+adding this tier needed no meta-search change.
+
+## Card identity
+
+A card's CID is a **`0x1007` card locator** — an identity-multihash CIDv1 whose
+payload is `varint(len(source)) ‖ source ‖ id`, e.g. `("tmdb", "tv:95479")`.
+
+- **Deterministic in `(source, id)`**, so any peer holding a tmdb id derives the
+  address offline, with no discovery round-trip — and the same card found by two
+  peers converges on one meta-core record for free.
+- **Ranks 5** (opaque-locator tier), so it never outranks a real digest. That is
+  what will make a future cross-source merge — one meta-core record carrying
+  *both* a TMDB and a MyAnimeList locator — safe.
+- **Never seeded to bitswap** (seeding is gated on sha2-256). Nothing to seed.
+
+Cards are treated as **immutable**: no TTL, no refresh, and no drifting values
+(rating, popularity) stored on them. Growth is append-only — a future *season
+card* is an addition beside the series card, never an edit to it.
+
+## Config
+
+Per-plugin config lives on the feeder (gateway invariant 12), served as a
+schema-driven form at `GET /config`:
+
+| key | meaning |
+|---|---|
+| `tmdb_token` | TMDB v4 read token. **Required** — without it the feeder soft-skips (stays healthy, returns no cards). |
+| `tmdb_rate_per_sec` | Sustained TMDB request rate. Default 20/s. |
+| `tmdb_burst` | Short-burst token ceiling. Default 20. |
+| `card_top_n` | Works returned per free-text search. Default 10. Costs no indexer requests. |
+
+Env (`TMDB_TOKEN`, `TMDB_RATE_PER_SEC`, `TMDB_BURST`, `CARD_TOP_N`) seeds the
+first boot; the config file wins from then on. No hot reload — bounce the feeder.
+
+## Build / run
+
+```bash
+cargo build --release --bin card-feeder
+cargo test -p card-feeder
+
+# Image (build context is the repo root so the vendored SDK path dep resolves)
+docker build -f feeder-plugin/card-feeder/Dockerfile -t ghcr.io/worph/meta-feeder-card:dev .
+```
+
+Register it with a gateway by adding it to that gateway's `gateway-config.json`
+`feeders` map (or via the dashboard) and restarting:
+
+```json
+{ "feeders": { "card-feeder": { "url": "http://card-feeder-hub:8080" } } }
+```
+
+## Layout
+
+```
+crates/meta-feeder-sdk/        vendored SDK copy (mirrored from meta-gateway)
+feeder-plugin/card-feeder/
+  src/tmdb.rs                  the FeederPlugin impl — routing + query + outcomes
+  src/card.rs                  the Card model + its DiscoveryRecord projection
+  src/resolve.rs               cache- and budget-gated TMDB resolution
+  src/tmdb_client.rs           TMDB v3 HTTP client + DTOs
+  src/tmdb_budget.rs           token bucket
+  src/consts.rs
+```
+
+Design rationale, including what was deliberately *not* built:
+`meta-gateway/docs/others/card-tier-search.md`.

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 +1,268 @@
+<!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; }
+  .req { color: #f0a0a0; margin-left: 3px; font-weight: 700; }
+  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. Fields marked <span class="req">*</span> are required.</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; }
+
+// A "*" marker for required fields (null for optional ones).
+function reqStar(f) { return f.required ? el("span", { class: "req", text: "*", title: "required" }) : null; }
+
+// Render one scalar field. `vals` is the object holding the value/`_set` flag.
+// Every returned widget exposes `_collect()` and `_key` so collect() can read it.
+function fieldInput(f, vals) {
+  if (f.type === "bool") {
+    const cb = el("input", { type: "checkbox" });
+    cb.checked = vals[f.key] === true;
+    const w = el("div", { class: "field" },
+      el("label", {}, cb, " ", el("span", { text: " " + f.label }), reqStar(f)),
+      helpEl(f));
+    w._collect = () => cb.checked;
+    w._key = f.key;
+    return w;
+  }
+  const label = el("label", { text: f.label });
+  const star = reqStar(f); if (star) label.append(star);
+  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;
+  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 lbl = el("label", { text: f.label });
+  const star = reqStar(f); if (star) lbl.append(star);
+  const wrap = el("div", { class: "field" }, lbl, 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) {
+      const detail = (await res.text().catch(() => "")).trim();
+      throw new Error(detail ? `the feeder rejected it (HTTP ${res.status}): ${detail}`
+                             : `the feeder returned HTTP ${res.status}`);
+    }
+    const j = await res.json().catch(() => ({}));
+    if (j.reloading) {
+      // The feeder is restarting itself (~2-3s) to apply the change. Poll it
+      // back up, then refresh the form so secret pills reflect stored state.
+      banner("warn", "Saved — reloading the feeder to apply (~3 s)…");
+      let tries = 6;
+      const poll = async () => {
+        try { await load(); banner("ok", "Reloaded — new config applied."); }
+        catch { if (--tries > 0) setTimeout(poll, 1500); else banner("warn", "Saved & reloading — reopen this panel shortly to confirm."); }
+      };
+      setTimeout(poll, 3000);
+    } else {
+      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) {
+    const msg = (e && e.message) ? e.message : String(e);
+    const friendly = /Failed to fetch|NetworkError|load failed/i.test(msg)
+      ? "couldn't reach the feeder — it may be restarting; wait a moment and try again."
+      : msg;
+    banner("err", "Could not save: " + friendly);
+  } 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 +1,614 @@
+//! 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())
+}
+
+/// Derive a co-bundled sidecar plugin URL from the feeder's own hostname, by the
+/// app-compose convention `<prefix>-feeder` → `<prefix>-<plugin>:8080`
+/// (e.g. `metafeedertorznab-feeder` → `http://metafeedertorznab-tmdb:8080`). Lets
+/// a feeder reach its bundled filename-parser / tmdb sidecars with **no explicit
+/// `*_PLUGIN_URL` env** — config lives in the UI, wiring is derived. Returns
+/// `None` when `HOSTNAME` is unset or not in `<prefix>-feeder` form, in which case
+/// the caller falls back to the env seed (then soft-skips that plugin).
+fn derive_sidecar_url(plugin: &str) -> Option<String> {
+    let host = env_nonempty("HOSTNAME")?;
+    let prefix = host.strip_suffix("-feeder")?;
+    Some(format!("http://{prefix}-{plugin}:8080"))
+}
+
+impl EnrichmentConfig {
+    /// Parse from the feeder env. Returns `None` when `META_CORE_URL` is unset —
+    /// the **bundled-feeder** path (records returned over HTTP, the gateway
+    /// dispatcher stores them). Self-publishing feeders that source meta-core from
+    /// their own dashboard config use [`from_meta_core`](Self::from_meta_core).
+    pub fn from_env() -> Option<Self> {
+        Some(Self::from_meta_core(env_nonempty("META_CORE_URL")?))
+    }
+
+    /// Build with an explicit meta-core root — sourced from the feeder's own
+    /// `config.json` (dashboard) when set, else the `META_CORE_URL` env seed. The
+    /// sidecar URLs come from their `*_PLUGIN_URL` env, falling back to the
+    /// hostname-derived co-bundled sidecar (see [`derive_sidecar_url`]); the rest
+    /// keep their env seeds / defaults. Secrets (tmdb token, language) are
+    /// overlaid by the caller from the per-feeder config.
+    pub fn from_meta_core(meta_core_url: String) -> Self {
+        EnrichmentConfig {
+            meta_core_url,
+            filename_parser_url: env_nonempty("FILENAME_PARSER_URL")
+                .or_else(|| derive_sidecar_url("filename-parser")),
+            tmdb_url: env_nonempty("TMDB_PLUGIN_URL").or_else(|| derive_sidecar_url("tmdb")),
+            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 +1,889 @@
+//! 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))
+}
+
+/// Custom multicodec for a self-describing **Newznab release** locator
+/// (`nzb-release`). MetaMesh-private, adjacent to the torrent-file codecs in the
+/// `0x10xx` range. Unlike a content hash, the cid *embeds* the indexer host +
+/// release id in an identity multihash, so the credentialed meta-share peer
+/// decodes `{host,id}` straight from the cid and grabs the `.nzb` via `t=get`
+/// only at playback — no side-table, no `.nzb` download at search time. (The
+/// `0x1004` `sha256(host‖id)` opaque hash and the `0x1003` `nzb-posting`
+/// variant were both removed; only this self-describing form is emitted.)
+pub const NZB_RELEASE_CODEC: u64 = 0x1005;
+
+/// Digest byte ceiling for an `nzb-release` cid, matching meta-share's
+/// `MAX_MULTIHASH_SIZE` — its `CidGeneric<64>` rejects a longer multihash.
+const NZB_RELEASE_MAX_DIGEST: usize = 64;
+
+/// Encode a self-describing `nzb-release` CID from a Newznab indexer's host +
+/// bare release id (the hex `<guid>` id). The multihash is *identity* (`0x00`)
+/// and its digest embeds the locator as `varint(host_len) ‖ host ‖ id_bytes`,
+/// where `id_bytes` is the hex-decoded id. meta-share's `decode_nzb_release_cid`
+/// is the exact inverse. Host-namespaced: the same release on two indexers
+/// yields different cids.
+///
+/// Returns `None` when the id isn't hex or the digest would overflow the
+/// multihash budget (an oversized host) — the caller drops that row.
+pub fn compute_nzb_release_cid(host: &str, release_id: &str) -> Option<String> {
+    let id_bytes = decode_hex_id(release_id)?;
+    let host_b = host.as_bytes();
+
+    // identity-multihash digest = varint(host_len) ‖ host ‖ id_bytes
+    let mut digest = Vec::with_capacity(2 + host_b.len() + id_bytes.len());
+    write_pb_varint(host_b.len() as u64, &mut digest);
+    digest.extend_from_slice(host_b);
+    digest.extend_from_slice(&id_bytes);
+    if digest.len() > NZB_RELEASE_MAX_DIGEST {
+        return None;
+    }
+
+    // CIDv1: [version=0x01][codec varint][mh code=0x00 identity][len varint][digest]
+    let mut wire = Vec::with_capacity(1 + 3 + 2 + digest.len());
+    wire.push(0x01);
+    write_pb_varint(NZB_RELEASE_CODEC, &mut wire);
+    wire.push(0x00); // multihash code: identity
+    write_pb_varint(digest.len() as u64, &mut wire);
+    wire.extend_from_slice(&digest);
+    Some(format!("b{}", base32_lower_no_padding(&wire)))
+}
+
+/// Hex-decode a Newznab release id. The guid parser only keeps hex runs, so the
+/// id is all-hex by construction; odd-length ids are left-padded with a `0`
+/// nibble. Returns `None` on any non-hex byte (defensive).
+fn decode_hex_id(id: &str) -> Option<Vec<u8>> {
+    let padded;
+    let s: &str = if id.len() % 2 == 1 {
+        padded = format!("0{id}");
+        &padded
+    } else {
+        id
+    };
+    let bytes = s.as_bytes();
+    let mut out = Vec::with_capacity(bytes.len() / 2);
+    for pair in bytes.chunks(2) {
+        let hi = (pair[0] as char).to_digit(16)?;
+        let lo = (pair[1] as char).to_digit(16)?;
+        out.push(((hi << 4) | lo) as u8);
+    }
+    Some(out)
+}
+
+/// Custom multicodec for a **card locator** — the identity of a *work* (a
+/// series, a film) as published by a metadata bridge, rather than of any
+/// bytes. MetaMesh-private, the next free slot after `0x1006` url. (`0x1004`
+/// nzb-release-hash stays **retired — do not reuse**. `0x1003` was retired too,
+/// but has since been **revived** for the self-scanned Usenet posting identity —
+/// see [`NZB_POSTING_CODEC`] and `METADATA_KEYS.md` §2.)
+///
+/// This is the first CID family in the platform where **no bytes exist
+/// anywhere, ever** — `0x1005`/`0x1006` are locators too, but both eventually
+/// resolve to bytes (NNTP, HTTP), and `0x1000` midhash256 is a digest over
+/// real file bytes. A card is pure identity: the record *is* the payload, and
+/// it lives in meta-core. Consequences, all already handled:
+///
+/// * Bitswap seeding is gated on `HashKind::Sha2_256` (gateway invariant 6),
+///   so a card is never offered to the swarm — nothing to seed.
+/// * meta-share's byte path 404s for it, which is correct.
+/// * It ranks `RANK_LOCATOR` (5), so it can never outrank a real digest for
+///   `canonical_cid` — which is what makes the future "one card, two CIDs"
+///   merge (a TMDB and a MyAnimeList locator on one meta-core record) safe.
+pub const CARD_LOCATOR_CODEC: u64 = 0x1007;
+
+/// Digest byte ceiling for a card locator, matching meta-share's
+/// `MAX_MULTIHASH_SIZE` — its `CidGeneric<64>` rejects a longer multihash.
+const CARD_LOCATOR_MAX_DIGEST: usize = 64;
+
+/// Encode a self-describing **card locator** CID from a metadata source and
+/// that source's own id for the work (`("tmdb", "tv:95479")`,
+/// `("mal", "52991")`). The multihash is *identity* (`0x00`) and its digest
+/// embeds the locator as `varint(source_len) ‖ source ‖ id`, both UTF-8 —
+/// exactly the shape [`compute_nzb_release_cid`] uses, minus the hex decode
+/// (a card id is arbitrary text, not a hex release id).
+///
+/// **The CID is a pure function of `(source, id)`**, which is the whole point:
+/// any peer that knows a tmdb id can derive the card's address offline and
+/// look it up locally, with no discovery round-trip. It also means the same
+/// card discovered by two different peers converges on one meta-core record
+/// for free, via the `cids/<bareCid>` reverse index.
+///
+/// Source-namespaced: the same work on TMDB and on MyAnimeList yields
+/// different cids. Merging those two into one record is a meta-core alias
+/// operation driven by an explicit cross-source id mapping — never inferred
+/// here (see `meta-gateway/docs/others/card-tier-search.md` §9).
+///
+/// Returns `None` when `source` is empty or the digest would overflow the
+/// multihash budget — the caller drops that card.
+pub fn compute_card_cid(source: &str, id: &str) -> Option<String> {
+    if source.is_empty() || id.is_empty() {
+        return None;
+    }
+    let source_b = source.as_bytes();
+    let id_b = id.as_bytes();
+
+    // identity-multihash digest = varint(source_len) ‖ source ‖ id
+    let mut digest = Vec::with_capacity(2 + source_b.len() + id_b.len());
+    write_pb_varint(source_b.len() as u64, &mut digest);
+    digest.extend_from_slice(source_b);
+    digest.extend_from_slice(id_b);
+    if digest.len() > CARD_LOCATOR_MAX_DIGEST {
+        return None;
+    }
+
+    // CIDv1: [version=0x01][codec varint][mh code=0x00 identity][len varint][digest]
+    let mut wire = Vec::with_capacity(1 + 3 + 2 + digest.len());
+    wire.push(0x01);
+    write_pb_varint(CARD_LOCATOR_CODEC, &mut wire);
+    wire.push(0x00); // multihash code: identity
+    write_pb_varint(digest.len() as u64, &mut wire);
+    wire.extend_from_slice(&digest);
+    Some(format!("b{}", base32_lower_no_padding(&wire)))
+}
+
+/// Custom multicodec for a **Usenet posting** identity (`nzb-posting`),
+/// minted from the article Message-IDs of a release we scanned ourselves.
+///
+/// Unlike [`NZB_RELEASE_CODEC`] (`0x1005`), this is **not** a locator: it
+/// embeds no indexer host, so any peer with a plain NNTP provider can fetch the
+/// articles — no `IndexerCred`, no `t=get` grab. That portability is the whole
+/// point of running our own header scanner.
+///
+/// It is also **not** fetchable by cid: the digest is over the Message-ID set,
+/// not over any block's bytes, so bitswap can never satisfy a want for it (the
+/// receiver derives a block's cid by hashing the block — see
+/// `beetswap::incoming_stream`). The `.nzb` manifest travels separately, as an
+/// ordinary sha2-256 IPFS cid in the record's `manifest` field. See
+/// `meta-gateway/docs/others/self-hosted-usenet-indexer-study.md` §5.
+pub const NZB_POSTING_CODEC: u64 = 0x1003;
+
+/// Mint an `nzb-posting` cid (`0x1003`) from a release's article Message-IDs.
+///
+/// # The normalisation rule IS the dedup contract
+///
+/// Mirrored verbatim in meta-share (`nzb/manifest.rs`), which re-derives this
+/// digest from the fetched manifest and rejects a mismatch. **Any change here
+/// must land on both sides together** — a drift silently makes every posting
+/// unplayable. The rule:
+///
+/// 1. Trim each id, and strip a surrounding `<…>` pair if present (NZB
+///    `<segment>` bodies carry the bare form, but be defensive — NNTP `BODY`
+///    re-adds the brackets).
+/// 2. Drop empties.
+/// 3. Deduplicate, then **sort** lexicographically by bytes.
+/// 4. Join with `\n` and SHA-256 the result.
+///
+/// Case is **preserved**: a Message-ID's local part is case-sensitive per RFC
+/// 5322, so lowercasing could collide two genuinely distinct articles.
+///
+/// Sorting is what makes the cid survive segment/file reordering between two
+/// NZBs of the same posting — the reason we hash the *id set* and never the raw
+/// `.nzb` bytes (the retired `0x1004` did the latter and broke dedup; see
+/// meta-share `nzb/mod.rs`).
+///
+/// **v1 scope:** every segment counts, par2 volumes included. Two NZBs for one
+/// posting that disagree about whether to bundle par2 therefore mint different
+/// cids. Acceptable while we are the only producer of these manifests; if a
+/// second producer ever appears, the fix is to filter par2 subjects here *and*
+/// in meta-share's mirror — do not do it on one side only.
+pub fn compute_nzb_posting_cid<S: AsRef<str>>(message_ids: &[S]) -> String {
+    let mut ids: Vec<&str> = message_ids
+        .iter()
+        .map(|s| {
+            let t = s.as_ref().trim();
+            t.strip_prefix('<')
+                .and_then(|r| r.strip_suffix('>'))
+                .unwrap_or(t)
+        })
+        .filter(|s| !s.is_empty())
+        .collect();
+    ids.sort_unstable();
+    ids.dedup();
+
+    let mut hasher = Sha256::new();
+    for (i, id) in ids.iter().enumerate() {
+        if i > 0 {
+            hasher.update(b"\n");
+        }
+        hasher.update(id.as_bytes());
+    }
+    let digest: [u8; 32] = hasher.finalize().into();
+
+    // CIDv1: [version=0x01][codec varint][mh code=0x12 sha2-256][len=32][digest]
+    let mut wire = Vec::with_capacity(1 + 3 + 2 + digest.len());
+    wire.push(0x01);
+    write_pb_varint(NZB_POSTING_CODEC, &mut wire);
+    wire.push(0x12); // multihash code: sha2-256
+    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);
+    }
+
+    // ---- card locator (0x1007) --------------------------------------------
+
+    /// Inverse of [`base32_lower_no_padding`], test-only. Lets the card tests
+    /// assert the actual CIDv1 bytes (codec, multihash code, digest) rather
+    /// than just round-tripping the encoder against itself.
+    fn base32_lower_decode(s: &str) -> Vec<u8> {
+        const ALPHABET: &[u8] = b"abcdefghijklmnopqrstuvwxyz234567";
+        let mut out = Vec::new();
+        let (mut acc, mut bits) = (0u32, 0u32);
+        for c in s.bytes() {
+            let v = ALPHABET.iter().position(|&a| a == c).expect("base32 char") as u32;
+            acc = (acc << 5) | v;
+            bits += 5;
+            if bits >= 8 {
+                bits -= 8;
+                out.push((acc >> bits) as u8);
+                acc &= (1 << bits) - 1;
+            }
+        }
+        out
+    }
+
+    /// The property the whole card design leans on: the CID is a pure function
+    /// of `(source, id)`, so any peer can derive a card's address offline. If
+    /// this literal ever changes, every already-published card silently forks
+    /// into a second identity — treat a failure here as a wire break.
+    #[test]
+    fn card_cid_is_deterministic() {
+        let cid = compute_card_cid("tmdb", "tv:95479").expect("card cid");
+        assert_eq!(cid, compute_card_cid("tmdb", "tv:95479").unwrap());
+        assert_eq!(cid, "bagdsaaanar2g2zdcor3duojvgq3ts");
+    }
+
+    /// Source- and kind-namespaced: the same numeric id on a different bridge,
+    /// or the same id under a different media kind, must not collide.
+    #[test]
+    fn card_cid_is_source_and_kind_namespaced() {
+        let tmdb = compute_card_cid("tmdb", "tv:95479").unwrap();
+        let mal = compute_card_cid("mal", "95479").unwrap();
+        let movie = compute_card_cid("tmdb", "movie:95479").unwrap();
+        assert_ne!(tmdb, mal);
+        assert_ne!(tmdb, movie);
+    }
+
+    /// Pins the on-the-wire CIDv1 layout: v1, codec `0x1007` (varint
+    /// `[0x87,0x20]`), **identity** multihash `0x00` (this family has no
+    /// digest — the locator itself is the payload), then
+    /// `varint(source_len) ‖ source ‖ id`.
+    #[test]
+    fn card_cid_wire_shape() {
+        let cid = compute_card_cid("tmdb", "tv:95479").unwrap();
+        assert!(cid.starts_with('b'), "multibase base32-lower prefix");
+        let wire = base32_lower_decode(&cid[1..]);
+
+        assert_eq!(wire[0], 0x01, "CIDv1");
+        assert_eq!(&wire[1..3], &[0x87, 0x20], "codec varint for 0x1007");
+        assert_eq!(wire[3], 0x00, "identity multihash — no digest");
+
+        let digest_len = wire[4] as usize;
+        let digest = &wire[5..5 + digest_len];
+        assert_eq!(digest[0], 4, "varint(len(\"tmdb\"))");
+        assert_eq!(&digest[1..5], b"tmdb");
+        assert_eq!(&digest[5..], b"tv:95479");
+    }
+
+    /// Empty inputs are rejected rather than encoded as an ambiguous locator,
+    /// and an id past the 64-byte multihash budget (meta-share's
+    /// `CidGeneric<64>`) returns `None` so the caller drops the card instead of
+    /// emitting a cid that peer would refuse to parse.
+    #[test]
+    fn card_cid_rejects_empty_and_oversized() {
+        assert!(compute_card_cid("", "tv:1").is_none());
+        assert!(compute_card_cid("tmdb", "").is_none());
+        let huge = "x".repeat(CARD_LOCATOR_MAX_DIGEST);
+        assert!(compute_card_cid("tmdb", &huge).is_none());
+        // Exactly at the ceiling still encodes.
+        let fits = "x".repeat(CARD_LOCATOR_MAX_DIGEST - 5); // varint(4) + "tmdb"
+        assert!(compute_card_cid("tmdb", &fits).is_some());
+    }
+
+    // ---- nzb-posting (0x1003) — the dedup contract -----------------------
+    //
+    // These pin the normalisation rule documented on `compute_nzb_posting_cid`.
+    // meta-share re-derives the same digest from the fetched manifest and
+    // rejects a mismatch, so a change that breaks one of these silently makes
+    // every self-scanned posting unplayable. Mirror any edit on both sides.
+
+    /// Segment/file **order must not matter** — that is the whole reason we
+    /// hash a sorted id set instead of the raw `.nzb` bytes (the retired
+    /// `0x1004` did the latter and broke dedup across reposts).
+    #[test]
+    fn nzb_posting_cid_is_order_independent() {
+        let a = ["[email protected]", "[email protected]", "[email protected]"];
+        let b = ["[email protected]", "[email protected]", "[email protected]"];
+        let c = ["[email protected]", "[email protected]", "[email protected]"];
+        assert_eq!(compute_nzb_posting_cid(&a), compute_nzb_posting_cid(&b));
+        assert_eq!(compute_nzb_posting_cid(&a), compute_nzb_posting_cid(&c));
+    }
+
+    /// Angle brackets, surrounding whitespace, empty entries and duplicates are
+    /// all normalised away. NZB `<segment>` bodies carry the bare form but NNTP
+    /// `BODY` re-adds the brackets, so both spellings must converge.
+    #[test]
+    fn nzb_posting_cid_normalises_brackets_blanks_and_dupes() {
+        let bare = ["[email protected]", "[email protected]"];
+        let noisy = [" <[email protected]> ", "[email protected]", "", "  ", "[email protected]"];
+        assert_eq!(
+            compute_nzb_posting_cid(&bare),
+            compute_nzb_posting_cid(&noisy)
+        );
+    }
+
+    /// Case is preserved: a Message-ID's local part is case-sensitive per RFC
+    /// 5322, so lowercasing could collide two genuinely distinct articles.
+    #[test]
+    fn nzb_posting_cid_is_case_sensitive() {
+        assert_ne!(
+            compute_nzb_posting_cid(&["[email protected]"]),
+            compute_nzb_posting_cid(&["[email protected]"])
+        );
+    }
+
+    /// A different article set is a different posting.
+    #[test]
+    fn nzb_posting_cid_differs_on_different_ids() {
+        assert_ne!(
+            compute_nzb_posting_cid(&["[email protected]", "[email protected]"]),
+            compute_nzb_posting_cid(&["[email protected]", "[email protected]"])
+        );
+    }
+
+    /// Wire shape: CIDv1, codec `0x1003` (varint `[0x83, 0x20]`), multihash
+    /// sha2-256 (`0x12`), 32-byte digest. Pinned because meta-share matches on
+    /// the **codec** slot — a drift here makes every posting undecodable there.
+    #[test]
+    fn nzb_posting_cid_wire_shape() {
+        let cid = compute_nzb_posting_cid(&["[email protected]"]);
+        assert!(cid.starts_with('b'), "multibase base32-lower prefix");
+
+        // Round-trip the base32 body back to bytes to assert the header.
+        let decoded = base32_lower_decode(&cid[1..]);
+        assert_eq!(decoded[0], 0x01, "CIDv1");
+        assert_eq!(&decoded[1..3], &[0x83, 0x20], "varint(0x1003)");
+        assert_eq!(decoded[3], 0x12, "multihash sha2-256");
+        assert_eq!(decoded[4], 0x20, "digest length 32");
+        assert_eq!(decoded.len(), 5 + 32);
+
+        // The digest is sha256 over the newline-joined sorted set.
+        let expect: [u8; 32] = Sha256::digest(b"[email protected]").into();
+        assert_eq!(&decoded[5..], &expect);
+    }
+}

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 +1,49 @@
+//! `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)]
+
+/// pcs-convention default meta-core root a self-publishing feeder targets when
+/// neither its dashboard config nor the `META_CORE_URL` env seed sets one. Keeps
+/// the modular storage seam out of compose: a standard install just works; a
+/// non-standard meta-core is set in the feeder's config UI. Matches the
+/// `metacore-app` service the AppStore feeder apps ship on the shared network.
+pub const DEFAULT_META_CORE_URL: &str = "http://metacore-app:9000";
+
+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 +1,223 @@
+//! 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.
+/// - `NzbRelease` — a Usenet release (a Newznab listing), custom multicodec
+///   `nzb-release` `0x1005`. Self-describing locator: the cid embeds `{host,
+///   id}` in an identity multihash, opaque to bitswap and redeemed to bytes
+///   only by a credentialed meta-share peer, which decodes the locator from the
+///   cid and grabs the `.nzb` via `t=get` — no KV side-table.
+/// - `CardLocator` — the identity of a *work* (a series, a film) as published by
+///   a metadata bridge, custom multicodec `0x1007`. The only family with **no
+///   bytes behind it at all**: the record IS the payload. Always paired with
+///   `bytes: None`, so it lands in the `(None, Some)` metadata-only branch
+///   above. Opaque to bitswap, and deliberately so — there is nothing to seed.
+/// - `NzbPosting` — a Usenet posting we **scanned ourselves**, custom multicodec
+///   `nzb-posting` `0x1003`. A content-derived digest over the article
+///   Message-ID set (not a locator: no host is embedded), so any peer with a
+///   plain NNTP provider can fetch it — no `IndexerCred`. Opaque to bitswap: the
+///   digest is over the id set, not over any block, so the `.nzb` manifest
+///   travels separately as an ordinary sha2-256 cid in the record's `manifest`
+///   field. Produced by `hash::compute_nzb_posting_cid`.
+#[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,
+    NzbRelease,
+    CardLocator,
+    NzbPosting,
+}
+
+/// 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 +1,759 @@
+//! `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,
+    NzbRelease,
+    CardLocator,
+    NzbPosting,
+}
+
+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,
+            HashKind::NzbRelease => HashKindDto::NzbRelease,
+            HashKind::CardLocator => HashKindDto::CardLocator,
+            HashKind::NzbPosting => HashKindDto::NzbPosting,
+        }
+    }
+}
+
+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,
+            HashKindDto::NzbRelease => HashKind::NzbRelease,
+            HashKindDto::CardLocator => HashKind::CardLocator,
+            HashKindDto::NzbPosting => HashKind::NzbPosting,
+        }
+    }
+}
+
+/// 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; scheduling self-restart to apply");
+    // Apply the new config by restarting the process. We can't hot-swap config
+    // in-place: plugins hold an exclusive redb cache handle for their whole life,
+    // so a fresh instance can't reopen the cache while the old one is alive. A
+    // clean process exit is the reliable apply path — Docker's restart policy
+    // (`unless-stopped`) brings the feeder back in ~2-3s and `configure()`
+    // re-reads config.json at startup. Delay briefly so THIS response flushes to
+    // the client before we exit.
+    tokio::spawn(async {
+        tokio::time::sleep(std::time::Duration::from_millis(800)).await;
+        info!(target: "meta-feeder", "restarting now to apply saved config");
+        std::process::exit(0);
+    });
+    Ok(Json(serde_json::json!({ "saved": true, "reloading": 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.
+///
+/// The feeder process never aborts on a config problem: a plugin whose
+/// `configure()` returns [`ConfigError::MissingConfig`] is **soft-skipped**
+/// (warn + dropped — an optional source the operator didn't ask for), and any
+/// other failure (bad/empty key, cache-dir setup, …) registers a
+/// [`DegradedPlugin`] that stays up and reports the reason via `/health` and
+/// `/manifest`. The container keeps running so the error is visible in the UX
+/// instead of crash-looping.
+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)
+}
+
+/// A plugin whose `configure()` (or cache-dir setup) failed. It stays
+/// **registered** so the failure is visible at `GET /health` (as
+/// [`PluginHealth::Degraded`]) and `GET /manifest`; any query/compute against it
+/// returns the captured reason as a permanent error. This is what keeps one
+/// misconfigured feeder from crash-looping the whole container — the operator
+/// sees the reason in the UX and fixes the config, no restart-loop required.
+struct DegradedPlugin {
+    id: &'static str,
+    reason: String,
+    file_types: &'static [&'static str],
+    content_kinds: &'static [&'static str],
+    /// Preserved from the wrapped plugin so a degraded/unconfigured feeder still
+    /// serves its config form — otherwise the dashboard shows "no configuration
+    /// required" and the operator can never fix the very config that's missing.
+    schema: crate::config::ConfigSchema,
+}
+
+#[async_trait::async_trait]
+impl FeederPlugin for DegradedPlugin {
+    fn upstream_id(&self) -> &'static str {
+        self.id
+    }
+
+    fn configure(&mut self, _cache_dir: &FsPath) -> Result<(), ConfigError> {
+        Ok(())
+    }
+
+    async fn handle_query(
+        &self,
+        _query: &GatewayQuery,
+        _max_results: usize,
+    ) -> Result<Vec<DiscoveryRecord>, GatewayError> {
+        Err(GatewayError::Permanent(self.reason.clone()))
+    }
+
+    async fn compute_outcomes(
+        &self,
+        _record_id: &str,
+    ) -> Result<Vec<HashOutcome>, GatewayError> {
+        Err(GatewayError::Permanent(self.reason.clone()))
+    }
+
+    fn health(&self) -> PluginHealth {
+        PluginHealth::Degraded {
+            reason: self.reason.clone(),
+        }
+    }
+
+    fn served_file_types(&self) -> &'static [&'static str] {
+        self.file_types
+    }
+
+    fn served_content_kinds(&self) -> &'static [&'static str] {
+        self.content_kinds
+    }
+
+    fn config_schema(&self) -> crate::config::ConfigSchema {
+        self.schema.clone()
+    }
+}
+
+/// Run each plugin's `configure()` against its per-plugin cache dir and return
+/// the plugins keyed by `upstream_id`. Never errors on a per-plugin problem:
+/// `MissingConfig` is soft-skipped, every other failure becomes a
+/// [`DegradedPlugin`] (see [`serve_feeders`]). 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();
+        // Captured up front (both are `'static`, so this borrow of `plugin`
+        // ends immediately) — lets us build a DegradedPlugin without holding a
+        // borrow across `configure(&mut self)` / the move into `out`.
+        let file_types = plugin.served_file_types();
+        let content_kinds = plugin.served_content_kinds();
+        // Preserve the config surface so a degraded stand-in still serves the
+        // real config form (see DegradedPlugin.schema).
+        let schema = plugin.config_schema();
+        let cache_dir = state_dir.join("gateway").join(id);
+
+        // Build (don't insert) a degraded stand-in: keeps the container up and
+        // surfaces the reason via /health instead of crash-looping.
+        let make_degraded = |reason: String| -> Arc<dyn FeederPlugin> {
+            warn!(
+                target: "meta-feeder",
+                upstream_id = id,
+                %reason,
+                "feeder plugin degraded: container stays up; error surfaced via /health"
+            );
+            Arc::new(DegradedPlugin {
+                id,
+                reason,
+                file_types,
+                content_kinds,
+                schema: schema.clone(),
+            })
+        };
+
+        if let Err(e) = std::fs::create_dir_all(&cache_dir) {
+            let reason = format!("cache dir {} setup failed: {e}", cache_dir.display());
+            out.insert(id.to_string(), make_degraded(reason));
+            continue;
+        }
+
+        match plugin.configure(&cache_dir) {
+            Ok(()) => {
+                out.insert(id.to_string(), Arc::from(plugin));
+            }
+            Err(ConfigError::MissingConfig { plugin: p, what }) => {
+                // Keep a degraded stand-in (don't drop it) so its config form
+                // stays reachable from the dashboard and the operator can supply
+                // the missing config — then restart the feeder to apply it.
+                warn!(
+                    target: "meta-feeder",
+                    upstream_id = %p,
+                    missing = %what,
+                    "feeder plugin degraded: required config not supplied; \
+                     config form still served so it can be fixed from the dashboard"
+                );
+                out.insert(
+                    id.to_string(),
+                    make_degraded(format!("required config not supplied: {what}")),
+                );
+            }
+            Err(other) => {
+                out.insert(id.to_string(), make_degraded(other.to_string()));
+            }
+        }
+    }
+    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/card-feeder/Cargo.toml

@@ -0,0 +1,42 @@
+[package]
+name = "card-feeder"
+version = "0.1.0"
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+publish = false
+
+# Card feeder sidecar: the DISCOVERY half of gateway search. Bridges metadata
+# sources (TMDB today; MyAnimeList later, as a second FeederPlugin in this same
+# binary) into `fileType=card` records — poster + description + an id bag
+# describing a WORK, never a file. It serves no bytes and never touches an
+# indexer: retrieval is a separate, second query answered by the content
+# feeders. See ../../../meta-gateway/docs/others/card-tier-search.md.
+
+[lib]
+name = "card_feeder"
+path = "src/lib.rs"
+
+[[bin]]
+name = "card-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"
+tokio = { version = "1.40", features = ["full", "test-util"] }

feeder-plugin/card-feeder/Dockerfile

@@ -0,0 +1,56 @@
+# syntax=docker/dockerfile:1
+# =============================================================================
+# card-feeder — metadata-bridge feeder sidecar (the card tier). Multi-stage.
+#
+# The build context must be the REPO ROOT (the workspace) so the
+# `meta-feeder-sdk` path dependency (../../crates/meta-feeder-sdk) resolves.
+# Same BuildKit cache-mount strategy as the other feeders: 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 rust-toolchain.toml ./
+COPY Cargo.lock* ./
+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-feeder-card-target,target=/build/target \
+    cargo build --release --bin card-feeder \
+    && cp target/release/card-feeder /card-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 /card-feeder /usr/local/bin/card-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/card-feeder"]

feeder-plugin/card-feeder/src/card.rs

@@ -0,0 +1,342 @@
+//! The card model and its projection to a `DiscoveryRecord`.
+//!
+//! A **card** identifies a *work* — a series, a film — not a file. It carries
+//! the display trio meta-watch's quality gate needs (title, poster,
+//! description) plus an **id bag**: every cross-source identifier the metadata
+//! bridge could resolve. Nothing here is a byte locator; the card's own CID is
+//! a `0x1007` card-locator derived from `(source, id)` and nothing is ever
+//! fetchable by it.
+//!
+//! This is the lifted-and-renamed `ResolvedAnchor` from the indexer feeder's
+//! `torznab/mod.rs`, promoted from a private query-time struct to the feeder's
+//! public output. The projection below is its `build_anchor_record`, minus the
+//! two fields that only made sense while a card was pretending to be a video
+//! (`anchorPlaceholder`, `videoType`) — see
+//! `meta-gateway/docs/others/card-tier-search.md` §2.1.
+
+use std::collections::BTreeMap;
+use std::sync::Arc;
+
+use meta_feeder_sdk::hash::compute_card_cid;
+use meta_feeder_sdk::types::DiscoveryRecord;
+
+use crate::consts::MAX_CARD_SEASONS;
+use crate::tmdb_client::{TmdbClient, TmdbKind, TmdbSeasonSummary};
+
+/// The metadata source that published a card. One variant per bridge; the
+/// string form is the CID's source namespace, so **changing it re-mints every
+/// card CID from that source** — treat these as wire constants.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum CardSource {
+    Tmdb,
+}
+
+impl CardSource {
+    pub(crate) fn as_str(self) -> &'static str {
+        match self {
+            CardSource::Tmdb => "tmdb",
+        }
+    }
+}
+
+/// A resolved work. Immutable by contract: cards carry no rating, popularity,
+/// or any other value that drifts, precisely so the identity and the content
+/// stay in agreement forever (design doc §5). A newly aired season is a *new*
+/// record, never an edit to this one.
+#[derive(Debug, Clone)]
+pub(crate) struct Card {
+    pub(crate) source: CardSource,
+    pub(crate) kind: TmdbKind,
+    /// Source-local id for the work — `"tv:95479"` / `"movie:27205"` for TMDB.
+    /// Half of the card CID's preimage; see [`Card::record_id`].
+    pub(crate) source_id: String,
+
+    // -- the id bag ---------------------------------------------------------
+    // Every cross-source identifier resolved for this work. The phase-2 query
+    // selects the best one for a given indexer's capabilities rather than
+    // assuming `tmdbid`, so a future MyAnimeList-only card needs no new filter
+    // name on the wire (design doc §10b).
+    pub(crate) tmdbid: Option<u64>,
+    /// `i64` to match TMDB's `external_ids` payload verbatim — no lossy cast.
+    pub(crate) tvdb_id: Option<i64>,
+    pub(crate) imdb_id: Option<String>,
+
+    // -- display ------------------------------------------------------------
+    pub(crate) title: String,
+    /// Canonical + original + AKA titles. Rides along so the *content* feeders
+    /// can run their resemblance guard against the card instead of re-querying
+    /// TMDB for it.
+    ///
+    /// Not read here yet — it is consumed by phase 2, which still resolves its
+    /// own anchor until the indexer feeder's `build_anchor_record` path is
+    /// retired. Resolving it now means the card is already complete when that
+    /// switch happens.
+    #[allow(dead_code)]
+    pub(crate) guard_titles: Arc<Vec<String>>,
+    pub(crate) overview: Option<String>,
+    pub(crate) poster_path: Option<String>,
+    pub(crate) year: Option<u16>,
+
+    // -- structure (TV only) ------------------------------------------------
+    pub(crate) seasons: u32,
+    /// Per-season episode counts. Surfaced on the record only as `seasonCount`
+    /// today; the full summaries are what a *season-scoped* phase 2 would fan
+    /// out over. Deliberately descoped for v1 (design doc §12.1) — note that a
+    /// season list is the one thing an immutable card cannot carry forever, so
+    /// when that lands it comes either from season cards or a live lookup, not
+    /// from a mutated series card.
+    #[allow(dead_code)]
+    pub(crate) season_summaries: Arc<Vec<TmdbSeasonSummary>>,
+}
+
+impl Card {
+    /// `contentKind` for this card: a work-level facet, orthogonal to
+    /// `fileType=card`. A whole show is `series` (an individual instalment
+    /// would be `episode`, which a card never is).
+    pub(crate) fn content_kind(&self) -> &'static str {
+        match self.kind {
+            TmdbKind::Tv => "series",
+            TmdbKind::Movie => "movie",
+        }
+    }
+
+    /// The feeder-scoped record id, `"<source>:<source_id>"` — also the exact
+    /// preimage half of the card CID, so `compute_outcomes` can round-trip a
+    /// record id straight back to a CID with no lookup.
+    pub(crate) fn record_id(&self) -> String {
+        format!("{}:{}", self.source.as_str(), self.source_id)
+    }
+
+    /// This card's `0x1007` locator CID. Deterministic in `(source, id)`, so
+    /// any peer holding a tmdb id derives the same address offline.
+    ///
+    /// `compute_outcomes` doesn't use this — it derives the CID straight from
+    /// the record id string (`split_record_id`), deliberately, so deriving an
+    /// address never needs a resolved `Card`. [`Card::to_record`] does use it,
+    /// to stamp the `cids/` key-set member (see there for why that matters).
+    pub(crate) fn cid(&self) -> Option<String> {
+        compute_card_cid(self.source.as_str(), &self.source_id)
+    }
+
+    /// Clamp a TMDB season count into the defensive ceiling.
+    pub(crate) fn clamp_seasons(n: u32) -> u32 {
+        n.clamp(1, MAX_CARD_SEASONS)
+    }
+
+    /// Project to the wire record.
+    ///
+    /// Returns `None` when the card lacks a poster or a description — it would
+    /// fail meta-watch's quality gate and render as a hole in the grid, so the
+    /// feeder declines to emit it at all rather than shipping an unrenderable
+    /// card.
+    ///
+    /// `query_filters` echoes back every structured filter the query carried
+    /// (except `languages`, which fails open on a record that has none) so the
+    /// record survives the gateway's and meta-search's `record_matches`
+    /// re-filter on the way out.
+    pub(crate) fn to_record(
+        &self,
+        tmdb: &TmdbClient,
+        query_filters: &BTreeMap<String, Vec<String>>,
+    ) -> Option<DiscoveryRecord> {
+        let overview = self.overview.as_deref().map(str::trim).filter(|s| !s.is_empty())?;
+        let poster_path = self.poster_path.as_deref().filter(|s| !s.is_empty())?;
+
+        let mut fields: BTreeMap<String, String> = BTreeMap::new();
+        // Type axes. `card` on the fileType axis and the work kind on the
+        // contentKind axis stay orthogonal, so `fileType:card contentKind:series`
+        // is a well-formed query and routing needs no special case.
+        fields.insert("fileType".to_string(), "card".to_string());
+        fields.insert("contentKind".to_string(), self.content_kind().to_string());
+        fields.insert("title".to_string(), self.title.clone());
+
+        // The id bag.
+        if let Some(id) = self.tmdbid {
+            fields.insert("tmdbid".to_string(), id.to_string());
+        }
+        if let Some(id) = self.tvdb_id {
+            fields.insert("tvdbid".to_string(), id.to_string());
+        }
+        if let Some(id) = self.imdb_id.as_deref().filter(|s| !s.is_empty()) {
+            fields.insert("imdbid".to_string(), id.to_string());
+        }
+        if let Some(y) = self.year {
+            fields.insert("movieYear".to_string(), y.to_string());
+        }
+        if self.kind == TmdbKind::Tv && self.seasons > 0 {
+            fields.insert("seasonCount".to_string(), self.seasons.to_string());
+        }
+
+        // Synopsis under the namespaced key meta-search indexes (not a flat
+        // `overview`).
+        fields.insert("description/eng".to_string(), overview.to_string());
+
+        // The bare-CID key-set member, exactly as every other feeder stamps on
+        // its search records (`indexer-feeder`'s `torznab/xml.rs` does the same
+        // with the release's infohash).
+        //
+        // **Load-bearing, and not merely cosmetic.** The gateway's search path
+        // persists a hit to meta-core only if `url_key_for` can find a CID on
+        // it (`dispatch.rs` — no key ⇒ `continue`, silently unpersisted). A card
+        // is the one record type whose CID isn't discovered from an upstream but
+        // *derived*, and `compute_outcomes` (the other place it's derived) is not
+        // on the search path — so without this line a card is never persisted at
+        // all. That interacts badly with the search-coverage gate, which stamps
+        // `(upstream, query)` as covered whenever the feeder produced records and
+        // thereafter skips the feeder for an hour, serving meta-core instead:
+        // coverage marked + nothing persisted = **an identical card search
+        // returns empty for the rest of the window**. Observed live before this
+        // was added.
+        //
+        // Stamping it also delivers §4's addressability and §9.4's free
+        // cross-peer convergence: two peers that resolve the same work publish
+        // the same locator, so meta-core's reverse index folds them into one
+        // record with no merge logic.
+        if let Some(cid) = self.cid() {
+            fields.insert(format!("cids/{cid}"), "true".to_string());
+        }
+
+        // Poster: the transient `posterPath` is rewritten to `poster_url` below,
+        // and the GATEWAY CORE seeds that URL into a content-addressed `poster`
+        // cid (`blockstore::seed_preview`). The feeder never fetches poster
+        // bytes — invariant 10.
+        fields.insert(
+            "poster_url".to_string(),
+            tmdb.poster_cdn_url(poster_path),
+        );
+
+        for (key, allowed) in query_filters {
+            if key == "languages" || fields.contains_key(key) || allowed.is_empty() {
+                continue;
+            }
+            fields.insert(key.clone(), allowed.join(","));
+        }
+
+        Some(DiscoveryRecord {
+            upstream_id: self.source.as_str().to_string(),
+            record_id: self.record_id(),
+            fields,
+        })
+    }
+}
+
+/// Split a card `record_id` (`"tmdb:tv:95479"`) back into `(source, source_id)`
+/// — the two halves of the CID preimage. `None` when the id carries no source
+/// prefix.
+pub(crate) fn split_record_id(record_id: &str) -> Option<(&str, &str)> {
+    record_id.split_once(':').filter(|(s, id)| !s.is_empty() && !id.is_empty())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn tv_card() -> Card {
+        Card {
+            source: CardSource::Tmdb,
+            kind: TmdbKind::Tv,
+            source_id: "tv:95479".to_string(),
+            tmdbid: Some(95479),
+            tvdb_id: Some(367189),
+            imdb_id: Some("tt22248376".to_string()),
+            title: "Frieren: Beyond Journey's End".to_string(),
+            guard_titles: Arc::new(vec!["Sousou no Frieren".to_string()]),
+            overview: Some("The story follows the elf mage Frieren.".to_string()),
+            poster_path: Some("/dqZENchTd7lp5zit1Q7Bkjzcxpi.jpg".to_string()),
+            year: Some(2023),
+            seasons: 1,
+            season_summaries: Arc::new(Vec::new()),
+        }
+    }
+
+    #[test]
+    fn record_id_round_trips_to_the_cid_preimage() {
+        let card = tv_card();
+        let record_id = card.record_id();
+        assert_eq!(record_id, "tmdb:tv:95479");
+        let (source, id) = split_record_id(&record_id).unwrap();
+        assert_eq!((source, id), ("tmdb", "tv:95479"));
+        // The id a peer would derive offline must equal the one we publish.
+        assert_eq!(compute_card_cid(source, id), card.cid());
+    }
+
+    #[test]
+    fn record_carries_both_type_axes_and_the_whole_id_bag() {
+        let tmdb = TmdbClient::new("token".to_string());
+        let rec = tv_card().to_record(&tmdb, &BTreeMap::new()).expect("record");
+        assert_eq!(rec.fields["fileType"], "card");
+        assert_eq!(rec.fields["contentKind"], "series");
+        assert_eq!(rec.fields["tmdbid"], "95479");
+        assert_eq!(rec.fields["tvdbid"], "367189");
+        assert_eq!(rec.fields["imdbid"], "tt22248376");
+        assert!(rec.fields["poster_url"].contains("dqZENchTd7lp5zit1Q7Bkjzcxpi.jpg"));
+        assert!(rec.fields.contains_key("description/eng"));
+        // A card is not a video pretending to have no stream — it is its own
+        // type. The two markers the placeholder needed are gone.
+        assert!(!rec.fields.contains_key("videoType"));
+        assert!(!rec.fields.contains_key("anchorPlaceholder"));
+        // And it must never carry a poster *path* (the transient field the
+        // gateway core would not know how to seed).
+        assert!(!rec.fields.contains_key("posterPath"));
+    }
+
+    /// Regression: the record MUST carry a `cids/` key-set member, because the
+    /// gateway's search-persist path silently drops any record `url_key_for`
+    /// finds no CID on. Without it a card is never persisted, and the
+    /// search-coverage gate then blanks every repeat search for an hour (see
+    /// the comment in `to_record`). This failed live before the fix.
+    #[test]
+    fn record_carries_its_locator_as_a_cids_keyset_member() {
+        let tmdb = TmdbClient::new("token".to_string());
+        let card = tv_card();
+        let rec = card.to_record(&tmdb, &BTreeMap::new()).expect("record");
+        let cid = card.cid().expect("locator");
+        assert_eq!(
+            rec.fields.get(&format!("cids/{cid}")).map(String::as_str),
+            Some("true"),
+            "a card must be persistable; see to_record's cids/ comment"
+        );
+        // And it must be the derived locator, not some other cid shape.
+        assert!(cid.starts_with('b'));
+    }
+
+    /// The quality gate lives here, not in the consumer: a card with no poster
+    /// or no synopsis would render as a hole in meta-watch's grid.
+    #[test]
+    fn card_without_poster_or_overview_is_not_emitted() {
+        let tmdb = TmdbClient::new("token".to_string());
+        let mut no_poster = tv_card();
+        no_poster.poster_path = None;
+        assert!(no_poster.to_record(&tmdb, &BTreeMap::new()).is_none());
+
+        let mut no_overview = tv_card();
+        no_overview.overview = Some("   ".to_string());
+        assert!(no_overview.to_record(&tmdb, &BTreeMap::new()).is_none());
+    }
+
+    #[test]
+    fn movie_card_uses_the_movie_content_kind() {
+        let mut card = tv_card();
+        card.kind = TmdbKind::Movie;
+        card.source_id = "movie:27205".to_string();
+        assert_eq!(card.content_kind(), "movie");
+        assert_eq!(card.record_id(), "tmdb:movie:27205");
+        // Distinct id namespace → distinct CID from the tv card.
+        assert_ne!(card.cid(), tv_card().cid());
+    }
+
+    /// Query filters are echoed so the record survives `record_matches` on the
+    /// way out, but `languages` is deliberately skipped (it fails open).
+    #[test]
+    fn query_filters_are_echoed_except_languages() {
+        let tmdb = TmdbClient::new("token".to_string());
+        let mut filters = BTreeMap::new();
+        filters.insert("fileType".to_string(), vec!["card".to_string()]);
+        filters.insert("genres".to_string(), vec!["fantasy".to_string()]);
+        filters.insert("languages".to_string(), vec!["jpn".to_string()]);
+        let rec = tv_card().to_record(&tmdb, &filters).expect("record");
+        assert_eq!(rec.fields["genres"], "fantasy");
+        assert_eq!(rec.fields["fileType"], "card", "own value wins the echo");
+        assert!(!rec.fields.contains_key("languages"));
+    }
+}

feeder-plugin/card-feeder/src/consts.rs

@@ -0,0 +1,52 @@
+//! Constants for the card feeder. The TMDB subset of the indexer feeder's
+//! `consts.rs` — lifted verbatim so a warm redb cache and the observed live
+//! behaviour carry over unchanged — plus the card-tier's own anchor knobs.
+
+/// User-Agent for outbound HTTP from the card feeder.
+pub(crate) const USER_AGENT: &str = concat!(
+    "meta-card/",
+    env!("CARGO_PKG_VERSION"),
+    " (gateway:card)"
+);
+
+/// TMDB REST API base (v3).
+pub(crate) const TMDB_API_BASE: &str = "https://api.themoviedb.org/3";
+/// TMDB image CDN base; callers append a size segment (e.g. `/w500/<path>`).
+pub(crate) const TMDB_IMAGE_BASE: &str = "https://image.tmdb.org/t/p";
+/// TMDB poster size segment inserted between [`TMDB_IMAGE_BASE`] and the
+/// poster path. Mandatory — TMDB 404s a path with no size. `w500` is the
+/// poster-grid sweet spot (≈122 KB vs `original`'s multi-MB).
+pub(crate) const TMDB_POSTER_SIZE: &str = "w500";
+
+/// TMDB search request timeout (seconds).
+pub(crate) const TMDB_SEARCH_TIMEOUT_SECS: u64 = 1000;
+/// TMDB poster download timeout (seconds). The card feeder never downloads
+/// poster bytes (the gateway core seeds them), but the client is shared with
+/// the indexer feeder's lift, so the knob stays.
+pub(crate) const TMDB_POSTER_TIMEOUT_SECS: u64 = 3000;
+/// Default Retry-After (seconds) applied when TMDB returns HTTP 429 with no
+/// (or an unparseable) `Retry-After` header.
+pub(crate) const TMDB_RATE_LIMIT_DEFAULT_RETRY_SECS: u64 = 5;
+/// Maximum time to wait for a TMDB budget permit before giving up (seconds).
+pub(crate) const TMDB_WAIT_DEADLINE_SECS: u64 = 3000;
+
+/// Default number of cards returned for a free-text query (top-N by TMDB
+/// popularity).
+///
+/// **This is no longer a fan-out multiplier.** In the pre-card design the same
+/// knob (`DEFAULT_ANCHOR_TOP_N`) multiplied the indexer request count — every
+/// anchor became its own paged search per indexer against a shared 1 req/s
+/// bucket, so ten anchors meant the show the user actually wanted got a tenth
+/// of the budget. A card query touches TMDB only, so N here costs at most one
+/// cached `search/multi`; the indexer budget is spent later, entirely on the
+/// single card the user clicks.
+pub(crate) const DEFAULT_CARD_TOP_N: usize = 10;
+
+/// How many ranked principal-search anchors to persist per query in the redb
+/// cache. Sized comfortably above any reasonable top-N so flipping the knob
+/// never needs a cache wipe — reads slice the cached list to the configured N.
+pub(crate) const CACHED_PRINCIPAL_DEPTH: usize = 30;
+
+/// Defensive ceiling on a card's reported season count (guards a corrupt
+/// `number_of_seasons`).
+pub(crate) const MAX_CARD_SEASONS: u32 = 50;

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

@@ -0,0 +1,18 @@
+//! `card-feeder` library surface.
+//!
+//! The **discovery** half of gateway search. Bridges metadata sources into
+//! `fileType=card` records — poster + description + an id bag describing a
+//! *work*, never a file — so the question "which work is this?" is answered
+//! cheaply and separately from "which releases exist for it?".
+//!
+//! One upstream today (`tmdb`). MyAnimeList will be a sibling `FeederPlugin`
+//! in this same binary, declaring the same `served_file_types = ["card"]`.
+//!
+//! Design: `meta-gateway/docs/others/card-tier-search.md`.
+
+pub mod card;
+pub mod consts;
+pub mod resolve;
+pub mod tmdb;
+pub mod tmdb_budget;
+pub mod tmdb_client;

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

@@ -0,0 +1,33 @@
+//! `card-feeder` — metadata-bridge feeder sidecar (the card tier).
+//!
+//! Hosts [`tmdb::TmdbCardPlugin`] 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`)
+//! - `TMDB_TOKEN`              — TMDB v4 read token (seed; the config page wins)
+//! - `TMDB_RATE_PER_SEC`, `TMDB_BURST`, `CARD_TOP_N` — optional seeds
+//! - `RUST_LOG`                — tracing filter (default `info`)
+
+use std::net::SocketAddr;
+
+use card_feeder::tmdb::TmdbCardPlugin;
+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(TmdbCardPlugin::new())], state_dir, listen).await
+}

feeder-plugin/card-feeder/src/resolve.rs

@@ -0,0 +1,331 @@
+//! Cache- and budget-gated TMDB resolution: free text → [`Card`]s.
+//!
+//! A slimmed lift of the indexer feeder's `enrich.rs` `TmdbEnricher`. Only the
+//! **anchor-resolution** half comes across — `principal_top_n`, `tv_details`,
+//! `movie_details`, `external_ids`. The per-release fuzzy title search and its
+//! single-flight map stay in the indexer feeder, which still owns release
+//! enrichment (design doc §8).
+//!
+//! The redb table names are unchanged from the indexer feeder's cache, so a
+//! warm cache carries straight over.
+
+use std::sync::Arc;
+use std::time::Duration;
+
+use meta_feeder_sdk::cache::MidhashCache;
+use tracing::debug;
+
+use crate::card::{Card, CardSource};
+use crate::consts::{CACHED_PRINCIPAL_DEPTH, TMDB_WAIT_DEADLINE_SECS};
+use crate::tmdb_budget::{Lease, TmdbBudget};
+use crate::tmdb_client::{
+    principal_top_n, TmdbCall, TmdbClient, TmdbExternalIds, TmdbHit, TmdbKind, TmdbTvDetails,
+};
+
+/// TMDB resolution with a persistent cache in front and a token budget behind.
+#[derive(Clone)]
+pub(crate) struct Resolver {
+    pub(crate) client: Arc<TmdbClient>,
+    pub(crate) cache: MidhashCache,
+    pub(crate) budget: Arc<TmdbBudget>,
+}
+
+impl Resolver {
+    pub(crate) fn new(client: Arc<TmdbClient>, cache: MidhashCache, budget: Arc<TmdbBudget>) -> Self {
+        Self {
+            client,
+            cache,
+            budget,
+        }
+    }
+
+    /// Acquire a budget token, run `call`, map a hit through `on_hit`. A 429
+    /// pauses the shared bucket globally rather than being retried here.
+    async fn budgeted<T, R>(
+        &self,
+        call: impl std::future::Future<Output = TmdbCall<T>>,
+        on_hit: impl FnOnce(T) -> R,
+    ) -> Option<R> {
+        if matches!(
+            self.budget
+                .acquire(Duration::from_secs(TMDB_WAIT_DEADLINE_SECS))
+                .await,
+            Lease::DeadlineExceeded
+        ) {
+            return None;
+        }
+        match call.await {
+            TmdbCall::Hit(x) => Some(on_hit(x)),
+            TmdbCall::Miss => None,
+            TmdbCall::RateLimited(retry) => {
+                self.budget.note_429(retry);
+                None
+            }
+        }
+    }
+
+    /// Cached `GET /3/tv/{id}`. Entries written before the display/AKA fields
+    /// existed self-heal with one refetch.
+    pub(crate) async fn tv_details(&self, tmdbid: u64) -> Option<TmdbTvDetails> {
+        let key = tmdbid.to_string();
+        if let Ok(Some(json)) = self.cache.get_tmdb_tvdetails(&key) {
+            if let Ok(details) = serde_json::from_str::<TmdbTvDetails>(&json) {
+                if details.has_display() && details.has_akas() {
+                    return Some(details);
+                }
+            }
+        }
+        self.budgeted(self.client.tv_details(tmdbid), |details| {
+            if let Ok(json) = serde_json::to_string(&details) {
+                let _ = self.cache.put_tmdb_tvdetails(&key, &json);
+            }
+            details
+        })
+        .await
+    }
+
+    /// Cached `GET /3/movie/{id}`.
+    pub(crate) async fn movie_hit(&self, tmdbid: u64) -> Option<TmdbHit> {
+        let key = tmdbid.to_string();
+        if let Ok(Some(json)) = self.cache.get_tmdb_moviedetails(&key) {
+            if json != "null" {
+                if let Ok(hit) = serde_json::from_str::<TmdbHit>(&json) {
+                    return Some(hit);
+                }
+            }
+        }
+        self.budgeted(self.client.movie_details(tmdbid), |details| {
+            let hit = details.into_hit();
+            if let Ok(json) = serde_json::to_string(&hit) {
+                let _ = self.cache.put_tmdb_moviedetails(&key, &json);
+            }
+            hit
+        })
+        .await
+    }
+
+    /// Cached `external_ids` — the source of the card's `tvdbid` / `imdbid`.
+    /// External ids are immutable, so a hit is cached forever.
+    pub(crate) async fn external_ids(
+        &self,
+        kind: TmdbKind,
+        tmdbid: u64,
+    ) -> Option<TmdbExternalIds> {
+        let key = tmdbid.to_string();
+        if let Ok(Some(json)) = self.cache.get_tmdb_extids(&key) {
+            return serde_json::from_str(&json).ok();
+        }
+        self.budgeted(self.client.external_ids(kind, tmdbid), |ids| {
+            if let Ok(json) = serde_json::to_string(&ids) {
+                let _ = self.cache.put_tmdb_extids(&key, &json);
+            }
+            ids
+        })
+        .await
+    }
+
+    /// Cached principal `search/multi`: free text → up to `n` confident
+    /// `(tmdbid, kind)` candidates, most popular first.
+    ///
+    /// The full ranked list (up to [`CACHED_PRINCIPAL_DEPTH`]) is persisted and
+    /// sliced on read, so changing the top-N knob never needs a cache wipe.
+    /// A transient TMDB failure is **not** negative-cached — the client folds
+    /// "no results" and "timeout" into the same `Miss`, and caching the latter
+    /// once left a keyword permanently card-less until the next cache wipe.
+    pub(crate) async fn principal_top_n(&self, query: &str, n: usize) -> Vec<(u64, TmdbKind)> {
+        let key = query.trim().to_lowercase();
+        if key.is_empty() || n == 0 {
+            return Vec::new();
+        }
+        if let Ok(Some(json)) = self.cache.get_tmdb_principal_topn(&key) {
+            let mut list = decode_principal_list(&json);
+            list.truncate(n);
+            return list;
+        }
+        if matches!(
+            self.budget
+                .acquire(Duration::from_secs(TMDB_WAIT_DEADLINE_SECS))
+                .await,
+            Lease::DeadlineExceeded
+        ) {
+            return Vec::new(); // transient — don't poison the cache
+        }
+        let ranked: Vec<(u64, TmdbKind)> = match self.client.search_multi(query).await {
+            TmdbCall::Hit(hits) => principal_top_n(&hits, query, CACHED_PRINCIPAL_DEPTH)
+                .into_iter()
+                .filter_map(|h| h.kind().map(|k| (h.id, k)))
+                .collect(),
+            TmdbCall::Miss => return Vec::new(),
+            TmdbCall::RateLimited(retry) => {
+                self.budget.note_429(retry);
+                return Vec::new();
+            }
+        };
+        // An empty list here is a genuine negative ("TMDB answered, nothing
+        // passed the relevance gate") and is safe to remember.
+        let _ = self
+            .cache
+            .put_tmdb_principal_topn(&key, &encode_principal_list(&ranked));
+        let mut out = ranked;
+        out.truncate(n);
+        out
+    }
+
+    /// Resolve one `(tmdbid, kind)` into a full [`Card`] — details plus the
+    /// cross-source id bag. `None` when the by-id lookup fails.
+    pub(crate) async fn card_by_id(&self, tmdbid: u64, kind: TmdbKind) -> Option<Card> {
+        let ext = self.external_ids(kind, tmdbid).await.unwrap_or_default();
+        match kind {
+            TmdbKind::Tv => {
+                let details = self.tv_details(tmdbid).await?;
+                let year = details
+                    .first_air_date
+                    .as_deref()
+                    .and_then(|d| d.get(0..4))
+                    .and_then(|y| y.parse::<u16>().ok());
+                Some(Card {
+                    source: CardSource::Tmdb,
+                    kind,
+                    source_id: format!("tv:{tmdbid}"),
+                    tmdbid: Some(tmdbid),
+                    tvdb_id: ext.tvdb_id,
+                    imdb_id: ext.imdb_id,
+                    title: details.name.clone(),
+                    guard_titles: Arc::new(dedup_titles(details.guard_titles())),
+                    overview: details.overview.clone().filter(|s| !s.is_empty()),
+                    poster_path: details.poster_path.clone().filter(|s| !s.is_empty()),
+                    year,
+                    seasons: Card::clamp_seasons(details.number_of_seasons),
+                    season_summaries: Arc::new(details.seasons.clone()),
+                })
+            }
+            TmdbKind::Movie => {
+                let hit = self.movie_hit(tmdbid).await?;
+                let mut titles = vec![hit.title.clone()];
+                if let Some(o) = &hit.original_title {
+                    titles.push(o.clone());
+                }
+                titles.extend(hit.alt_titles.iter().cloned());
+                Some(Card {
+                    source: CardSource::Tmdb,
+                    kind,
+                    source_id: format!("movie:{tmdbid}"),
+                    tmdbid: Some(tmdbid),
+                    tvdb_id: None,
+                    imdb_id: ext.imdb_id,
+                    title: hit.title,
+                    guard_titles: Arc::new(dedup_titles(titles)),
+                    overview: hit.overview.filter(|s| !s.is_empty()),
+                    poster_path: hit.poster_path.filter(|s| !s.is_empty()),
+                    year: hit.year,
+                    seasons: 0,
+                    season_summaries: Arc::new(Vec::new()),
+                })
+            }
+        }
+    }
+
+    /// Free text → up to `top_n` cards, most popular first.
+    pub(crate) async fn cards_for_text(&self, free_text: &str, top_n: usize) -> Vec<Card> {
+        let mut cards = Vec::new();
+        for (id, kind) in self.principal_top_n(free_text, top_n).await {
+            if let Some(card) = self.card_by_id(id, kind).await {
+                cards.push(card);
+            }
+        }
+        debug!(
+            target: "meta-share::gateway",
+            upstream = "tmdb",
+            query = free_text,
+            count = cards.len(),
+            "resolved cards"
+        );
+        cards
+    }
+}
+
+/// Case-insensitive de-dup preserving first-seen order.
+fn dedup_titles(titles: Vec<String>) -> Vec<String> {
+    let mut seen = std::collections::HashSet::new();
+    titles
+        .into_iter()
+        .map(|t| t.trim().to_string())
+        .filter(|t| !t.is_empty() && seen.insert(t.to_lowercase()))
+        .collect()
+}
+
+/// JSON shape persisted in the principal-search cache. Kept byte-identical to
+/// the indexer feeder's so a warm cache is shared, not re-earned.
+#[derive(serde::Serialize, serde::Deserialize)]
+struct PrincipalEntry {
+    tmdbid: u64,
+    kind: String,
+}
+
+fn kind_str(k: TmdbKind) -> &'static str {
+    match k {
+        TmdbKind::Movie => "movie",
+        TmdbKind::Tv => "tv",
+    }
+}
+
+fn kind_from_str(s: &str) -> Option<TmdbKind> {
+    match s {
+        "movie" => Some(TmdbKind::Movie),
+        "tv" => Some(TmdbKind::Tv),
+        _ => None,
+    }
+}
+
+fn encode_principal_list(list: &[(u64, TmdbKind)]) -> String {
+    let entries: Vec<PrincipalEntry> = list
+        .iter()
+        .map(|&(tmdbid, kind)| PrincipalEntry {
+            tmdbid,
+            kind: kind_str(kind).to_string(),
+        })
+        .collect();
+    serde_json::to_string(&entries).unwrap_or_else(|_| "[]".to_string())
+}
+
+fn decode_principal_list(json: &str) -> Vec<(u64, TmdbKind)> {
+    serde_json::from_str::<Vec<PrincipalEntry>>(json)
+        .unwrap_or_default()
+        .into_iter()
+        .filter_map(|e| kind_from_str(&e.kind).map(|k| (e.tmdbid, k)))
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The cache format must stay byte-compatible with the indexer feeder's —
+    /// the two share redb table names, so a drift here silently invalidates a
+    /// warm cache instead of failing loudly.
+    #[test]
+    fn principal_list_round_trips() {
+        let list = vec![(95479, TmdbKind::Tv), (27205, TmdbKind::Movie)];
+        let json = encode_principal_list(&list);
+        assert_eq!(json, r#"[{"tmdbid":95479,"kind":"tv"},{"tmdbid":27205,"kind":"movie"}]"#);
+        assert_eq!(decode_principal_list(&json), list);
+    }
+
+    #[test]
+    fn principal_list_tolerates_garbage_and_unknown_kinds() {
+        assert!(decode_principal_list("not json").is_empty());
+        assert!(decode_principal_list(r#"[{"tmdbid":1,"kind":"person"}]"#).is_empty());
+        assert!(decode_principal_list("[]").is_empty());
+    }
+
+    #[test]
+    fn dedup_titles_is_case_insensitive_and_order_preserving() {
+        let out = dedup_titles(vec![
+            "Frieren".to_string(),
+            "  frieren  ".to_string(),
+            "Sousou no Frieren".to_string(),
+            "".to_string(),
+        ]);
+        assert_eq!(out, vec!["Frieren", "Sousou no Frieren"]);
+    }
+}

feeder-plugin/card-feeder/src/tmdb.rs

@@ -0,0 +1,408 @@
+//! The `tmdb` upstream — a **card** feeder plugin.
+//!
+//! Answers the discovery half of gateway search: free text in, `fileType=card`
+//! records out. It never calls an indexer and never serves bytes, which is the
+//! whole point — a card query costs at most one cached TMDB `search/multi`, so
+//! the expensive indexer budget is spent later, entirely on the single card the
+//! user clicks (design doc §1.2).
+//!
+//! When MyAnimeList lands it becomes a sibling `FeederPlugin` in this same
+//! binary, declaring the same `served_file_types = ["card"]`.
+
+use std::path::Path;
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use meta_feeder_sdk::cache::MidhashCache;
+use meta_feeder_sdk::config::ConfigSchema;
+use meta_feeder_sdk::plugin::{
+    ConfigError, FeederPlugin, GatewayQuery, HashKind, HashOutcome,
+};
+use meta_feeder_sdk::types::{DiscoveryRecord, GatewayError, Hash, PluginHealth};
+use tracing::warn;
+
+use crate::card::{split_record_id, Card, CardSource};
+use crate::consts::DEFAULT_CARD_TOP_N;
+use crate::resolve::Resolver;
+use crate::tmdb_budget::{TmdbBudget, DEFAULT_TMDB_BURST, DEFAULT_TMDB_REFILL_PER_SEC};
+use crate::tmdb_client::TmdbClient;
+
+/// Config the operator supplies via the feeder's own web form (invariant 12 —
+/// per-plugin config lives on the feeder, never on the gateway).
+#[derive(Clone, Debug, Default, serde::Serialize, serde::Deserialize)]
+pub struct TmdbConfig {
+    #[serde(default)]
+    pub tmdb_token: String,
+    #[serde(default)]
+    pub tmdb_rate_per_sec: Option<f64>,
+    #[serde(default)]
+    pub tmdb_burst: Option<f64>,
+    #[serde(default)]
+    pub card_top_n: Option<usize>,
+}
+
+impl TmdbConfig {
+    /// Env seed, used on first boot before any `config.json` exists.
+    pub fn from_env() -> Self {
+        Self {
+            tmdb_token: std::env::var("TMDB_TOKEN").unwrap_or_default(),
+            tmdb_rate_per_sec: std::env::var("TMDB_RATE_PER_SEC")
+                .ok()
+                .and_then(|v| v.parse().ok()),
+            tmdb_burst: std::env::var("TMDB_BURST").ok().and_then(|v| v.parse().ok()),
+            card_top_n: std::env::var("CARD_TOP_N").ok().and_then(|v| v.parse().ok()),
+        }
+    }
+}
+
+pub struct TmdbCardPlugin {
+    config: TmdbConfig,
+    /// `None` until `configure` runs, and stays `None` when no TMDB token was
+    /// supplied — the soft-skip path. The feeder still serves `/health` so
+    /// `depends_on` is satisfied; queries just return nothing.
+    resolver: Option<Resolver>,
+    top_n: usize,
+    /// Test hook: `(api_base, image_base)` overriding TMDB's real endpoints, so
+    /// the contract test can point the client at a wiremock server.
+    api_bases: Option<(String, String)>,
+}
+
+impl Default for TmdbCardPlugin {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl TmdbCardPlugin {
+    pub fn new() -> Self {
+        Self {
+            config: TmdbConfig::from_env(),
+            resolver: None,
+            top_n: DEFAULT_CARD_TOP_N,
+            api_bases: None,
+        }
+    }
+
+    /// Test constructor: a fixed token plus TMDB endpoint overrides, so the
+    /// contract test can drive the real router against a wiremock TMDB.
+    pub fn with_api_base(token: &str, api_base: String, image_base: String) -> Self {
+        let mut p = Self::new();
+        p.config.tmdb_token = token.to_string();
+        p.api_bases = Some((api_base, image_base));
+        p
+    }
+
+    /// File wins over env, exactly like every other feeder: `config.json` in the
+    /// per-plugin cache dir is read at `configure` time, falling back to the env
+    /// seed. No hot reload — bounce the feeder.
+    fn load_config(&mut self, cache_dir: &Path) {
+        let path = cache_dir.join("config.json");
+        if let Ok(bytes) = std::fs::read(&path) {
+            if let Ok(file_cfg) = serde_json::from_slice::<TmdbConfig>(&bytes) {
+                if !file_cfg.tmdb_token.trim().is_empty() {
+                    self.config.tmdb_token = file_cfg.tmdb_token;
+                }
+                if file_cfg.tmdb_rate_per_sec.is_some() {
+                    self.config.tmdb_rate_per_sec = file_cfg.tmdb_rate_per_sec;
+                }
+                if file_cfg.tmdb_burst.is_some() {
+                    self.config.tmdb_burst = file_cfg.tmdb_burst;
+                }
+                if file_cfg.card_top_n.is_some() {
+                    self.config.card_top_n = file_cfg.card_top_n;
+                }
+            }
+        }
+    }
+
+    fn resolver(&self) -> Option<&Resolver> {
+        self.resolver.as_ref()
+    }
+}
+
+#[async_trait]
+impl FeederPlugin for TmdbCardPlugin {
+    fn upstream_id(&self) -> &'static str {
+        "tmdb"
+    }
+
+    /// **The routing declaration.** `card` on the fileType axis is what makes
+    /// meta-search fan a `fileType:card` query out to this gateway — no
+    /// meta-search change was needed to add the tier (design doc §6). The
+    /// content kinds are the *work* kinds a card can describe.
+    fn served_file_types(&self) -> &'static [&'static str] {
+        &["card"]
+    }
+
+    fn served_content_kinds(&self) -> &'static [&'static str] {
+        &["movie", "series"]
+    }
+
+    fn configure(&mut self, cache_dir: &Path) -> Result<(), ConfigError> {
+        self.load_config(cache_dir);
+        let cache: MidhashCache = meta_feeder_sdk::common::open_midhash_cache(cache_dir, "tmdb")?;
+        self.top_n = self.config.card_top_n.filter(|n| *n > 0).unwrap_or(DEFAULT_CARD_TOP_N);
+
+        let token = self.config.tmdb_token.trim().to_string();
+        if token.is_empty() {
+            // Soft-skip, matching every other opt-in upstream: stay healthy so
+            // the gateway's `depends_on` is satisfied, but answer nothing.
+            warn!(
+                target: "meta-share::gateway",
+                upstream = "tmdb",
+                "no TMDB token configured; card queries will return nothing \
+                 (set it on the feeder's config page or via TMDB_TOKEN)"
+            );
+            return Ok(());
+        }
+        // `TmdbBudget::new` already returns an `Arc`.
+        let budget = TmdbBudget::new(
+            self.config
+                .tmdb_rate_per_sec
+                .filter(|v| *v > 0.0)
+                .unwrap_or(DEFAULT_TMDB_REFILL_PER_SEC),
+            self.config
+                .tmdb_burst
+                .filter(|v| *v > 0.0)
+                .unwrap_or(DEFAULT_TMDB_BURST),
+        );
+        let client = match &self.api_bases {
+            Some((api, image)) => TmdbClient::with_bases(token, api.clone(), image.clone()),
+            None => TmdbClient::new(token),
+        };
+        self.resolver = Some(Resolver::new(Arc::new(client), cache, budget));
+        Ok(())
+    }
+
+    fn health(&self) -> PluginHealth {
+        PluginHealth::Ok
+    }
+
+    async fn handle_query(
+        &self,
+        query: &GatewayQuery,
+        max_results: usize,
+    ) -> Result<Vec<DiscoveryRecord>, GatewayError> {
+        // Layer A early-return: this plugin only ever serves cards.
+        if !meta_feeder_sdk::query_eval::query_accepts_plugin(
+            query,
+            self.served_file_types(),
+            self.served_content_kinds(),
+        ) {
+            return Ok(Vec::new());
+        }
+        let Some(resolver) = self.resolver() else {
+            return Ok(Vec::new()); // soft-skipped (no token)
+        };
+
+        // An explicit `tmdbid:` filter resolves exactly one card — the direct
+        // "give me this work" lookup a deep link uses. Otherwise the free text
+        // goes through the principal search.
+        let cards: Vec<Card> = if let Some(id) = query
+            .filters
+            .get("tmdbid")
+            .and_then(|v| v.first())
+            .and_then(|s| s.trim().parse::<u64>().ok())
+        {
+            let kind = kind_hint(query, resolver, id).await;
+            resolver.card_by_id(id, kind).await.into_iter().collect()
+        } else {
+            let free_text = query.free_text.trim();
+            if free_text.is_empty() {
+                return Ok(Vec::new());
+            }
+            let n = self.top_n.min(max_results.max(1));
+            resolver.cards_for_text(free_text, n).await
+        };
+
+        let client = &resolver.client;
+        Ok(cards
+            .iter()
+            .filter_map(|c| c.to_record(client, &query.filters))
+            .take(max_results)
+            .collect())
+    }
+
+    /// Resolve a card `record_id` back into its locator CID.
+    ///
+    /// This is **not** a stub for a byte-less plugin — it is the card path. The
+    /// outcome carries `bytes: None`, so the gateway core's three-branch
+    /// auto-store routes it to `store_metadata_only`: the record lands in
+    /// meta-core, nothing is written to WebDAV, and nothing is seeded to
+    /// bitswap (seeding is gated on `Sha2_256`). The record IS the payload.
+    ///
+    /// The CID needs no network call — it is a pure function of the record id,
+    /// which is itself `"<source>:<source_id>"`. We re-resolve the card anyway
+    /// so the stored record is complete; a resolution failure still yields the
+    /// CID, because the identity is derivable regardless.
+    async fn compute_outcomes(&self, record_id: &str) -> Result<Vec<HashOutcome>, GatewayError> {
+        let (source, source_id) = split_record_id(record_id).ok_or_else(|| {
+            GatewayError::Permanent(format!("malformed card record_id '{record_id}'"))
+        })?;
+        if source != CardSource::Tmdb.as_str() {
+            return Err(GatewayError::NotFound);
+        }
+        let hash = meta_feeder_sdk::hash::compute_card_cid(source, source_id).ok_or_else(|| {
+            GatewayError::Permanent(format!("card id too long to encode: '{record_id}'"))
+        })?;
+
+        let record = match (self.resolver(), parse_source_id(source_id)) {
+            (Some(resolver), Some((kind, tmdbid))) => resolver
+                .card_by_id(tmdbid, kind)
+                .await
+                .and_then(|c| c.to_record(&resolver.client, &Default::default())),
+            _ => None,
+        };
+
+        Ok(vec![HashOutcome {
+            hash: Hash(hash),
+            hash_kind: HashKind::CardLocator,
+            bytes: None,
+            record,
+            file_extension: None,
+        }])
+    }
+
+    fn config_schema(&self) -> ConfigSchema {
+        use meta_feeder_sdk::config::{ConfigField as F, ConfigSchema};
+        ConfigSchema {
+            fields: vec![
+                F::secret("tmdb_token", "TMDB v4 token").with_help(
+                    "Read-access token from themoviedb.org. Without it this feeder \
+                     soft-skips and card searches return nothing.",
+                ),
+                F::number("tmdb_rate_per_sec", "TMDB rate (req/s)").with_help(
+                    "Sustained TMDB request rate. Blank keeps the built-in default \
+                     (20/s). Takes effect on the next feeder restart.",
+                ),
+                F::number("tmdb_burst", "TMDB burst (tokens)").with_help(
+                    "Short-burst TMDB token ceiling. Blank keeps the built-in \
+                     default (20). Takes effect on the next feeder restart.",
+                ),
+                F::number("card_top_n", "Cards per search").with_help(
+                    "How many works a free-text search returns, most popular first. \
+                     Blank keeps the built-in default (10). Unlike the old anchor \
+                     top-N this costs no indexer requests — only TMDB.",
+                ),
+            ],
+        }
+    }
+
+    fn config_values(&self) -> serde_json::Value {
+        serde_json::json!({
+            "tmdb_token": self.config.tmdb_token,
+            "tmdb_rate_per_sec": self.config.tmdb_rate_per_sec,
+            "tmdb_burst": self.config.tmdb_burst,
+            "card_top_n": self.config.card_top_n,
+        })
+    }
+}
+
+/// Decide whether an explicit `tmdbid:` refers to a show or a film: an explicit
+/// `contentKind` filter wins, else probe TV details (a movie id has none).
+async fn kind_hint(query: &GatewayQuery, resolver: &Resolver, id: u64) -> crate::tmdb_client::TmdbKind {
+    use crate::tmdb_client::TmdbKind;
+    match query
+        .filters
+        .get("contentKind")
+        .and_then(|v| v.first())
+        .map(|s| s.as_str())
+    {
+        Some("movie") => return TmdbKind::Movie,
+        Some("series") | Some("episode") | Some("tvshow") => return TmdbKind::Tv,
+        _ => {}
+    }
+    match resolver.tv_details(id).await {
+        Some(d) if d.number_of_seasons > 0 => TmdbKind::Tv,
+        _ => TmdbKind::Movie,
+    }
+}
+
+/// `"tv:95479"` / `"movie:27205"` → `(kind, tmdbid)`.
+fn parse_source_id(source_id: &str) -> Option<(crate::tmdb_client::TmdbKind, u64)> {
+    use crate::tmdb_client::TmdbKind;
+    let (kind, id) = source_id.split_once(':')?;
+    let kind = match kind {
+        "tv" => TmdbKind::Tv,
+        "movie" => TmdbKind::Movie,
+        _ => return None,
+    };
+    Some((kind, id.trim().parse().ok()?))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::tmdb_client::TmdbKind;
+
+    #[test]
+    fn source_id_parses_both_kinds() {
+        assert!(matches!(
+            parse_source_id("tv:95479"),
+            Some((TmdbKind::Tv, 95479))
+        ));
+        assert!(matches!(
+            parse_source_id("movie:27205"),
+            Some((TmdbKind::Movie, 27205))
+        ));
+        assert!(parse_source_id("person:123").is_none());
+        assert!(parse_source_id("tv:notanumber").is_none());
+        assert!(parse_source_id("95479").is_none());
+    }
+
+    /// The routing declaration is load-bearing: `card` here is the only reason
+    /// meta-search fans a `fileType:card` query out to this gateway.
+    #[test]
+    fn declares_the_card_file_type() {
+        let p = TmdbCardPlugin::new();
+        assert_eq!(p.served_file_types(), &["card"]);
+        assert_eq!(p.served_content_kinds(), &["movie", "series"]);
+        assert_eq!(p.upstream_id(), "tmdb");
+    }
+
+    /// A card CID is derivable with no network and no resolver, so
+    /// `compute_outcomes` must still produce one when the feeder is
+    /// soft-skipped — only the record is missing.
+    #[tokio::test]
+    async fn compute_outcomes_yields_a_byteless_locator_without_a_token() {
+        let p = TmdbCardPlugin::new();
+        let outcomes = p.compute_outcomes("tmdb:tv:95479").await.expect("outcomes");
+        assert_eq!(outcomes.len(), 1);
+        let o = &outcomes[0];
+        assert_eq!(o.hash_kind, HashKind::CardLocator);
+        assert!(o.bytes.is_none(), "a card has no bytes, ever");
+        assert!(o.file_extension.is_none());
+        assert_eq!(
+            o.hash.0,
+            meta_feeder_sdk::hash::compute_card_cid("tmdb", "tv:95479").unwrap()
+        );
+    }
+
+    #[tokio::test]
+    async fn compute_outcomes_rejects_a_foreign_or_malformed_id() {
+        let p = TmdbCardPlugin::new();
+        assert!(p.compute_outcomes("mal:52991").await.is_err());
+        assert!(p.compute_outcomes("nosource").await.is_err());
+    }
+
+    /// Soft-skip: no token means no cards, not a hard failure — the feeder
+    /// stays healthy so the gateway's `depends_on` is satisfied.
+    #[tokio::test]
+    async fn queries_return_nothing_without_a_token() {
+        let p = TmdbCardPlugin::new();
+        let q = GatewayQuery::from_free_text("frieren");
+        assert!(p.handle_query(&q, 10).await.unwrap().is_empty());
+        assert!(matches!(p.health(), PluginHealth::Ok));
+    }
+
+    /// Layer-A routing: a query for something this plugin cannot serve returns
+    /// immediately without touching TMDB.
+    #[tokio::test]
+    async fn rejects_a_query_for_a_type_it_does_not_serve() {
+        let p = TmdbCardPlugin::new();
+        let mut q = GatewayQuery::from_free_text("frieren");
+        q.filters
+            .insert("fileType".to_string(), vec!["video".to_string()]);
+        assert!(p.handle_query(&q, 10).await.unwrap().is_empty());
+    }
+}

feeder-plugin/card-feeder/src/tmdb_budget.rs

@@ -0,0 +1,328 @@
+//! Process-global TMDB request budget.
+//!
+//! TMDB's free tier caps at ~40 requests / 10 s. The gateway can field many
+//! concurrent searches, each wanting to enrich several records, so without a
+//! shared throttle a burst of discover queries collectively blows the limit
+//! and TMDB starts returning 429s — which surfaces as empty home rows in
+//! meta-watch. This is a single hand-rolled token bucket shared by every
+//! enrichment task in the process (one `Arc<TmdbBudget>` injected into the
+//! torznab plugin via [`meta_feeder_sdk::plugin::FeederPlugin::set_tmdb_budget`]).
+//!
+//! Design notes:
+//! - **Lazy refill.** Tokens accrue on each `acquire` from elapsed wall-clock,
+//!   so there is no background timer task and no idle cost.
+//! - **Cancel-safe.** The token is decremented synchronously under the lock
+//!   right before `acquire` returns [`Lease::Granted`]; there is no `.await`
+//!   between the decision to grant and the return, so a dropped `acquire`
+//!   future (consumer disconnected) never spends a token.
+//! - **Global 429 pause.** A single upstream 429 freezes *all* grants until
+//!   the Retry-After window elapses via [`TmdbBudget::note_429`].
+//!
+//! The critical sections hold a plain `std::sync::Mutex` (no awaits inside),
+//! while the waiting happens outside the lock on a `tokio::time::sleep` raced
+//! against a `tokio::sync::Notify`.
+
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+
+use tokio::sync::Notify;
+
+/// Default sustained TMDB grant rate (requests/second). TMDB removed its old
+/// 40-req/10-s hard limit years ago (now effectively ~50 req/s, "be reasonable"),
+/// and the `tmdb_inflight` single-flight dedup already collapses same-show lookups
+/// to one search — so a conservative 20/s is well under the ceiling while removing
+/// the enrichment serializer that made the first poster land ~25 s late. Operator-
+/// overridable via the `tmdb_rate_per_sec` feeder web-config field.
+pub const DEFAULT_TMDB_REFILL_PER_SEC: f64 = 20.0;
+/// Default TMDB burst ceiling (tokens) — a short burst of concurrent searches
+/// is served immediately, then throttled to the sustained rate. Overridable via
+/// the `tmdb_burst` feeder web-config field.
+pub const DEFAULT_TMDB_BURST: f64 = 20.0;
+
+/// Outcome of an [`TmdbBudget::acquire`] call.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Lease {
+    /// A token was granted; the caller may make one TMDB API call.
+    Granted,
+    /// The wait deadline elapsed before a token became available. The caller
+    /// should proceed without enrichment (best-effort degrade).
+    DeadlineExceeded,
+}
+
+struct BucketState {
+    tokens: f64,
+    capacity: f64,
+    refill_per_sec: f64,
+    last_refill: Instant,
+    /// When set and in the future, all grants are frozen (global 429 pause).
+    paused_until: Option<Instant>,
+}
+
+impl BucketState {
+    /// Accrue tokens for elapsed time since the last refill. Called under the
+    /// lock before inspecting `tokens`.
+    fn refill(&mut self, now: Instant) {
+        let elapsed = now
+            .saturating_duration_since(self.last_refill)
+            .as_secs_f64();
+        if elapsed > 0.0 {
+            self.tokens = (self.tokens + elapsed * self.refill_per_sec).min(self.capacity);
+            self.last_refill = now;
+        }
+    }
+}
+
+/// Shared token bucket gating TMDB API calls across the whole process.
+pub struct TmdbBudget {
+    state: Mutex<BucketState>,
+    /// Woken whenever a grant *might* now be possible (a token was returned to
+    /// the bucket, or a 429 pause was cleared). Waiters re-check under the
+    /// lock after waking; a missed wake only costs an extra bounded sleep.
+    wake: Notify,
+}
+
+impl TmdbBudget {
+    /// Build a budget that sustains `refill_per_sec` grants/second with a burst
+    /// ceiling of `capacity` tokens. The bucket starts full so the first burst
+    /// is served immediately.
+    pub fn new(refill_per_sec: f64, capacity: f64) -> Arc<Self> {
+        Arc::new(Self {
+            state: Mutex::new(BucketState {
+                tokens: capacity,
+                capacity,
+                refill_per_sec,
+                last_refill: Instant::now(),
+                paused_until: None,
+            }),
+            wake: Notify::new(),
+        })
+    }
+
+    /// Wait for one token, giving up after `deadline`. Cancel-safe: dropping
+    /// the returned future before it resolves [`Lease::Granted`] spends no
+    /// token.
+    pub async fn acquire(&self, deadline: Duration) -> Lease {
+        let give_up_at = Instant::now() + deadline;
+        loop {
+            // Decide-or-compute-wait, all under the lock (no awaits inside).
+            let sleep_for = {
+                let mut st = self.state.lock().expect("tmdb budget mutex poisoned");
+                let now = Instant::now();
+                if let Some(until) = st.paused_until {
+                    if now >= until {
+                        st.paused_until = None;
+                    }
+                }
+                let paused = st.paused_until;
+                if paused.is_none() {
+                    st.refill(now);
+                    if st.tokens >= 1.0 {
+                        // Atomic grant: decrement and return with no await between.
+                        st.tokens -= 1.0;
+                        return Lease::Granted;
+                    }
+                }
+                if now >= give_up_at {
+                    return Lease::DeadlineExceeded;
+                }
+                let until_deadline = give_up_at.saturating_duration_since(now);
+                let wait = match paused {
+                    // Frozen: wake when the pause is scheduled to lift.
+                    Some(until) => until.saturating_duration_since(now),
+                    // Throttled: wake when the next whole token should accrue.
+                    None => {
+                        let needed = 1.0 - st.tokens;
+                        let secs = if st.refill_per_sec > 0.0 {
+                            needed / st.refill_per_sec
+                        } else {
+                            until_deadline.as_secs_f64()
+                        };
+                        Duration::from_secs_f64(secs.max(0.0))
+                    }
+                };
+                wait.min(until_deadline)
+            };
+            tokio::select! {
+                _ = tokio::time::sleep(sleep_for) => {}
+                _ = self.wake.notified() => {}
+            }
+            if Instant::now() >= give_up_at {
+                return Lease::DeadlineExceeded;
+            }
+        }
+    }
+
+    /// Feed back an upstream 429: pause all grants until `now + retry_after`.
+    /// Extends an existing pause but never shortens it.
+    pub fn note_429(&self, retry_after: Duration) {
+        let until = Instant::now() + retry_after;
+        {
+            let mut st = self.state.lock().expect("tmdb budget mutex poisoned");
+            st.paused_until = Some(match st.paused_until {
+                Some(existing) if existing > until => existing,
+                _ => until,
+            });
+        }
+        // Wake waiters so they recompute their sleep against the new pause.
+        self.wake.notify_waiters();
+    }
+}
+
+/// Per-indexer registry of [`TmdbBudget`] token buckets.
+///
+/// Replaces the single global indexer limiter: each indexer is rate-limited
+/// independently, so a fast Usenet indexer isn't throttled to a slow torrent
+/// tracker's cadence, and N indexers fetched concurrently give ~N× aggregate
+/// throughput. A 429/reset on one indexer (via [`TmdbBudget::note_429`] on its
+/// own bucket) pauses only that indexer. Buckets are created lazily, keyed by
+/// `IndexerConfig.name` (unique per tracker after `effective_indexers`
+/// expansion). Shared by `Arc` across every concurrent search's plugin clone, so
+/// the per-indexer rate is real — all in-flight searches contend on the same
+/// bucket for a given indexer.
+pub struct IndexerLimiterRegistry {
+    inner: Mutex<RegistryInner>,
+}
+
+struct RegistryInner {
+    buckets: HashMap<String, Arc<TmdbBudget>>,
+    default_rps: f64,
+    default_burst: f64,
+}
+
+impl IndexerLimiterRegistry {
+    /// Build a registry that seeds each lazily-created bucket at
+    /// `default_rps` / `default_burst`.
+    pub fn new(default_rps: f64, default_burst: f64) -> Arc<Self> {
+        Arc::new(Self {
+            inner: Mutex::new(RegistryInner {
+                buckets: HashMap::new(),
+                default_rps,
+                default_burst,
+            }),
+        })
+    }
+
+    /// Override the default per-indexer rate (the `indexer_rate_per_sec`
+    /// web-config field). Applied at `configure()` time — before any bucket is
+    /// created — and clears existing buckets so a reconfigure re-seeds at the new
+    /// rate. Non-positive `rps` is ignored (keeps the compiled default).
+    pub fn set_default_rps(&self, rps: f64) {
+        if rps > 0.0 {
+            let mut inner = self.inner.lock().expect("indexer limiter registry poisoned");
+            inner.default_rps = rps;
+            inner.buckets.clear();
+        }
+    }
+
+    /// The token bucket for `indexer_name`, creating it on first use.
+    pub fn for_indexer(&self, indexer_name: &str) -> Arc<TmdbBudget> {
+        let mut inner = self.inner.lock().expect("indexer limiter registry poisoned");
+        if let Some(bucket) = inner.buckets.get(indexer_name) {
+            return bucket.clone();
+        }
+        let bucket = TmdbBudget::new(inner.default_rps, inner.default_burst);
+        inner
+            .buckets
+            .insert(indexer_name.to_string(), bucket.clone());
+        bucket
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test(start_paused = true)]
+    async fn grants_up_to_capacity_immediately() {
+        let b = TmdbBudget::new(3.0, 3.0);
+        // Three tokens available at t0 — all granted without waiting.
+        for _ in 0..3 {
+            assert_eq!(b.acquire(Duration::from_secs(10)).await, Lease::Granted);
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn fourth_token_waits_for_refill() {
+        let b = TmdbBudget::new(2.0, 2.0); // 2/s sustained
+        for _ in 0..2 {
+            assert_eq!(b.acquire(Duration::from_secs(10)).await, Lease::Granted);
+        }
+        // Bucket empty; next grant must wait ~0.5s for one token at 2/s.
+        let start = Instant::now();
+        assert_eq!(b.acquire(Duration::from_secs(10)).await, Lease::Granted);
+        let waited = start.elapsed();
+        assert!(
+            waited >= Duration::from_millis(450),
+            "expected to wait ~0.5s for refill, waited {waited:?}"
+        );
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn deadline_exceeded_when_drained() {
+        let b = TmdbBudget::new(0.001, 1.0); // effectively no refill within the test
+        assert_eq!(b.acquire(Duration::from_secs(1)).await, Lease::Granted);
+        // Bucket empty, refill far slower than the deadline.
+        assert_eq!(
+            b.acquire(Duration::from_millis(200)).await,
+            Lease::DeadlineExceeded
+        );
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn note_429_pauses_grants_globally() {
+        let b = TmdbBudget::new(100.0, 100.0); // plenty of tokens
+        b.note_429(Duration::from_secs(2));
+        // Even with tokens available, grants are frozen during the pause —
+        // a short deadline times out.
+        assert_eq!(
+            b.acquire(Duration::from_millis(500)).await,
+            Lease::DeadlineExceeded
+        );
+        // After the pause window, grants resume.
+        assert_eq!(b.acquire(Duration::from_secs(5)).await, Lease::Granted);
+    }
+
+    #[test]
+    fn registry_lazily_keys_by_indexer_name() {
+        let reg = IndexerLimiterRegistry::new(1.0, 3.0);
+        let a1 = reg.for_indexer("nzb.life");
+        let a2 = reg.for_indexer("nzb.life");
+        let b = reg.for_indexer("nyaa.si");
+        // Same name → same bucket instance; different name → distinct bucket.
+        assert!(Arc::ptr_eq(&a1, &a2));
+        assert!(!Arc::ptr_eq(&a1, &b));
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn per_indexer_429_does_not_pause_other_indexers() {
+        let reg = IndexerLimiterRegistry::new(1.0, 3.0);
+        let hot = reg.for_indexer("hot");
+        let cold = reg.for_indexer("cold");
+        // A 429 on `hot` pauses only `hot`.
+        hot.note_429(Duration::from_secs(5));
+        assert_eq!(
+            hot.acquire(Duration::from_millis(200)).await,
+            Lease::DeadlineExceeded
+        );
+        // `cold` is untouched — its bucket starts full and grants immediately.
+        assert_eq!(
+            cold.acquire(Duration::from_millis(200)).await,
+            Lease::Granted
+        );
+    }
+
+    #[test]
+    fn registry_set_default_rps_affects_new_buckets_only() {
+        let reg = IndexerLimiterRegistry::new(1.0, 1.0);
+        let _early = reg.for_indexer("early"); // created at 1.0
+        reg.set_default_rps(50.0); // clears buckets; new default 50/s
+                                   // Non-positive is ignored (keeps 50).
+        reg.set_default_rps(0.0);
+        // A freshly created bucket uses the updated default (burst 1 → one
+        // immediate grant, then throttled at 50/s).
+        let b = reg.for_indexer("late");
+        // Smoke: the bucket exists and is usable (full-bucket first grant).
+        assert!(Arc::ptr_eq(&b, &reg.for_indexer("late")));
+    }
+}

feeder-plugin/card-feeder/src/tmdb_client.rs

@@ -0,0 +1,978 @@
+//! TMDB HTTP client + DTOs + season/episode bounds check.
+//!
+//! **Verbatim lift** of the indexer feeder's `tmdb.rs` (which was itself split
+//! out of the monolithic `torznab.rs`). Kept byte-compatible with that copy on
+//! purpose: TMDB's API handling is fiddly — 429/`Retry-After` semantics, the
+//! self-healing cache-shape checks, the AKA append — and a fix on either side
+//! should port across as a straight copy. The redb cache table names match for
+//! the same reason, so a warm cache is shared rather than re-earned.
+//!
+//! Consequence: this file carries surface the **card feeder does not use**.
+//! `search` / `discovery_list` (fuzzy title search, home rows),
+//! `season_episode_bounds` / `episode_in_season` (release season/episode
+//! reconciliation) and the language helpers all belong to *retrieval*, which
+//! stays in the indexer feeder. They are dead here by design, not by neglect —
+//! hence the module-wide allow rather than a trim that would fork the file.
+#![allow(dead_code)]
+
+use meta_feeder_sdk::common::urlencode;
+use std::time::Duration;
+
+use tracing::debug;
+
+// -- TMDB metadata enrichment ------------------------------------------------
+//
+// TMDB v4 bearer-token client. Two endpoints used:
+//
+//   GET /3/search/{movie,tv}?query=<urlencoded>&include_adult=false[&year=<YYYY>]
+//   GET <image_base>/<poster_path>
+//
+// The token-bearing `Authorization` header gates both. Free-tier rate
+// limit is ~40 req / 10 s — we don't hit it from a dev box but it's
+// worth knowing. All TMDB failures degrade silently: enrichment is a
+// best-effort augmentation, never a hard dep.
+
+/// Lightweight TMDB v4 client. Cheap to clone (the underlying
+/// `reqwest::Client` is `Arc`-backed).
+#[derive(Clone)]
+pub struct TmdbClient {
+    http: reqwest::Client,
+    bearer_token: String,
+    api_base: String,
+    image_base: String,
+}
+
+impl TmdbClient {
+    pub fn new(bearer_token: String) -> Self {
+        let http = reqwest::Client::builder()
+            .timeout(Duration::from_secs(TMDB_POSTER_TIMEOUT_SECS))
+            .user_agent(USER_AGENT)
+            .build()
+            .expect("rustls reqwest client build infallible");
+        Self {
+            http,
+            bearer_token,
+            api_base: TMDB_API_BASE.to_string(),
+            image_base: TMDB_IMAGE_BASE.to_string(),
+        }
+    }
+
+    /// **The one deviation from the verbatim lift**: the indexer feeder gates
+    /// this on `#[cfg(test)]`. The card feeder's contract test is an
+    /// *integration* test, which compiles the lib without `cfg(test)`, so it
+    /// needs a real constructor to point the client at a wiremock TMDB.
+    pub fn with_bases(token: String, api_base: String, image_base: String) -> Self {
+        let http = reqwest::Client::builder()
+            .timeout(Duration::from_secs(TMDB_POSTER_TIMEOUT_SECS))
+            .user_agent(USER_AGENT)
+            .build()
+            .expect("rustls reqwest client build infallible");
+        Self {
+            http,
+            bearer_token: token,
+            api_base,
+            image_base,
+        }
+    }
+
+    /// Search TMDB for a movie/TV title. Returns [`TmdbCall::Hit`] with the
+    /// top result, [`TmdbCall::Miss`] when there are no matches (or any
+    /// transient/timeout error — best-effort degrade), or
+    /// [`TmdbCall::RateLimited`] carrying the Retry-After window on a 429 so
+    /// the caller can pause the shared budget globally. Wraps the whole call
+    /// (including JSON decode) in [`TMDB_SEARCH_TIMEOUT_SECS`].
+    pub(crate) async fn search(
+        &self,
+        kind: TmdbKind,
+        title: &str,
+        year: Option<u16>,
+    ) -> TmdbCall<TmdbHit> {
+        let endpoint = match kind {
+            TmdbKind::Movie => "search/movie",
+            TmdbKind::Tv => "search/tv",
+        };
+        let mut url = format!(
+            "{}/{}?query={}&include_adult=false",
+            self.api_base.trim_end_matches('/'),
+            endpoint,
+            urlencode(title),
+        );
+        if let Some(y) = year {
+            // Movie uses `year`, TV uses `first_air_date_year`; both
+            // are documented but movies accept `year` for backward-
+            // compat which is what we want.
+            url.push_str(&format!("&year={y}"));
+        }
+        let fut = self
+            .http
+            .get(&url)
+            .bearer_auth(&self.bearer_token)
+            .header("accept", "application/json")
+            .send();
+        let resp =
+            match tokio::time::timeout(Duration::from_secs(TMDB_SEARCH_TIMEOUT_SECS), fut).await {
+                Ok(Ok(r)) => r,
+                _ => return TmdbCall::Miss,
+            };
+        if resp.status().as_u16() == 429 {
+            return TmdbCall::RateLimited(parse_tmdb_retry_after(&resp));
+        }
+        if !resp.status().is_success() {
+            debug!(
+                target: "meta-share::gateway",
+                upstream = "prowlarr",
+                tmdb_url = %url,
+                status = %resp.status(),
+                "tmdb search non-2xx; degrading"
+            );
+            return TmdbCall::Miss;
+        }
+        let body: TmdbSearchResponse =
+            match tokio::time::timeout(Duration::from_secs(TMDB_SEARCH_TIMEOUT_SECS), resp.json())
+                .await
+            {
+                Ok(Ok(b)) => b,
+                _ => return TmdbCall::Miss,
+            };
+        match body.results.into_iter().next() {
+            Some(top) => TmdbCall::Hit(top.into_hit(kind)),
+            None => TmdbCall::Miss,
+        }
+    }
+
+    /// Build the public TMDB poster CDN URL for `poster_path` (no fetch). In the
+    /// feeder model the feeder no longer fetches+stores the poster bytes; it
+    /// emits this URL as a `poster_url` field and the **gateway core** seeds it
+    /// into a content-addressed `poster` cid (same path as giphy/wikicommons
+    /// previews). Inserts the required [`TMDB_POSTER_SIZE`] segment.
+    pub(crate) fn poster_cdn_url(&self, poster_path: &str) -> String {
+        format!(
+            "{}/{}/{}",
+            self.image_base.trim_end_matches('/'),
+            TMDB_POSTER_SIZE,
+            poster_path.trim_start_matches('/'),
+        )
+    }
+
+    /// Fetch a poster image's raw bytes. `poster_path` is the
+    /// TMDB-supplied path (typically `"/abcdef.jpg"` with leading slash).
+    ///
+    /// `image_base` is the bare CDN root (`https://image.tmdb.org/t/p`); TMDB
+    /// requires a **size segment** (`/w500`, `/original`, …) between the root
+    /// and the file or it 404s. The const comment documents this contract and
+    /// this caller honors it by inserting [`TMDB_POSTER_SIZE`].
+    #[cfg_attr(not(test), allow(dead_code))]
+    pub(crate) async fn fetch_poster(&self, poster_path: &str) -> Option<bytes::Bytes> {
+        let url = format!(
+            "{}/{}/{}",
+            self.image_base.trim_end_matches('/'),
+            TMDB_POSTER_SIZE,
+            poster_path.trim_start_matches('/'),
+        );
+        let resp = self.http.get(&url).send().await.ok()?;
+        if !resp.status().is_success() {
+            debug!(
+                target: "meta-share::gateway",
+                upstream = "prowlarr",
+                poster_url = %url,
+                status = %resp.status(),
+                "tmdb poster fetch non-2xx; skipping"
+            );
+            return None;
+        }
+        resp.bytes().await.ok()
+    }
+
+    /// Fetch a TV show's authoritative structure (`number_of_seasons`
+    /// plus the per-season `episode_count`s) via `GET /3/tv/{id}`. Used
+    /// only to bounds-check a title-parsed season/episode — see
+    /// [`season_episode_bounds`]. Wrapped in the same search timeout;
+    /// any failure returns `None` so validation degrades to "accept".
+    pub(crate) async fn tv_details(&self, tmdbid: u64) -> TmdbCall<TmdbTvDetails> {
+        let url = format!(
+            "{}/tv/{}?append_to_response=alternative_titles",
+            self.api_base.trim_end_matches('/'),
+            tmdbid,
+        );
+        let fut = self
+            .http
+            .get(&url)
+            .bearer_auth(&self.bearer_token)
+            .header("accept", "application/json")
+            .send();
+        let resp =
+            match tokio::time::timeout(Duration::from_secs(TMDB_SEARCH_TIMEOUT_SECS), fut).await {
+                Ok(Ok(r)) => r,
+                _ => return TmdbCall::Miss,
+            };
+        if resp.status().as_u16() == 429 {
+            return TmdbCall::RateLimited(parse_tmdb_retry_after(&resp));
+        }
+        if !resp.status().is_success() {
+            debug!(
+                target: "meta-share::gateway",
+                upstream = "prowlarr",
+                tmdb_url = %url,
+                status = %resp.status(),
+                "tmdb tv-details non-2xx; skipping season/episode validation"
+            );
+            return TmdbCall::Miss;
+        }
+        match tokio::time::timeout(Duration::from_secs(TMDB_SEARCH_TIMEOUT_SECS), resp.json()).await
+        {
+            Ok(Ok(d)) => TmdbCall::Hit(d),
+            _ => TmdbCall::Miss,
+        }
+    }
+
+    /// Fetch cross-database ids (`tvdb_id`, `imdb_id`) for a known tmdbid via
+    /// `GET /3/{tv,movie}/{id}/external_ids`. The anchored torznab path needs
+    /// `tvdb_id` to query indexers by `tvsearch&tvdbid=` (TMDB's id is not a
+    /// standard `tvsearch` param). Same timeout/429 contract as [`tv_details`].
+    pub(crate) async fn external_ids(
+        &self,
+        kind: TmdbKind,
+        tmdbid: u64,
+    ) -> TmdbCall<TmdbExternalIds> {
+        let seg = match kind {
+            TmdbKind::Movie => "movie",
+            TmdbKind::Tv => "tv",
+        };
+        let url = format!(
+            "{}/{}/{}/external_ids",
+            self.api_base.trim_end_matches('/'),
+            seg,
+            tmdbid,
+        );
+        self.get_json(&url, "external_ids").await
+    }
+
+    /// Fetch a movie's canonical details via `GET /3/movie/{id}` (the anchored
+    /// enrichment source for a known movie tmdbid — resolves title/overview/
+    /// poster/year/imdb_id without a fuzzy title search). Same contract as
+    /// [`tv_details`].
+    pub(crate) async fn movie_details(&self, tmdbid: u64) -> TmdbCall<TmdbMovieDetails> {
+        let url = format!(
+            "{}/movie/{}?append_to_response=alternative_titles",
+            self.api_base.trim_end_matches('/'),
+            tmdbid
+        );
+        self.get_json(&url, "movie-details").await
+    }
+
+    /// Principal search via `GET /3/search/multi` — maps a bare keyword to a
+    /// canonical tmdbid + media type so the gateway can then query indexers
+    /// structurally (the "TMDB as the front door" identification step
+    /// Sonarr/Radarr use). Returns the mixed movie/tv/person result list; the
+    /// caller picks the top-N confident anchors via [`principal_top_n`].
+    pub(crate) async fn search_multi(&self, query: &str) -> TmdbCall<Vec<TmdbMultiItem>> {
+        let url = format!(
+            "{}/search/multi?query={}&include_adult=false",
+            self.api_base.trim_end_matches('/'),
+            urlencode(query),
+        );
+        match self
+            .get_json::<TmdbMultiResponse>(&url, "search/multi")
+            .await
+        {
+            TmdbCall::Hit(r) => TmdbCall::Hit(r.results),
+            TmdbCall::Miss => TmdbCall::Miss,
+            TmdbCall::RateLimited(d) => TmdbCall::RateLimited(d),
+        }
+    }
+
+    /// Fetch a TMDB **discovery list** (trending / popular / top-rated / a
+    /// `/discover` query) and map each result to a [`TmdbHit`]. Powers the
+    /// keyword-less discovery branch (see [`super::discovery`]): the caller
+    /// composes `path_and_query` (e.g. `"movie/popular"`,
+    /// `"trending/tv/week"`, or `"discover/tv?with_keywords=210024&sort_by=…"`)
+    /// and passes the `kind` so movie-vs-TV title/date fields decode correctly.
+    /// All these endpoints share the `{results: [TmdbSearchItem]}` shape, so one
+    /// method covers them. Returns the page's hits, or `Miss`/`RateLimited` under
+    /// the usual best-effort contract.
+    pub(crate) async fn discovery_list(
+        &self,
+        kind: TmdbKind,
+        path_and_query: &str,
+    ) -> TmdbCall<Vec<TmdbHit>> {
+        let url = format!(
+            "{}/{}",
+            self.api_base.trim_end_matches('/'),
+            path_and_query.trim_start_matches('/'),
+        );
+        match self.get_json::<TmdbSearchResponse>(&url, "discovery").await {
+            TmdbCall::Hit(r) => {
+                TmdbCall::Hit(r.results.into_iter().map(|i| i.into_hit(kind)).collect())
+            }
+            TmdbCall::Miss => TmdbCall::Miss,
+            TmdbCall::RateLimited(d) => TmdbCall::RateLimited(d),
+        }
+    }
+
+    /// Shared `GET <url>` → JSON helper with the standard bearer auth, search
+    /// timeout, 429→`RateLimited`, non-2xx/transient→`Miss` contract. Factored
+    /// out of [`tv_details`]/[`external_ids`]/[`movie_details`] (identical
+    /// modulo the decoded type and the log label).
+    async fn get_json<T: serde::de::DeserializeOwned>(&self, url: &str, what: &str) -> TmdbCall<T> {
+        let fut = self
+            .http
+            .get(url)
+            .bearer_auth(&self.bearer_token)
+            .header("accept", "application/json")
+            .send();
+        let resp =
+            match tokio::time::timeout(Duration::from_secs(TMDB_SEARCH_TIMEOUT_SECS), fut).await {
+                Ok(Ok(r)) => r,
+                // Transport error / timeout. Logged at debug so the otherwise
+                // silent best-effort degrade is diagnosable.
+                Ok(Err(e)) => {
+                    debug!(target: "meta-share::gateway", upstream="prowlarr", tmdb_url=%url, what, error=%e, "tmdb call send error; degrading");
+                    return TmdbCall::Miss;
+                }
+                Err(_) => {
+                    debug!(target: "meta-share::gateway", upstream="prowlarr", tmdb_url=%url, what, "tmdb call send timeout; degrading");
+                    return TmdbCall::Miss;
+                }
+            };
+        if resp.status().as_u16() == 429 {
+            return TmdbCall::RateLimited(parse_tmdb_retry_after(&resp));
+        }
+        if !resp.status().is_success() {
+            debug!(
+                target: "meta-share::gateway",
+                upstream = "prowlarr",
+                tmdb_url = %url,
+                status = %resp.status(),
+                what,
+                "tmdb call non-2xx; degrading"
+            );
+            return TmdbCall::Miss;
+        }
+        match tokio::time::timeout(Duration::from_secs(TMDB_SEARCH_TIMEOUT_SECS), resp.json()).await {
+            Ok(Ok(d)) => TmdbCall::Hit(d),
+            // A decode failure here folds into `Miss` (best-effort), but log it
+            // at debug: a single malformed element fails the whole `Vec` decode,
+            // which would otherwise silently disable e.g. multi-anchor.
+            Ok(Err(e)) => {
+                debug!(target: "meta-share::gateway", upstream="prowlarr", tmdb_url=%url, what, error=%e, "tmdb call decode error; degrading");
+                TmdbCall::Miss
+            }
+            Err(_) => {
+                debug!(target: "meta-share::gateway", upstream="prowlarr", tmdb_url=%url, what, "tmdb call decode timeout; degrading");
+                TmdbCall::Miss
+            }
+        }
+    }
+}
+
+/// Outcome of a TMDB API call. Distinguishes a 429 rate-limit (which feeds
+/// `Retry-After` into the shared [`TmdbBudget`] so all enrichment pauses
+/// globally) from a plain miss or transient error (best-effort degrade).
+pub(crate) enum TmdbCall<T> {
+    Hit(T),
+    Miss,
+    RateLimited(Duration),
+}
+
+/// Parse a TMDB `Retry-After` header into a pause duration, defaulting to
+/// [`TMDB_RATE_LIMIT_DEFAULT_RETRY_SECS`] when absent/unparseable. Mirrors
+/// the indexer-path logic in [`map_status`].
+pub(crate) fn parse_tmdb_retry_after(resp: &reqwest::Response) -> Duration {
+    let secs = resp
+        .headers()
+        .get(reqwest::header::RETRY_AFTER)
+        .and_then(|v| v.to_str().ok())
+        .and_then(|v| v.parse::<u64>().ok())
+        .unwrap_or(TMDB_RATE_LIMIT_DEFAULT_RETRY_SECS);
+    Duration::from_secs(secs)
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) enum TmdbKind {
+    Movie,
+    Tv,
+}
+
+#[derive(serde::Deserialize)]
+pub(crate) struct TmdbSearchResponse {
+    #[serde(default)]
+    pub(crate) results: Vec<TmdbSearchItem>,
+}
+
+#[derive(serde::Deserialize)]
+pub(crate) struct TmdbSearchItem {
+    pub(crate) id: u64,
+    #[serde(default)]
+    pub(crate) title: Option<String>, // movie
+    #[serde(default)]
+    pub(crate) name: Option<String>, // tv
+    #[serde(default)]
+    pub(crate) original_title: Option<String>, // movie
+    #[serde(default)]
+    pub(crate) original_name: Option<String>, // tv
+    #[serde(default)]
+    pub(crate) overview: Option<String>,
+    #[serde(default)]
+    pub(crate) release_date: Option<String>, // movie, YYYY-MM-DD
+    #[serde(default)]
+    pub(crate) first_air_date: Option<String>, // tv,    YYYY-MM-DD
+    #[serde(default)]
+    pub(crate) poster_path: Option<String>,
+    #[serde(default)]
+    pub(crate) genre_ids: Vec<u32>,
+    /// ISO 639-1 (2-letter) language the title was originally produced in.
+    /// Mapped to `lang3` and used to file `original_title` under
+    /// `titles/{lang3}` (METADATA_KEYS.md §3).
+    #[serde(default)]
+    pub(crate) original_language: Option<String>,
+}
+
+impl TmdbSearchItem {
+    pub(crate) fn into_hit(self, kind: TmdbKind) -> TmdbHit {
+        let title = match kind {
+            TmdbKind::Movie => self.title,
+            TmdbKind::Tv => self.name,
+        }
+        .unwrap_or_else(|| "(untitled)".to_string());
+        let original_title = match kind {
+            TmdbKind::Movie => self.original_title,
+            TmdbKind::Tv => self.original_name,
+        };
+        let date = match kind {
+            TmdbKind::Movie => self.release_date,
+            TmdbKind::Tv => self.first_air_date,
+        };
+        let year = date
+            .as_deref()
+            .filter(|d| d.len() >= 4)
+            .and_then(|d| d[..4].parse::<u16>().ok());
+        TmdbHit {
+            tmdbid: self.id,
+            title,
+            original_title,
+            original_language: self.original_language,
+            overview: self.overview,
+            year,
+            poster_path: self.poster_path,
+            genre_ids: self.genre_ids,
+            // search-list items carry no AKAs (a details/append fetch does)
+            alt_titles: Vec::new(),
+        }
+    }
+}
+
+/// Map an ISO 639-1 (2-letter) code to its ISO 639-3 (`lang3`) equivalent,
+/// matching the store's convention (`eng`, `jpn`, `fra`; the 639-2/T variant
+/// where B/T differ, e.g. `deu` not `ger`). Covers the languages TMDB
+/// commonly returns; unknown codes return `None` so the caller skips the
+/// `titles/{lang3}` write rather than persisting a non-`lang3` key.
+pub(crate) fn iso639_1_to_3(code: &str) -> Option<&'static str> {
+    Some(match code.to_ascii_lowercase().as_str() {
+        "en" => "eng",
+        "ja" => "jpn",
+        "fr" => "fra",
+        "de" => "deu",
+        "es" => "spa",
+        "it" => "ita",
+        "ru" => "rus",
+        "ko" => "kor",
+        "zh" => "zho",
+        "pt" => "por",
+        "nl" => "nld",
+        "sv" => "swe",
+        "no" => "nor",
+        "da" => "dan",
+        "fi" => "fin",
+        "pl" => "pol",
+        "tr" => "tur",
+        "ar" => "ara",
+        "hi" => "hin",
+        "th" => "tha",
+        "vi" => "vie",
+        "id" => "ind",
+        "cs" => "ces",
+        "el" => "ell",
+        "he" => "heb",
+        "hu" => "hun",
+        "ro" => "ron",
+        "uk" => "ukr",
+        "fa" => "fas",
+        "ms" => "msa",
+        "tl" => "tgl",
+        "ca" => "cat",
+        "nb" => "nob",
+        _ => return None,
+    })
+}
+
+// `Clone` is required so the single-flight `Shared` future (whose `Output`
+// must be `Clone`) can hand the same hit to every coalesced caller; serde so
+// hits persist in the redb TMDB-search cache.
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub(crate) struct TmdbHit {
+    pub(crate) tmdbid: u64,
+    pub(crate) title: String,
+    #[serde(default)]
+    pub(crate) original_title: Option<String>,
+    #[serde(default)]
+    pub(crate) original_language: Option<String>,
+    #[serde(default)]
+    pub(crate) overview: Option<String>,
+    #[serde(default)]
+    pub(crate) year: Option<u16>,
+    #[serde(default)]
+    pub(crate) poster_path: Option<String>,
+    #[serde(default)]
+    #[allow(dead_code)] // surfaced as CSV of ids; name resolution would need a second call
+    pub(crate) genre_ids: Vec<u32>,
+    /// TMDB alternative / AKA titles (romaji + foreign), used by the id-based-job
+    /// relevance guard so a differently-titled release still resembles the show.
+    /// Empty for hits built from a search list (which carries no AKAs).
+    #[serde(default)]
+    pub(crate) alt_titles: Vec<String>,
+}
+
+impl TmdbHit {
+    /// TMDB `original_language` (ISO 639-1) mapped to the store's `lang3`
+    /// (ISO 639-3), or `None` when the language is outside the common set or
+    /// absent. Used to file `original_title` under `titles/{lang3}` (§3).
+    pub(crate) fn original_lang3(&self) -> Option<&'static str> {
+        iso639_1_to_3(self.original_language.as_deref()?.trim())
+    }
+}
+
+/// Authoritative TV structure from `GET /3/tv/{id}`, used to bounds-check
+/// a title-parsed season/episode. Only the two structural fields are
+/// decoded; everything else in the (large) details payload is ignored.
+// `Clone` + `Serialize` added so details persist in the redb TMDB-tvdetails
+// cache and can be cheaply handed around.
+#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
+pub(crate) struct TmdbTvDetails {
+    #[serde(default)]
+    pub(crate) number_of_seasons: u32,
+    /// One entry per season TMDB knows about, including season 0
+    /// ("Specials"). `episode_count` is the released-episode total.
+    #[serde(default)]
+    pub(crate) seasons: Vec<TmdbSeasonSummary>,
+    // -- Display fields ------------------------------------------------------
+    // Decoded from the same `GET /3/tv/{id}` payload so an anchored TV record
+    // (known tmdbid) enriches directly from its canonical entry instead of a
+    // fuzzy title search. Old cache entries (written before these fields
+    // existed) deserialize them to empty/None; [`TmdbTvDetails::has_display`]
+    // detects that so the cached-fetch path can self-heal with one refetch.
+    #[serde(default)]
+    pub(crate) name: String,
+    #[serde(default)]
+    pub(crate) original_name: Option<String>,
+    #[serde(default)]
+    pub(crate) original_language: Option<String>,
+    #[serde(default)]
+    pub(crate) overview: Option<String>,
+    #[serde(default)]
+    pub(crate) first_air_date: Option<String>,
+    #[serde(default)]
+    pub(crate) poster_path: Option<String>,
+    /// AKA titles from `?append_to_response=alternative_titles`. `None` on a
+    /// pre-AKA cache entry → [`TmdbTvDetails::has_akas`] is false so the cached
+    /// fetch self-heals with one refetch (same pattern as the display fields).
+    #[serde(default)]
+    pub(crate) alternative_titles: Option<TmdbAltTitles>,
+}
+
+impl TmdbTvDetails {
+    /// True when the cached payload carries the display fields (i.e. was
+    /// written by the current decoder). A `false` here on a cache hit means a
+    /// pre-display entry — the caller refetches once to upgrade it.
+    pub(crate) fn has_display(&self) -> bool {
+        !self.name.trim().is_empty()
+    }
+
+    /// True when the AKA field was populated by the current (append-enabled)
+    /// decoder. A `false` on a cache hit means a pre-AKA entry → the cached
+    /// fetch refetches once to upgrade it (mirrors [`has_display`]).
+    pub(crate) fn has_akas(&self) -> bool {
+        self.alternative_titles.is_some()
+    }
+
+    /// The show's canonical + original + AKA titles, for the id-based-job
+    /// relevance guard.
+    pub(crate) fn guard_titles(&self) -> Vec<String> {
+        let mut out = vec![self.name.clone()];
+        if let Some(o) = &self.original_name {
+            out.push(o.clone());
+        }
+        if let Some(a) = &self.alternative_titles {
+            out.extend(a.all());
+        }
+        out
+    }
+
+    /// Build a [`TmdbHit`] (resolved by id) from the canonical TV entry, for
+    /// the anchored enrichment path. `None` when display fields are absent.
+    pub(crate) fn as_hit(&self, tmdbid: u64) -> Option<TmdbHit> {
+        if !self.has_display() {
+            return None;
+        }
+        let year = self
+            .first_air_date
+            .as_deref()
+            .filter(|d| d.len() >= 4)
+            .and_then(|d| d[..4].parse::<u16>().ok());
+        Some(TmdbHit {
+            tmdbid,
+            title: self.name.clone(),
+            original_title: self.original_name.clone(),
+            original_language: self.original_language.clone(),
+            overview: self.overview.clone(),
+            year,
+            poster_path: self.poster_path.clone(),
+            genre_ids: Vec::new(),
+            alt_titles: self
+                .alternative_titles
+                .as_ref()
+                .map(TmdbAltTitles::all)
+                .unwrap_or_default(),
+        })
+    }
+}
+
+/// Alternative / AKA titles from `?append_to_response=alternative_titles`.
+/// TMDB nests them under `results` for `/tv/{id}` and under `titles` for
+/// `/movie/{id}`; both entries carry a `title`. Decoded so the id-based-job
+/// relevance guard ([`super::torznab::apply_record_tag`]) can match a
+/// romaji/foreign release title (e.g. `Sousou no Frieren`) against a show whose
+/// canonical `name` is the English one (`Frieren: Beyond Journey's End`).
+///
+/// `Option<TmdbAltTitles>` on the details structs doubles as a fetched-marker:
+/// `None` means "written by a decoder that didn't append AKAs" (an old cache
+/// entry) and triggers a one-shot self-heal refetch; `Some` (even empty) means
+/// "AKAs were fetched, this show simply has none".
+#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
+pub(crate) struct TmdbAltTitles {
+    #[serde(default)]
+    pub(crate) results: Vec<TmdbAltTitle>, // tv
+    #[serde(default)]
+    pub(crate) titles: Vec<TmdbAltTitle>, // movie
+}
+
+#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
+pub(crate) struct TmdbAltTitle {
+    #[serde(default)]
+    pub(crate) title: String,
+}
+
+impl TmdbAltTitles {
+    /// The non-empty AKA titles across both the TV (`results`) and movie
+    /// (`titles`) shapes.
+    pub(crate) fn all(&self) -> Vec<String> {
+        self.results
+            .iter()
+            .chain(self.titles.iter())
+            .map(|t| t.title.trim().to_string())
+            .filter(|t| !t.is_empty())
+            .collect()
+    }
+}
+
+/// TMDB cross-database ids from `GET /3/{tv,movie}/{id}/external_ids`. Only the
+/// two the anchored torznab path needs are decoded.
+#[derive(Debug, Clone, Default, serde::Serialize, serde::Deserialize)]
+pub(crate) struct TmdbExternalIds {
+    #[serde(default)]
+    pub(crate) tvdb_id: Option<i64>,
+    #[serde(default)]
+    pub(crate) imdb_id: Option<String>,
+}
+
+/// Authoritative movie structure from `GET /3/movie/{id}`. Decodes only the
+/// display fields the anchored path needs; `imdb_id` (movie details carry it
+/// inline, no separate `external_ids` call) feeds the `t=movie&imdbid=` query
+/// fallback.
+#[derive(Debug, Clone, serde::Deserialize)]
+pub(crate) struct TmdbMovieDetails {
+    pub(crate) id: u64,
+    #[serde(default)]
+    pub(crate) title: Option<String>,
+    #[serde(default)]
+    pub(crate) original_title: Option<String>,
+    #[serde(default)]
+    pub(crate) original_language: Option<String>,
+    #[serde(default)]
+    pub(crate) overview: Option<String>,
+    #[serde(default)]
+    pub(crate) release_date: Option<String>,
+    #[serde(default)]
+    pub(crate) poster_path: Option<String>,
+    #[serde(default)]
+    pub(crate) alternative_titles: Option<TmdbAltTitles>,
+}
+
+impl TmdbMovieDetails {
+    pub(crate) fn into_hit(self) -> TmdbHit {
+        let year = self
+            .release_date
+            .as_deref()
+            .filter(|d| d.len() >= 4)
+            .and_then(|d| d[..4].parse::<u16>().ok());
+        let alt_titles = self
+            .alternative_titles
+            .as_ref()
+            .map(TmdbAltTitles::all)
+            .unwrap_or_default();
+        TmdbHit {
+            tmdbid: self.id,
+            title: self.title.unwrap_or_else(|| "(untitled)".to_string()),
+            original_title: self.original_title,
+            original_language: self.original_language,
+            overview: self.overview,
+            year,
+            poster_path: self.poster_path,
+            genre_ids: Vec::new(),
+            alt_titles,
+        }
+    }
+}
+
+#[derive(serde::Deserialize)]
+pub(crate) struct TmdbMultiResponse {
+    #[serde(default)]
+    pub(crate) results: Vec<TmdbMultiItem>,
+}
+
+/// One `search/multi` result. `media_type` discriminates movie / tv / person;
+/// we anchor only on movie/tv.
+#[derive(Debug, Clone, serde::Deserialize)]
+pub(crate) struct TmdbMultiItem {
+    pub(crate) id: u64,
+    #[serde(default)]
+    pub(crate) media_type: Option<String>,
+    #[serde(default)]
+    pub(crate) title: Option<String>, // movie
+    #[serde(default)]
+    pub(crate) name: Option<String>, // tv
+    #[serde(default)]
+    pub(crate) popularity: f64,
+}
+
+impl TmdbMultiItem {
+    pub(crate) fn kind(&self) -> Option<TmdbKind> {
+        match self.media_type.as_deref() {
+            Some("movie") => Some(TmdbKind::Movie),
+            Some("tv") => Some(TmdbKind::Tv),
+            _ => None,
+        }
+    }
+    pub(crate) fn display_title(&self) -> &str {
+        self.title.as_deref().or(self.name.as_deref()).unwrap_or("")
+    }
+}
+
+/// Lowercase alphanumeric word tokens of `s` (separators dropped). Shared by
+/// the principal-search confidence check.
+pub(crate) fn norm_tokens(s: &str) -> Vec<String> {
+    s.split(|c: char| !c.is_alphanumeric())
+        .filter(|t| !t.is_empty())
+        .map(|t| t.to_lowercase())
+        .collect()
+}
+
+/// Pick up to `n` **confident** anchors from `search/multi` hits for `query`,
+/// most-popular first (or `[]` → the caller falls back to today's generic
+/// `t=search&q=`). For each movie/tv hit, accept it only if its title shares at
+/// least half of the query's word tokens (the per-candidate relevance gate),
+/// then rank the survivors by popularity and take the top `n`. This binds a
+/// vague keyword like "black" to the *several* real shows TMDB returns (Black
+/// Mirror, Black Butler, Black Lagoon, …) instead of hijacking the whole query
+/// onto the single most-popular title, while still dropping unrelated noise
+/// (e.g. "zzz qqq www" → nothing). `n == 1` recovers the old single-anchor
+/// "principal_confident" behaviour (modulo the gate being applied before the
+/// popularity pick rather than after).
+pub(crate) fn principal_top_n<'a>(
+    hits: &'a [TmdbMultiItem],
+    query: &str,
+    n: usize,
+) -> Vec<&'a TmdbMultiItem> {
+    let q_tokens = norm_tokens(query);
+    if q_tokens.is_empty() || n == 0 {
+        return Vec::new();
+    }
+    let mut cands: Vec<&TmdbMultiItem> = hits
+        .iter()
+        .filter(|h| h.kind().is_some())
+        .filter(|h| {
+            let t_tokens = norm_tokens(h.display_title());
+            let overlap = q_tokens.iter().filter(|t| t_tokens.contains(t)).count();
+            (overlap as f64) >= (q_tokens.len() as f64) * 0.5
+        })
+        .collect();
+    cands.sort_by(|a, b| {
+        b.popularity
+            .partial_cmp(&a.popularity)
+            .unwrap_or(std::cmp::Ordering::Equal)
+    });
+    cands.truncate(n);
+    cands
+}
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub(crate) struct TmdbSeasonSummary {
+    #[serde(default)]
+    pub(crate) season_number: i64,
+    #[serde(default)]
+    pub(crate) episode_count: u32,
+}
+
+/// Verdict of [`season_episode_bounds`] — how a title-parsed `(season,
+/// episode)` lines up against TMDB's actual structure.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum SeasonEpisodeBounds {
+    /// Consistent with TMDB, or unverifiable — keep the record as-is.
+    Ok,
+    /// The parsed season sits past TMDB's known seasons and isn't otherwise
+    /// listed in `seasons[]`. The canonical case is anime where fansubbers
+    /// number a later cour `S2` of a show TMDB models as a **single** season
+    /// (Frieren, Jujutsu Kaisen, …). The TMDB match itself is sound — only the
+    /// season number is a naming-convention artifact — so the caller keeps the
+    /// record but strips the misleading `season` (the frontend buckets
+    /// season-less episodes into a flat "Episodes" list).
+    SeasonOverflow,
+    /// The season IS a real TMDB season, but the parsed episode runs past that
+    /// season's `episode_count`. The canonical case is absolute-numbered anime
+    /// that an indexer has relabelled onto season 1 (`S01E96` on a 28-episode
+    /// show): the `S01` is a numbering artifact, not a trustworthy token, so —
+    /// exactly like [`SeasonOverflow`] — the caller strips the season and keeps
+    /// the record (season-less episode, bucketed under the show's tmdbid). We
+    /// deliberately do NOT drop it: an out-of-bounds episode means the season
+    /// was never reliable, not that the whole release is bogus.
+    EpisodeOverflow,
+    /// A contradiction neither overflow allowance can explain — currently only a
+    /// negative season (codec noise like `x.265` → "S0E265" is caught upstream;
+    /// a genuinely garbled parse). The title parse is wrong — the caller drops
+    /// the record.
+    Contradiction,
+}
+
+/// Pure bounds check: is a title-parsed `(season, episode)` consistent
+/// with what TMDB says the show actually has? Deliberately lenient —
+/// it only returns a verdict other than [`SeasonEpisodeBounds::Ok`] on a
+/// *positive* divergence, so anything ambiguous or unverifiable is accepted:
+///
+/// - No parsed season → [`Ok`]. Episode-only titles (absolute-numbered
+///   anime, trailing `- 117`) are legitimately unbounded; the season-
+///   based bound doesn't apply.
+/// - TMDB gave us no structure (`number_of_seasons == 0` and no
+///   `seasons[]`) → [`Ok`]; we can't contradict what we don't know.
+/// - Season exceeds `number_of_seasons` AND isn't otherwise listed in
+///   `seasons[]` → [`SeasonOverflow`] (the `S2`-of-a-1-TMDB-season anime
+///   case; keep but strip `season`).
+/// - Negative season → [`Contradiction`].
+/// - Episode exceeds the matched season's `episode_count` (when that
+///   count is known) → [`EpisodeOverflow`]. This fires for explicit `SxxEyy`
+///   titles whose episode overruns the season — overwhelmingly absolute-
+///   numbered anime relabelled onto season 1 — so we strip the season and keep
+///   the record rather than dropping it.
+///
+/// [`Ok`]: SeasonEpisodeBounds::Ok
+/// [`SeasonOverflow`]: SeasonEpisodeBounds::SeasonOverflow
+/// [`EpisodeOverflow`]: SeasonEpisodeBounds::EpisodeOverflow
+/// [`Contradiction`]: SeasonEpisodeBounds::Contradiction
+pub(crate) fn season_episode_bounds(
+    details: &TmdbTvDetails,
+    season: Option<i64>,
+    episode: Option<i64>,
+) -> SeasonEpisodeBounds {
+    use SeasonEpisodeBounds::*;
+    let season = match season {
+        Some(s) => s,
+        None => return Ok,
+    };
+    if details.number_of_seasons == 0 && details.seasons.is_empty() {
+        return Ok;
+    }
+    if season < 0 {
+        return Contradiction;
+    }
+    let matched = details.seasons.iter().find(|s| s.season_number == season);
+    if season > details.number_of_seasons as i64 && matched.is_none() {
+        return SeasonOverflow;
+    }
+    if let (Some(ep), Some(sm)) = (episode, matched) {
+        if sm.episode_count > 0 && ep > sm.episode_count as i64 {
+            return EpisodeOverflow;
+        }
+    }
+    Ok
+}
+
+/// True when `episode` is a valid episode of `season` per TMDB's per-season
+/// `episode_count` (the season must be listed and the count known). Used to
+/// decide whether a bare title number is an in-season episode before falling
+/// back to absolute-number interpretation.
+pub(crate) fn episode_in_season(details: &TmdbTvDetails, season: i64, episode: i64) -> bool {
+    details
+        .seasons
+        .iter()
+        .find(|s| s.season_number == season)
+        .map(|s| s.episode_count > 0 && episode >= 1 && episode <= s.episode_count as i64)
+        .unwrap_or(false)
+}
+
+use crate::consts::*;
+
+#[cfg(test)]
+mod season_bounds_tests {
+    use super::*;
+
+    fn details(num_seasons: u32, seasons: &[(i64, u32)]) -> TmdbTvDetails {
+        TmdbTvDetails {
+            number_of_seasons: num_seasons,
+            seasons: seasons
+                .iter()
+                .map(|&(season_number, episode_count)| TmdbSeasonSummary {
+                    season_number,
+                    episode_count,
+                })
+                .collect(),
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn bounds_ok_within_season() {
+        let d = details(2, &[(1, 26), (2, 24)]);
+        assert_eq!(season_episode_bounds(&d, Some(1), Some(5)), SeasonEpisodeBounds::Ok);
+        assert_eq!(season_episode_bounds(&d, Some(2), Some(24)), SeasonEpisodeBounds::Ok);
+    }
+
+    #[test]
+    fn bounds_no_season_or_unknown_structure_is_ok() {
+        let d = details(2, &[(1, 26), (2, 24)]);
+        assert_eq!(season_episode_bounds(&d, None, Some(999)), SeasonEpisodeBounds::Ok);
+        let unknown = details(0, &[]);
+        assert_eq!(season_episode_bounds(&unknown, Some(5), Some(40)), SeasonEpisodeBounds::Ok);
+    }
+
+    #[test]
+    fn bounds_season_past_total_is_overflow() {
+        // Fansub "S2" of a show TMDB models as a single season → strip-and-keep.
+        let d = details(1, &[(1, 28)]);
+        assert_eq!(season_episode_bounds(&d, Some(2), Some(3)), SeasonEpisodeBounds::SeasonOverflow);
+    }
+
+    #[test]
+    fn bounds_episode_past_matched_season_is_episode_overflow() {
+        // "S1E27" on a show whose season 1 has 26 episodes → the `S01` is an
+        // absolute-numbering artifact, not a trustworthy token. The enrich path
+        // strips the season and KEEPS the record (season-less absolute episode),
+        // rather than dropping it.
+        let d = details(4, &[(1, 26), (2, 13), (3, 13), (4, 12)]);
+        assert_eq!(season_episode_bounds(&d, Some(1), Some(27)), SeasonEpisodeBounds::EpisodeOverflow);
+    }
+
+    #[test]
+    fn bounds_negative_season_is_contradiction() {
+        // A genuinely garbled parse (negative season) is still a hard drop.
+        let d = details(4, &[(1, 26), (2, 13), (3, 13), (4, 12)]);
+        assert_eq!(season_episode_bounds(&d, Some(-1), None), SeasonEpisodeBounds::Contradiction);
+    }
+}

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

@@ -0,0 +1,233 @@
+//! End-to-end contract test for the card feeder's HTTP surface: boots the real
+//! `serve_feeders` router (card-feeder hosting the TMDB plugin) on an ephemeral
+//! port, pointed at a wiremock TMDB, and drives `/manifest`, `/query` and
+//! `/compute` over HTTP. Deterministic — no live TMDB.
+//!
+//! What this pins is the *card contract* specifically:
+//! - the manifest advertises `fileType=card`, which is the only reason
+//!   meta-search routes a `fileType:card` query here;
+//! - a query yields records carrying both type axes plus the whole id bag;
+//! - `/compute` yields a **byte-less** `card_locator` outcome — the shape that
+//!   lands in the gateway core's metadata-only auto-store branch.
+
+use card_feeder::tmdb::TmdbCardPlugin;
+use meta_feeder_sdk::{
+    configure_plugins, router, ComputeRequest, ComputeResponse, GatewayQuery, HashKindDto,
+    ManifestResponse, QueryRequest, QueryResponse,
+};
+use wiremock::matchers::{method, path, query_param};
+use wiremock::{Mock, MockServer, ResponseTemplate};
+
+/// Spawn the feeder router (TMDB pointed at `upstream`) on an ephemeral port.
+/// 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 = TmdbCardPlugin::with_api_base("test-token", upstream.uri(), upstream.uri());
+    let plugins = configure_plugins(vec![Box::new(plugin)], dir.path()).expect("configure plugins");
+    let app = router(plugins, "card-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)
+}
+
+/// `search/multi` → one confident TV hit for "frieren".
+fn multi_json() -> serde_json::Value {
+    serde_json::json!({
+        "results": [{
+            "id": 95479,
+            "media_type": "tv",
+            "name": "Frieren: Beyond Journey's End",
+            "original_name": "\u{845}\u{9001}\u{306e}\u{30d5}\u{30ea}\u{30fc}\u{30ec}\u{30f3}",
+            "overview": "The story follows the elf mage Frieren.",
+            "poster_path": "/poster.jpg",
+            "first_air_date": "2023-09-29",
+            "popularity": 300.0,
+            "vote_count": 900
+        }]
+    })
+}
+
+fn tv_details_json() -> serde_json::Value {
+    serde_json::json!({
+        "id": 95479,
+        "name": "Frieren: Beyond Journey's End",
+        "original_name": "\u{845}\u{9001}\u{306e}\u{30d5}\u{30ea}\u{30fc}\u{30ec}\u{30f3}",
+        "original_language": "ja",
+        "overview": "The story follows the elf mage Frieren.",
+        "first_air_date": "2023-09-29",
+        "poster_path": "/poster.jpg",
+        "number_of_seasons": 1,
+        "seasons": [{"season_number": 1, "episode_count": 28}],
+        "alternative_titles": { "results": [{"title": "Sousou no Frieren"}] }
+    })
+}
+
+fn external_ids_json() -> serde_json::Value {
+    serde_json::json!({ "imdb_id": "tt22248376", "tvdb_id": 367189 })
+}
+
+async fn mount_tmdb(upstream: &MockServer) {
+    Mock::given(method("GET"))
+        .and(path("/search/multi"))
+        .respond_with(ResponseTemplate::new(200).set_body_json(multi_json()))
+        .mount(upstream)
+        .await;
+    Mock::given(method("GET"))
+        .and(path("/tv/95479"))
+        .respond_with(ResponseTemplate::new(200).set_body_json(tv_details_json()))
+        .mount(upstream)
+        .await;
+    Mock::given(method("GET"))
+        .and(path("/tv/95479/external_ids"))
+        .respond_with(ResponseTemplate::new(200).set_body_json(external_ids_json()))
+        .mount(upstream)
+        .await;
+}
+
+fn card_query(text: &str) -> GatewayQuery {
+    let mut q = GatewayQuery::from_free_text(text);
+    q.filters
+        .insert("fileType".to_string(), vec!["card".to_string()]);
+    q
+}
+
+/// The routing declaration. `served_file_types = ["card"]` is what makes
+/// meta-search fan a `fileType:card` query out to this gateway at all.
+#[tokio::test]
+async fn manifest_advertises_the_card_file_type() {
+    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, "tmdb");
+    assert_eq!(p.served_file_types, vec!["card".to_string()]);
+    assert_eq!(
+        p.served_content_kinds,
+        vec!["movie".to_string(), "series".to_string()]
+    );
+}
+
+#[tokio::test]
+async fn query_returns_a_card_with_both_axes_and_the_id_bag() {
+    let upstream = MockServer::start().await;
+    mount_tmdb(&upstream).await;
+    let (base, _dir) = spawn_feeder(&upstream).await;
+
+    let resp: QueryResponse = reqwest::Client::new()
+        .post(format!("{base}/query"))
+        .json(&QueryRequest {
+            upstream_id: "tmdb".to_string(),
+            query: card_query("frieren"),
+            max_results: 10,
+        })
+        .send()
+        .await
+        .expect("POST /query")
+        .json()
+        .await
+        .expect("decode query response");
+
+    assert_eq!(resp.records.len(), 1, "one confident TMDB hit → one card");
+    let f = &resp.records[0].fields;
+    assert_eq!(f["fileType"], "card");
+    assert_eq!(f["contentKind"], "series");
+    assert_eq!(f["title"], "Frieren: Beyond Journey's End");
+    // The id bag — phase 2 selects from this rather than assuming `tmdbid`.
+    assert_eq!(f["tmdbid"], "95479");
+    assert_eq!(f["tvdbid"], "367189");
+    assert_eq!(f["imdbid"], "tt22248376");
+    // Display trio that meta-watch's quality gate requires.
+    assert!(f.contains_key("description/eng"));
+    assert!(f["poster_url"].contains("poster.jpg"));
+    // The record id is the CID preimage.
+    assert_eq!(resp.records[0].record_id, "tmdb:tv:95479");
+}
+
+/// The core of the design: a card resolves to a locator with **no bytes**.
+/// That `bytes_b64: None` is what routes it to `store_metadata_only` in the
+/// gateway core's three-branch auto-store.
+#[tokio::test]
+async fn compute_returns_a_byteless_card_locator() {
+    let upstream = MockServer::start().await;
+    mount_tmdb(&upstream).await;
+    let (base, _dir) = spawn_feeder(&upstream).await;
+
+    let resp: ComputeResponse = reqwest::Client::new()
+        .post(format!("{base}/compute"))
+        .json(&ComputeRequest {
+            upstream_id: "tmdb".to_string(),
+            record_id: "tmdb:tv:95479".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!(matches!(o.hash_kind, HashKindDto::CardLocator));
+    assert!(o.bytes_b64.is_none(), "a card has no bytes, ever");
+    assert!(o.file_extension.is_none());
+    assert!(o.record.is_some(), "the record IS the payload");
+    // Deterministic in (source, id) — any peer derives the same address.
+    assert_eq!(
+        o.hash,
+        meta_feeder_sdk::hash::compute_card_cid("tmdb", "tv:95479").unwrap()
+    );
+}
+
+/// A card query never reaches an indexer, and a *typed* query for something the
+/// card tier does not serve short-circuits before TMDB is touched at all.
+#[tokio::test]
+async fn a_video_query_does_not_hit_tmdb() {
+    let upstream = MockServer::start().await;
+    // Deliberately mount nothing: any outbound call would 404 and surface as an
+    // empty result, but the assertion below is that we never even get there.
+    Mock::given(method("GET"))
+        .and(path("/search/multi"))
+        .and(query_param("query", "frieren"))
+        .respond_with(ResponseTemplate::new(500))
+        .expect(0)
+        .mount(&upstream)
+        .await;
+    let (base, _dir) = spawn_feeder(&upstream).await;
+
+    let mut q = GatewayQuery::from_free_text("frieren");
+    q.filters
+        .insert("fileType".to_string(), vec!["video".to_string()]);
+
+    let resp: QueryResponse = reqwest::Client::new()
+        .post(format!("{base}/query"))
+        .json(&QueryRequest {
+            upstream_id: "tmdb".to_string(),
+            query: q,
+            max_results: 10,
+        })
+        .send()
+        .await
+        .expect("POST /query")
+        .json()
+        .await
+        .expect("decode query response");
+
+    assert!(resp.records.is_empty());
+    // `expect(0)` above is asserted on MockServer drop.
+}

rust-toolchain.toml

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