add transport feature: meta-share transport-plugin contract, client, harness, focus replica, testkit

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "meta-feeder-sdk"
-version = "1.3.0"
+version = "1.4.0"
 edition = "2021"
 rust-version = "1.85"
 license = "MIT"
@@ -31,6 +31,14 @@ publish = false
 # 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.
 
+[features]
+# The meta-share transport-plugin contract (`src/transport/`): the hull's
+# client, the plugins' axum harness, the focus replica, shared error/range
+# types. Off by default so feeders build exactly what they always did.
+transport = ["dep:cid", "dep:libc", "dep:tokio-util"]
+# Conformance checks any transport implementation must pass.
+testkit = ["transport"]
+
 [dependencies]
 # Async runtime + HTTP server (the serve_feeders harness)
 tokio = { version = "1.40", features = ["full"] }
@@ -58,6 +66,14 @@ regex = "1"
 # cache.rs. Pure-Rust single-process embedded store.
 redb = "2"
 
+# CID parsing for the transport contract only (`transport::cid`). Optional so
+# feeders don't pull it.
+cid = { version = "0.11", optional = true }
+# statvfs(2) for `transport::fsutil::available_bytes`.
+libc = { version = "0.2", optional = true }
+# `ReaderStream` for `transport::file`.
+tokio-util = { version = "0.7", features = ["io"], optional = true }
+
 [dev-dependencies]
 tempfile = "3"
 tokio = { version = "1.40", features = ["full", "test-util"] }

src/lib.rs

@@ -8,6 +8,10 @@
 //! hashing-into-the-blockstore, and the meta-core store-back to itself.
 //!
 //! Deliberately libp2p-free and blockstore-free — see this crate's `Cargo.toml`.
+//!
+//! With the `transport` feature it also carries the **transport-plugin**
+//! contract ([`transport`]): the same split, applied to meta-share — a
+//! protocol-agnostic hull and one sidecar per byte-moving protocol.
 
 // `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
@@ -36,6 +40,8 @@ pub mod plugin;
 pub mod query;
 pub mod query_eval;
 pub mod serve;
+#[cfg(feature = "transport")]
+pub mod transport;
 pub mod types;
 
 pub use budget::RateBudget;

src/transport/cid.rs

@@ -0,0 +1,368 @@
+//! CID parsing and the CID-*shape* knowledge every transport process shares.
+//!
+//! Not protocol logic — these are pure decoders over the multihash standard,
+//! which is the one thing meta-share's hull stays tied to. Moved verbatim from
+//! meta-share (`blockstore.rs` for the parse, `bt.rs` for the torrent shapes) so
+//! the hull and the plugins classify a cid identically: a drift here would route
+//! a torrent cid to the IPFS walker or vice versa.
+
+use std::str::FromStr;
+
+use anyhow::{Context, Result};
+use cid::CidGeneric;
+
+/// Multihash byte ceiling. Must match the gateway and meta-search side — it
+/// feeds beetswap's `Behaviour<const MAX_MULTIHASH_SIZE, B>` in the ipfs plugin.
+pub const MAX_MULTIHASH_SIZE: usize = 64;
+
+/// The CID type every MetaMesh transport speaks.
+pub type MsCid = CidGeneric<MAX_MULTIHASH_SIZE>;
+
+/// Parse a `Record.cid` into an [`MsCid`].
+///
+/// **Strict**: a bare CIDv1 multibase string and nothing else, matching
+/// `cid.Decode` in meta-core's `internal/cid/rank.go`. The `<algo>:<cid>` token
+/// form and CIDv0 (`Qm…`) are both rejected — see meta-share's
+/// `blockstore::parse_record_cid` history for why each one bit.
+pub fn parse_record_cid(cid_str: &str) -> Result<MsCid> {
+    if cid_str.contains(':') {
+        anyhow::bail!(
+            "parse record cid `{cid_str}`: the `<algo>:<cid>` token form was removed; \
+             pass the bare CIDv1 (the multicodec already names the algorithm)"
+        );
+    }
+    let parsed =
+        MsCid::from_str(cid_str).with_context(|| format!("parse record cid `{cid_str}`"))?;
+    if parsed.version() != cid::Version::V1 {
+        anyhow::bail!(
+            "parse record cid `{cid_str}`: only CIDv1 is supported, got {:?}",
+            parsed.version()
+        );
+    }
+    Ok(parsed)
+}
+
+// ---------------------------------------------------------------------------
+// BitTorrent cid shapes
+// ---------------------------------------------------------------------------
+
+/// Multihash code for sha1 — a v1 infohash is the sha1 of the info dict.
+pub const MH_SHA1: u64 = 0x11;
+const BT_V1_INFOHASH_LEN: usize = 20;
+const BT_V2_INFOHASH_LEN: usize = 32;
+/// `btih-v1-file`: one file inside a v1 torrent; digest `<infohash> ‖ varint(index)`.
+pub const BTIH_V1_FILE_CODEC: u64 = 0x1001;
+/// `btih-v2-file`: recognised so the error is precise, not yet fetchable.
+pub const BTIH_V2_FILE_CODEC: u64 = 0x1002;
+
+/// Is this a whole-torrent v1 infohash cid (sha1 multihash, 20 bytes)? The raw
+/// codec `0x55` is shared with IPFS raw leaves, so key off the multihash.
+pub fn is_btih(mscid: &MsCid) -> bool {
+    let mh = mscid.hash();
+    mh.code() == MH_SHA1 && mh.digest().len() == BT_V1_INFOHASH_LEN
+}
+
+/// The raw 20-byte v1 infohash of a `btih:` cid. Errors (a `400` upstream) for
+/// anything that isn't sha1/20.
+pub fn infohash20_from_mscid(mscid: &MsCid) -> Result<[u8; 20]> {
+    let mh = mscid.hash();
+    anyhow::ensure!(
+        mh.code() == MH_SHA1,
+        "not a v1 BitTorrent infohash CID (multihash code 0x{:x}, want 0x11/sha1)",
+        mh.code()
+    );
+    <[u8; 20]>::try_from(mh.digest())
+        .map_err(|_| anyhow::anyhow!("infohash digest is {} bytes, want 20", mh.digest().len()))
+}
+
+/// A decoded per-file torrent cid.
+pub struct BtFileRef {
+    pub infohash20: [u8; 20],
+    pub file_index: usize,
+}
+
+/// Is this a per-file torrent cid (`0x1001` / `0x1002`)?
+pub fn is_btih_file(mscid: &MsCid) -> bool {
+    matches!(mscid.codec(), BTIH_V1_FILE_CODEC | BTIH_V2_FILE_CODEC)
+}
+
+/// Decode a `btih-v1-file` cid into `(infohash, file_index)`; v2 is rejected
+/// pending swarm support.
+pub fn decode_btih_file(mscid: &MsCid) -> Result<BtFileRef> {
+    let infohash_len = match mscid.codec() {
+        BTIH_V1_FILE_CODEC => BT_V1_INFOHASH_LEN,
+        BTIH_V2_FILE_CODEC => {
+            let _ = BT_V2_INFOHASH_LEN; // documents the v2 digest shape
+            anyhow::bail!("btih-v2-file CIDs are not yet supported by the fetch tier");
+        }
+        other => anyhow::bail!("not a btih-file CID (codec 0x{other:x})"),
+    };
+    let digest = mscid.hash().digest();
+    anyhow::ensure!(
+        digest.len() > infohash_len,
+        "btih-file digest is {} bytes, need > {} (infohash + file-index varint)",
+        digest.len(),
+        infohash_len
+    );
+    let (ih, idx_bytes) = digest.split_at(infohash_len);
+    let infohash20 = <[u8; 20]>::try_from(ih)
+        .map_err(|_| anyhow::anyhow!("btih-file infohash slice is not 20 bytes"))?;
+    let (file_index, consumed) = read_uvarint(idx_bytes)?;
+    anyhow::ensure!(
+        consumed == idx_bytes.len(),
+        "btih-file digest has {} trailing byte(s) after the file-index varint",
+        idx_bytes.len() - consumed
+    );
+    let file_index = usize::try_from(file_index)
+        .map_err(|_| anyhow::anyhow!("btih-file index {file_index} overflows usize"))?;
+    Ok(BtFileRef { infohash20, file_index })
+}
+
+/// Minimal unsigned-LEB128 reader → `(value, bytes_consumed)`. Rejects truncated
+/// and overflowing encodings.
+pub fn read_uvarint(bytes: &[u8]) -> Result<(u64, usize)> {
+    let mut result: u64 = 0;
+    let mut shift: u32 = 0;
+    for (i, &b) in bytes.iter().enumerate() {
+        anyhow::ensure!(shift < 64, "varint overflows u64");
+        result |= u64::from(b & 0x7f) << shift;
+        if b & 0x80 == 0 {
+            return Ok((result, i + 1));
+        }
+        shift += 7;
+    }
+    anyhow::bail!("varint is truncated (no terminating byte)")
+}
+
+/// Is this cid a torrent of either shape?
+pub fn is_torrent_cid(mscid: &MsCid) -> bool {
+    is_btih(mscid) || is_btih_file(mscid)
+}
+
+/// A torrent cid as `(infohash, file_index)`; a whole-torrent cid is file 0.
+/// `None` for anything else.
+pub fn torrent_ref_of_cid(cid: &str) -> Option<([u8; 20], usize)> {
+    let mscid = parse_record_cid(cid).ok()?;
+    if let Ok(r) = decode_btih_file(&mscid) {
+        return Some((r.infohash20, r.file_index));
+    }
+    infohash20_from_mscid(&mscid).ok().map(|ih| (ih, 0))
+}
+
+/// A torrent cid's v1 infohash (whole-torrent and per-file share it).
+pub fn infohash_of_cid(cid: &str) -> Option<[u8; 20]> {
+    torrent_ref_of_cid(cid).map(|(ih, _)| ih)
+}
+
+/// Lowercase hex.
+pub fn to_hex(bytes: &[u8]) -> String {
+    use std::fmt::Write;
+    let mut s = String::with_capacity(bytes.len() * 2);
+    for b in bytes {
+        let _ = write!(s, "{b:02x}");
+    }
+    s
+}
+
+/// Lowercase hex of a v1 infohash — the magnet-cache and seed-unit key.
+pub fn infohash20_hex(infohash20: &[u8; 20]) -> String {
+    to_hex(infohash20)
+}
+
+/// Parse a 40-char hex infohash.
+pub fn infohash20_from_hex(s: &str) -> Option<[u8; 20]> {
+    if s.len() != 40 {
+        return None;
+    }
+    let mut out = [0u8; 20];
+    for (i, pair) in s.as_bytes().chunks(2).enumerate() {
+        let hi = (pair[0] as char).to_digit(16)?;
+        let lo = (pair[1] as char).to_digit(16)?;
+        out[i] = ((hi << 4) | lo) as u8;
+    }
+    Some(out)
+}
+
+/// Best-effort MIME from a filename extension; `application/octet-stream` when
+/// unknown so the browser sniffs.
+pub fn content_type_from_name(name: &str) -> String {
+    let ext = name.rsplit('.').next().unwrap_or("").to_ascii_lowercase();
+    let ct = match ext.as_str() {
+        "mkv" => "video/x-matroska",
+        "mp4" | "m4v" => "video/mp4",
+        "webm" => "video/webm",
+        "avi" => "video/x-msvideo",
+        "mov" => "video/quicktime",
+        "ts" | "mts" | "m2ts" => "video/mp2t",
+        "mpg" | "mpeg" => "video/mpeg",
+        "wmv" => "video/x-ms-wmv",
+        "flv" => "video/x-flv",
+        "ogv" => "video/ogg",
+        "mp3" => "audio/mpeg",
+        "flac" => "audio/flac",
+        "aac" => "audio/aac",
+        "ogg" | "oga" => "audio/ogg",
+        "wav" => "audio/wav",
+        "pdf" => "application/pdf",
+        _ => "application/octet-stream",
+    };
+    ct.to_string()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn base32_lower(input: &[u8]) -> String {
+        const ALPHABET: &[u8; 32] = b"abcdefghijklmnopqrstuvwxyz234567";
+        let mut out = String::new();
+        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;
+                out.push(ALPHABET[((buffer >> bits) & 0x1F) as usize] as char);
+            }
+        }
+        if bits > 0 {
+            out.push(ALPHABET[((buffer << (5 - bits)) & 0x1F) as usize] as char);
+        }
+        out
+    }
+
+    fn make_cid(codec: u8, mh_code: u8, digest: &[u8]) -> MsCid {
+        let mut wire = vec![0x01u8, codec, mh_code, digest.len() as u8];
+        wire.extend_from_slice(digest);
+        MsCid::from_str(&format!("b{}", base32_lower(&wire))).expect("cid parse")
+    }
+
+    fn write_uvarint(mut v: u64, out: &mut Vec<u8>) {
+        loop {
+            let mut byte = (v & 0x7f) as u8;
+            v >>= 7;
+            if v != 0 {
+                byte |= 0x80;
+            }
+            out.push(byte);
+            if v == 0 {
+                break;
+            }
+        }
+    }
+
+    fn make_cid_varint(codec: u64, mh_code: u64, digest: &[u8]) -> MsCid {
+        let mut wire = vec![0x01u8];
+        write_uvarint(codec, &mut wire);
+        write_uvarint(mh_code, &mut wire);
+        write_uvarint(digest.len() as u64, &mut wire);
+        wire.extend_from_slice(digest);
+        MsCid::from_str(&format!("b{}", base32_lower(&wire))).expect("cid parse")
+    }
+
+    fn bt_file_digest(infohash: &[u8], index: u64) -> Vec<u8> {
+        let mut d = infohash.to_vec();
+        write_uvarint(index, &mut d);
+        d
+    }
+
+    fn v1_file_cid(infohash: &[u8; 20], index: u64) -> MsCid {
+        make_cid_varint(BTIH_V1_FILE_CODEC, BTIH_V1_FILE_CODEC, &bt_file_digest(infohash, index))
+    }
+
+    #[test]
+    fn btih_cid_is_recognised_and_decodes() {
+        let infohash = [7u8; 20];
+        let cid = make_cid(0x55, MH_SHA1 as u8, &infohash);
+        assert!(is_btih(&cid));
+        assert_eq!(infohash20_from_mscid(&cid).unwrap(), infohash);
+    }
+
+    #[test]
+    fn ipfs_sha256_cid_is_not_btih() {
+        let cid = make_cid(0x55, 0x12, &[0u8; 32]);
+        assert!(!is_btih(&cid));
+        assert!(infohash20_from_mscid(&cid).is_err());
+    }
+
+    #[test]
+    fn dagpb_cid_is_not_btih() {
+        assert!(!is_btih(&make_cid(0x70, 0x12, &[0u8; 32])));
+    }
+
+    #[test]
+    fn btih_v1_file_cid_is_recognised_and_decodes() {
+        let infohash = [7u8; 20];
+        let cid = v1_file_cid(&infohash, 4);
+        assert!(is_btih_file(&cid));
+        assert!(!is_btih(&cid));
+        let r = decode_btih_file(&cid).unwrap();
+        assert_eq!(r.infohash20, infohash);
+        assert_eq!(r.file_index, 4);
+    }
+
+    #[test]
+    fn btih_v1_file_index_zero_and_multibyte_roundtrip() {
+        assert_eq!(decode_btih_file(&v1_file_cid(&[0xab; 20], 0)).unwrap().file_index, 0);
+        assert_eq!(decode_btih_file(&v1_file_cid(&[0x11; 20], 200)).unwrap().file_index, 200);
+    }
+
+    #[test]
+    fn btih_v2_file_cid_recognised_but_rejected() {
+        let cid = make_cid_varint(BTIH_V2_FILE_CODEC, BTIH_V2_FILE_CODEC, &bt_file_digest(&[0x22; 32], 1));
+        assert!(is_btih_file(&cid));
+        assert!(decode_btih_file(&cid).is_err());
+    }
+
+    #[test]
+    fn whole_torrent_cid_is_index_zero_and_not_a_file_cid() {
+        let infohash = [9u8; 20];
+        let cid = make_cid(0x55, MH_SHA1 as u8, &infohash);
+        assert!(is_btih(&cid));
+        assert!(!is_btih_file(&cid));
+        assert!(decode_btih_file(&cid).is_err());
+        assert_eq!(torrent_ref_of_cid(&cid.to_string()), Some((infohash, 0)));
+    }
+
+    #[test]
+    fn decode_btih_file_rejects_missing_index_varint() {
+        let cid = make_cid_varint(BTIH_V1_FILE_CODEC, BTIH_V1_FILE_CODEC, &[3u8; 20]);
+        assert!(decode_btih_file(&cid).is_err());
+    }
+
+    #[test]
+    fn read_uvarint_rejects_truncated_and_reads_multibyte() {
+        assert!(read_uvarint(&[0x80]).is_err());
+        assert_eq!(read_uvarint(&[0xAC, 0x02]).unwrap(), (300, 2));
+        assert_eq!(read_uvarint(&[0x05, 0xFF]).unwrap(), (5, 1));
+    }
+
+    #[test]
+    fn torrent_ref_of_a_file_cid_keeps_its_index() {
+        let cid = v1_file_cid(&[5u8; 20], 3).to_string();
+        assert_eq!(torrent_ref_of_cid(&cid), Some(([5u8; 20], 3)));
+        assert_eq!(infohash_of_cid(&cid), Some([5u8; 20]));
+        assert_eq!(torrent_ref_of_cid("not a cid"), None);
+    }
+
+    #[test]
+    fn hex_round_trips() {
+        let ih = [0xabu8; 20];
+        assert_eq!(infohash20_from_hex(&infohash20_hex(&ih)), Some(ih));
+        assert_eq!(infohash20_from_hex("zz"), None);
+    }
+
+    #[test]
+    fn parse_rejects_token_form_and_cidv0() {
+        assert!(parse_record_cid("midhash256:bafy").is_err());
+        assert!(parse_record_cid("QmYwAPJzv5CZsnA625s3Xf2nemtYgPpHdWEz79ojWnPbdG").is_err());
+    }
+
+    #[test]
+    fn content_type_guesses_common_media() {
+        assert_eq!(content_type_from_name("movie.mkv"), "video/x-matroska");
+        assert_eq!(content_type_from_name("clip.MP4"), "video/mp4");
+        assert_eq!(content_type_from_name("noext"), "application/octet-stream");
+    }
+}

src/transport/client.rs

@@ -0,0 +1,317 @@
+//! The hull side of the contract: an HTTP client for one transport plugin.
+//!
+//! [`RemoteTransport`] covers the common routes and gives typed helpers for a
+//! plugin's extra ones (`get_json`, `post_json`, …). Byte responses come back as
+//! an axum [`Response`] whose body streams straight from the plugin, so the hull
+//! relays a multi-GB range without buffering it.
+
+use std::time::Duration;
+
+use axum::body::Body;
+use axum::http::{HeaderMap, HeaderName, HeaderValue, Method, StatusCode};
+use axum::response::{IntoResponse, Response};
+use serde::de::DeserializeOwned;
+use serde::Serialize;
+
+use super::dto::{
+    Deleted, Health, Job, Jobs, Manifest, ReconcileReport, ReconcileRequest,
+};
+use super::error::ApiError;
+use super::focus::FocusSnapshot;
+
+/// Request headers never forwarded to a plugin (hop-by-hop, or describing the
+/// client's connection to the hull rather than the hull's to the plugin).
+const HOP_BY_HOP: &[&str] = &[
+    "connection",
+    "keep-alive",
+    "proxy-authenticate",
+    "proxy-authorization",
+    "te",
+    "trailer",
+    "transfer-encoding",
+    "upgrade",
+    "host",
+    "content-length",
+];
+
+/// Client for one plugin at `base` (e.g. `http://metashare-torrent:3000`).
+#[derive(Clone, Debug)]
+pub struct RemoteTransport {
+    id: String,
+    base: String,
+    http: reqwest::Client,
+    /// For control calls (jobs, focus, reconcile). `/raw` has no overall
+    /// timeout — a body can legitimately stream for the length of a film.
+    control_timeout: Duration,
+}
+
+impl RemoteTransport {
+    /// `id` names the tier in error messages (`torrent`, `nzb`, `ipfs`).
+    pub fn new(id: impl Into<String>, base: impl Into<String>, http: reqwest::Client) -> Self {
+        Self {
+            id: id.into(),
+            base: base.into().trim_end_matches('/').to_string(),
+            http,
+            control_timeout: Duration::from_secs(30),
+        }
+    }
+
+    pub fn with_control_timeout(mut self, t: Duration) -> Self {
+        self.control_timeout = t;
+        self
+    }
+
+    pub fn id(&self) -> &str {
+        &self.id
+    }
+
+    pub fn base(&self) -> &str {
+        &self.base
+    }
+
+    pub fn url(&self, path: &str) -> String {
+        format!("{}{}", self.base, path)
+    }
+
+    fn unreachable(&self, e: impl std::fmt::Display) -> ApiError {
+        ApiError::Upstream(format!("{} plugin unreachable: {e}", self.id))
+    }
+
+    // ---- common routes ---------------------------------------------------
+
+    pub async fn manifest(&self) -> Result<Manifest, ApiError> {
+        self.get_json("/manifest").await
+    }
+
+    pub async fn health(&self) -> Result<Health, ApiError> {
+        let r = self
+            .http
+            .get(self.url("/health"))
+            .timeout(self.control_timeout)
+            .send()
+            .await
+            .map_err(|e| self.unreachable(e))?;
+        r.json::<Health>().await.map_err(|e| self.unreachable(e))
+    }
+
+    /// `GET /raw/:cid`, forwarding the client's request headers plus `extra`.
+    /// Returns the plugin's response as-is (any status), with its body streaming.
+    pub async fn raw(
+        &self,
+        cid: &str,
+        client_headers: &HeaderMap,
+        extra: &[(&str, String)],
+    ) -> Result<Response, ApiError> {
+        let mut req = self.http.get(self.url(&format!("/raw/{cid}")));
+        req = req.headers(forwardable(client_headers));
+        for (k, v) in extra {
+            req = req.header(*k, v);
+        }
+        let resp = req.send().await.map_err(|e| self.unreachable(e))?;
+        Ok(relay(resp))
+    }
+
+    pub async fn jobs(&self) -> Result<Vec<Job>, ApiError> {
+        Ok(self.get_json::<Jobs>("/jobs").await?.jobs)
+    }
+
+    pub async fn delete(&self, cid: &str, keep: &[String]) -> Result<Deleted, ApiError> {
+        let r = self
+            .http
+            .delete(self.url(&format!("/jobs/{cid}")))
+            .query(&[("keep", keep.join(","))])
+            .timeout(self.control_timeout)
+            .send()
+            .await
+            .map_err(|e| self.unreachable(e))?;
+        decode(r).await
+    }
+
+    pub async fn reconcile(&self, req: &ReconcileRequest) -> Result<ReconcileReport, ApiError> {
+        // A reconcile can delete many units; give it more room than a poll.
+        let r = self
+            .http
+            .post(self.url("/reconcile"))
+            .json(req)
+            .timeout(self.control_timeout * 10)
+            .send()
+            .await
+            .map_err(|e| self.unreachable(e))?;
+        decode(r).await
+    }
+
+    pub async fn push_focus(&self, snap: &FocusSnapshot) -> Result<(), ApiError> {
+        let r = self
+            .http
+            .put(self.url("/focus"))
+            .json(snap)
+            .timeout(self.control_timeout)
+            .send()
+            .await
+            .map_err(|e| self.unreachable(e))?;
+        if r.status().is_success() {
+            Ok(())
+        } else {
+            Err(ApiError::Upstream(format!("{} plugin /focus: {}", self.id, r.status())))
+        }
+    }
+
+    // ---- typed helpers for extra routes ------------------------------------
+
+    pub async fn get_json<T: DeserializeOwned>(&self, path: &str) -> Result<T, ApiError> {
+        let r = self
+            .http
+            .get(self.url(path))
+            .timeout(self.control_timeout)
+            .send()
+            .await
+            .map_err(|e| self.unreachable(e))?;
+        decode(r).await
+    }
+
+    pub async fn post_json<B: Serialize + ?Sized, T: DeserializeOwned>(
+        &self,
+        path: &str,
+        body: &B,
+    ) -> Result<T, ApiError> {
+        self.post_json_timeout(path, body, self.control_timeout).await
+    }
+
+    pub async fn post_json_timeout<B: Serialize + ?Sized, T: DeserializeOwned>(
+        &self,
+        path: &str,
+        body: &B,
+        timeout: Duration,
+    ) -> Result<T, ApiError> {
+        let r = self
+            .http
+            .post(self.url(path))
+            .json(body)
+            .timeout(timeout)
+            .send()
+            .await
+            .map_err(|e| self.unreachable(e))?;
+        decode(r).await
+    }
+
+    pub async fn put_json<B: Serialize + ?Sized>(&self, path: &str, body: &B) -> Result<(), ApiError> {
+        let r = self
+            .http
+            .put(self.url(path))
+            .json(body)
+            .timeout(self.control_timeout)
+            .send()
+            .await
+            .map_err(|e| self.unreachable(e))?;
+        if r.status().is_success() {
+            Ok(())
+        } else {
+            Err(error_from(r).await)
+        }
+    }
+
+    /// Forward an arbitrary request and relay the response verbatim — for the
+    /// public routes a plugin now answers (`/ipfs/:cid`, `/api/peers`, …).
+    pub async fn forward(
+        &self,
+        method: Method,
+        path_and_query: &str,
+        client_headers: &HeaderMap,
+        body: Option<bytes::Bytes>,
+    ) -> Result<Response, ApiError> {
+        let mut req = self
+            .http
+            .request(method, self.url(path_and_query))
+            .headers(forwardable(client_headers));
+        if let Some(b) = body {
+            req = req.body(b);
+        }
+        let resp = req.send().await.map_err(|e| self.unreachable(e))?;
+        Ok(relay(resp))
+    }
+}
+
+/// Client request headers worth forwarding to a plugin.
+fn forwardable(h: &HeaderMap) -> reqwest::header::HeaderMap {
+    let mut out = reqwest::header::HeaderMap::new();
+    for (k, v) in h {
+        if HOP_BY_HOP.contains(&k.as_str()) {
+            continue;
+        }
+        if let (Ok(k), Ok(v)) = (
+            reqwest::header::HeaderName::from_bytes(k.as_str().as_bytes()),
+            reqwest::header::HeaderValue::from_bytes(v.as_bytes()),
+        ) {
+            out.append(k, v);
+        }
+    }
+    out
+}
+
+/// A plugin response as an axum response: same status, same headers (minus
+/// hop-by-hop), body streamed.
+pub fn relay(resp: reqwest::Response) -> Response {
+    let status = StatusCode::from_u16(resp.status().as_u16()).unwrap_or(StatusCode::BAD_GATEWAY);
+    let mut headers = HeaderMap::new();
+    for (k, v) in resp.headers() {
+        let name = k.as_str();
+        if matches!(name, "connection" | "keep-alive" | "transfer-encoding" | "trailer" | "upgrade") {
+            continue;
+        }
+        if let (Ok(k), Ok(v)) = (
+            HeaderName::from_bytes(name.as_bytes()),
+            HeaderValue::from_bytes(v.as_bytes()),
+        ) {
+            headers.append(k, v);
+        }
+    }
+    let body = Body::from_stream(resp.bytes_stream());
+    let mut out = (status, body).into_response();
+    *out.headers_mut() = headers;
+    out
+}
+
+async fn decode<T: DeserializeOwned>(r: reqwest::Response) -> Result<T, ApiError> {
+    if !r.status().is_success() {
+        return Err(error_from(r).await);
+    }
+    r.json::<T>()
+        .await
+        .map_err(|e| ApiError::Upstream(format!("plugin returned malformed JSON: {e}")))
+}
+
+/// Map a plugin error response back onto [`ApiError`], keeping its message so
+/// a relayed error reads the same as the in-process one did.
+pub async fn error_from(r: reqwest::Response) -> ApiError {
+    let status = r.status().as_u16();
+    let msg = r
+        .json::<serde_json::Value>()
+        .await
+        .ok()
+        .and_then(|v| v.get("error").and_then(|e| e.as_str()).map(str::to_string))
+        .unwrap_or_else(|| format!("plugin error {status}"));
+    match status {
+        400 => ApiError::BadRequest(msg),
+        404 => ApiError::NotFound,
+        504 => ApiError::GatewayTimeout(msg),
+        503 if msg == "swarm task is gone" => ApiError::SwarmGone,
+        _ => ApiError::Upstream(msg),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn hop_by_hop_headers_are_not_forwarded() {
+        let mut h = HeaderMap::new();
+        h.insert("range", HeaderValue::from_static("bytes=0-1"));
+        h.insert("host", HeaderValue::from_static("x"));
+        h.insert("connection", HeaderValue::from_static("close"));
+        let f = forwardable(&h);
+        assert!(f.contains_key("range"));
+        assert!(!f.contains_key("host"));
+        assert!(!f.contains_key("connection"));
+    }
+}

src/transport/dto.rs

@@ -0,0 +1,274 @@
+//! Wire types of the transport-plugin contract (hull ⇄ plugin, JSON over HTTP).
+//!
+//! Only what is common to every plugin lives here. A plugin's own extra routes
+//! (ipfs `/share`, torrent `/commits`, …) carry their DTOs in the per-protocol
+//! modules next to this one, so the hull and the plugin still compile against
+//! one definition.
+
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+
+/// Bumped on any breaking change to the common routes below. The hull refuses
+/// a plugin whose `/manifest` reports a different major contract.
+pub const CONTRACT_VERSION: u32 = 1;
+
+/// `X-MetaMesh-Lane`: the hull's lane decision for a `/raw` request.
+pub const HDR_LANE: &str = "x-metamesh-lane";
+/// `X-MetaMesh-Player`: forwarded verbatim from the client — marks a genuine
+/// play (commits, provisional focus).
+pub const HDR_PLAYER: &str = "x-metamesh-player";
+/// `X-MetaMesh-Magnet`: the magnet the hull resolved for a torrent cid.
+pub const HDR_MAGNET: &str = "x-metamesh-magnet";
+/// Seed facts a plugin attaches to a successful `/raw` response, so the hull
+/// records the seed row *before* it relays the body (the monolith wrote the
+/// row inside the fetch, ahead of the first byte). Internal: the hull strips
+/// every `x-metamesh-seed-*` header before relaying.
+pub const HDR_SEED_NAME: &str = "x-metamesh-seed-name";
+pub const HDR_SEED_SIZE: &str = "x-metamesh-seed-size";
+/// Prefix of the seed headers the hull strips from a relayed response.
+pub const INTERNAL_HEADER_PREFIX: &str = "x-metamesh-seed-";
+/// `x-metamesh-meter: egress` on a `/raw` response: the hull floors this body
+/// against the playback focus when it relays it. Set exactly where the
+/// monolith metered (`focus::gated_stream`) — torrent ranges, Usenet file
+/// reads, IPFS dag ranges — and *not* on single-block IPFS answers (posters),
+/// which the monolith never throttled. Internal; stripped by the hull.
+pub const HDR_METER: &str = "x-metamesh-meter";
+pub const METER_EGRESS: &str = "egress";
+
+/// Encode free text (a file name) for an internal header value: every byte
+/// outside a conservative visible-ASCII set becomes `%XX`. Inverse:
+/// [`header_text_decode`].
+pub fn header_text_encode(s: &str) -> String {
+    let mut out = String::with_capacity(s.len());
+    for b in s.bytes() {
+        if b.is_ascii_alphanumeric() || b" ._-()[]+,".contains(&b) {
+            out.push(b as char);
+        } else {
+            out.push_str(&format!("%{b:02X}"));
+        }
+    }
+    out
+}
+
+/// Inverse of [`header_text_encode`]. Invalid escapes pass through literally.
+pub fn header_text_decode(s: &str) -> String {
+    let bytes = s.as_bytes();
+    let mut out = Vec::with_capacity(bytes.len());
+    let mut i = 0;
+    while i < bytes.len() {
+        if bytes[i] == b'%' && i + 2 < bytes.len() {
+            let hi = (bytes[i + 1] as char).to_digit(16);
+            let lo = (bytes[i + 2] as char).to_digit(16);
+            if let (Some(hi), Some(lo)) = (hi, lo) {
+                out.push((hi * 16 + lo) as u8);
+                i += 3;
+                continue;
+            }
+        }
+        out.push(bytes[i]);
+        i += 1;
+    }
+    String::from_utf8_lossy(&out).into_owned()
+}
+
+/// `GET /manifest`.
+#[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq)]
+pub struct Manifest {
+    /// Protocol id: `torrent` | `nzb` | `ipfs` (a competing implementation of a
+    /// protocol reports the same id — that's what makes it a drop-in).
+    pub id: String,
+    /// Implementation name + version, for the dashboard.
+    pub implementation: String,
+    pub version: String,
+    pub contract: u32,
+    pub capabilities: Capabilities,
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct Capabilities {
+    /// Can fetch bytes for a cid (`/raw`).
+    pub fetch: bool,
+    /// Serves what it fetched to other peers (seeds).
+    pub share: bool,
+}
+
+/// `GET /health`.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct Health {
+    pub ok: bool,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub detail: Option<Value>,
+}
+
+impl Health {
+    pub fn ok() -> Self {
+        Self { ok: true, detail: None }
+    }
+}
+
+/// One unit of bytes a plugin holds — the eviction unit as the plugin knows it
+/// (a torrent, an NZB job). `GET /jobs`.
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq)]
+pub struct Job {
+    /// Stable unit key (a torrent's infohash hex, an nzb job's cid).
+    pub id: String,
+    /// The cids this unit answers for, as far as the plugin can name them.
+    #[serde(default)]
+    pub cids: Vec<String>,
+    /// Bytes on disk right now.
+    pub bytes_on_disk: u64,
+    /// Every selected byte is present.
+    pub complete: bool,
+    /// Downloading in the background on its own (exempt from eviction).
+    pub filling: bool,
+    /// Plugin-defined state label (`live`, `paused`, `streaming`, …).
+    #[serde(default)]
+    pub state: String,
+    /// Plugin-specific detail the hull renders opaquely (live transfer stats).
+    #[serde(default, skip_serializing_if = "Value::is_null")]
+    pub extra: Value,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default)]
+pub struct Jobs {
+    pub jobs: Vec<Job>,
+}
+
+/// `DELETE /jobs/:cid` query: the *other* cids of the same unit the hull still
+/// wants (sibling files of one torrent). The plugin keeps those.
+#[derive(Serialize, Deserialize, Clone, Debug, Default)]
+pub struct DeleteQuery {
+    /// Comma-separated cids.
+    #[serde(default)]
+    pub keep: String,
+}
+
+impl DeleteQuery {
+    pub fn keep_list(&self) -> Vec<String> {
+        self.keep
+            .split(',')
+            .map(str::trim)
+            .filter(|s| !s.is_empty())
+            .map(str::to_string)
+            .collect()
+    }
+}
+
+/// `DELETE /jobs/:cid` → the bytes actually unlinked, which is what the hull's
+/// eviction budget credits.
+#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct Deleted {
+    pub bytes_freed: u64,
+}
+
+/// `POST /reconcile`: delete every unit no wanted cid names, and any on-disk
+/// leftover no unit owns older than `grace_secs`.
+#[derive(Serialize, Deserialize, Clone, Debug, Default)]
+pub struct ReconcileRequest {
+    pub wanted: Vec<String>,
+    pub grace_secs: u64,
+    /// Only meaningful for plugins whose state lives in a lazily-started
+    /// session (torrent): bring it up even if nothing asked for bytes yet,
+    /// because persisted units are on disk and only the session can say whose.
+    #[serde(default)]
+    pub start_session: bool,
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct ReconcileReport {
+    /// Units deleted because nothing wanted them.
+    pub units_deleted: u64,
+    pub unit_bytes_freed: u64,
+    /// On-disk leftovers removed.
+    pub orphans_deleted: u64,
+    pub orphan_bytes_freed: u64,
+    /// The plugin couldn't tell what it owns (session down): nothing deleted.
+    #[serde(default)]
+    pub skipped: bool,
+}
+
+/// A file inside a promoted cache container.
+#[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq)]
+pub struct PromotedFile {
+    /// Path relative to the container directory.
+    pub rel: String,
+    /// Absolute path on the shared `/data` volume.
+    pub path: String,
+    pub size: u64,
+}
+
+/// Plugin → hull notifications, `POST {hull}/internal/events` (see
+/// [`super::hull::HullClient`]). The hull acknowledges once it has acted, so a
+/// plugin that must not proceed before the hull indexed something can wait.
+#[derive(Serialize, Deserialize, Clone, Debug, PartialEq)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum Event {
+    /// The plugin now holds bytes under `cid`: the hull upserts a seed row.
+    Seed {
+        cid: String,
+        /// `torrent` | `ipfs` (the hull's `SeedKind`).
+        kind: String,
+        #[serde(default)]
+        name: String,
+        #[serde(default)]
+        size_bytes: u64,
+        /// The container the bytes live in, when there is one.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        container: Option<String>,
+    },
+    /// A completed fetch was published as a cache container (`rename(2)` into
+    /// `cache/`). The hull indexes it in the material index.
+    Promoted {
+        container: String,
+        cids: Vec<String>,
+        files: Vec<PromotedFile>,
+    },
+    /// The content behind `cid` turned out to be an archive: the hull marks
+    /// the meta-core record `fileType=archive`.
+    Archive { cid: String },
+    /// Bytes behind these cids are gone: the hull drops their seed rows.
+    Vanished { cids: Vec<String> },
+    /// What the plugin knows about `cid` changed (a new availability verdict):
+    /// the hull drops its cached probe grade.
+    ProbeStale { cid: String },
+    /// A fetch finished and its file is complete at `path` (still under
+    /// `tmp/`). The hull indexes it for local-first reads and decides what to
+    /// do next (the Usenet → IPFS re-seed).
+    Ready {
+        cid: String,
+        path: String,
+        size: u64,
+    },
+    /// Any event type this hull doesn't know yet. Ignored.
+    #[serde(other)]
+    Unknown,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn header_text_round_trips_unicode_and_reserved() {
+        for s in ["Plain Name.mkv", "Été 100% \"quoted\"/x.mkv", "", "%", "a%2"] {
+            assert_eq!(header_text_decode(&header_text_encode(s)), s);
+        }
+        assert!(header_text_encode("é/").is_ascii());
+    }
+
+    #[test]
+    fn delete_query_splits_and_trims() {
+        let q = DeleteQuery { keep: " a, ,b,".into() };
+        assert_eq!(q.keep_list(), vec!["a".to_string(), "b".to_string()]);
+        assert!(DeleteQuery::default().keep_list().is_empty());
+    }
+
+    #[test]
+    fn events_are_tagged_and_unknown_tolerated() {
+        let e = Event::Archive { cid: "b".into() };
+        let j = serde_json::to_string(&e).unwrap();
+        assert_eq!(j, r#"{"type":"archive","cid":"b"}"#);
+        let u: Event = serde_json::from_str(r#"{"type":"from_the_future","x":1}"#).unwrap();
+        assert_eq!(u, Event::Unknown);
+    }
+}

src/transport/error.rs

@@ -0,0 +1,60 @@
+//! The one error type every transport HTTP surface answers with.
+//!
+//! Moved verbatim from meta-share `crates/meta-share/src/api/errors.rs`. It lives
+//! here, not in the hull, because a plugin's error body is relayed to meta-share's
+//! public API unchanged: the JSON shape `{ "error": "<message>" }` and the
+//! status per variant are the contract every dashboard and peer depends on, and
+//! the hull and its plugins must not be able to drift on it.
+
+use axum::{http::StatusCode, response::IntoResponse, Json};
+
+#[derive(Debug)]
+pub enum ApiError {
+    BadRequest(String),
+    NotFound,
+    SwarmGone,
+    /// This peer is consumer-only (no `META_CORE_URL`) so it can't satisfy a
+    /// raw-bytes request directly. The cross-peer proxy turns this into a
+    /// clearer 502 with hint text; the direct route returns 503.
+    ConsumerOnly,
+    Upstream(String),
+    /// Gateway-side compute exceeded its deadline. Distinct from `Upstream` so
+    /// the dashboard can surface a clearer "the source fetcher is slow" message
+    /// instead of a generic 502.
+    GatewayTimeout(String),
+}
+
+impl IntoResponse for ApiError {
+    fn into_response(self) -> axum::response::Response {
+        let (status, msg) = match self {
+            ApiError::BadRequest(m) => (StatusCode::BAD_REQUEST, m),
+            ApiError::NotFound => (StatusCode::NOT_FOUND, "not found".into()),
+            ApiError::SwarmGone => (
+                StatusCode::SERVICE_UNAVAILABLE,
+                "swarm task is gone".into(),
+            ),
+            ApiError::ConsumerOnly => (
+                StatusCode::SERVICE_UNAVAILABLE,
+                "this peer has no META_CORE_URL; raw bytes are only served by the source peer".into(),
+            ),
+            ApiError::Upstream(m) => (StatusCode::BAD_GATEWAY, m),
+            ApiError::GatewayTimeout(m) => (StatusCode::GATEWAY_TIMEOUT, m),
+        };
+        (status, Json(serde_json::json!({ "error": msg }))).into_response()
+    }
+}
+
+impl std::fmt::Display for ApiError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            ApiError::BadRequest(m) => write!(f, "bad request: {m}"),
+            ApiError::NotFound => write!(f, "not found"),
+            ApiError::SwarmGone => write!(f, "swarm task is gone"),
+            ApiError::ConsumerOnly => write!(f, "consumer-only peer"),
+            ApiError::Upstream(m) => write!(f, "upstream: {m}"),
+            ApiError::GatewayTimeout(m) => write!(f, "timeout: {m}"),
+        }
+    }
+}
+
+impl std::error::Error for ApiError {}

src/transport/file.rs

@@ -0,0 +1,114 @@
+//! Serving a byte range straight off a file on the shared `/data` volume.
+//!
+//! The hull's local-first read and the usenet plugin's materialised/partial
+//! file reads answer with the same headers, so they share this. Moved from
+//! meta-share's `nzb::serve::serve_file_range`, minus the focus metering: the
+//! caller meters the body (the hull wraps every response it serves, a plugin
+//! doesn't meter its `/raw` at all — the hull does on relay).
+
+use std::path::Path;
+
+use axum::body::Body;
+use axum::http::{header, HeaderMap, HeaderValue, StatusCode};
+use axum::response::{IntoResponse, Response};
+use tokio::io::{AsyncReadExt, AsyncSeekExt};
+use tokio_util::io::ReaderStream;
+
+use super::range::{parse_range_header, RangeOutcome};
+
+/// Stream a complete on-disk file of `total` bytes, honouring a single
+/// `Range:` (206) else 200; 416 for an unsatisfiable range.
+///
+/// `Err` carries the IO error that stopped the open/seek, so a caller can tell
+/// "the file moved under us" apart from a served response.
+pub async fn serve_file_range(
+    path: &Path,
+    total: u64,
+    range: Option<&HeaderValue>,
+) -> std::io::Result<Response> {
+    let ctype = content_type(path);
+    match parse_range_header(range, total) {
+        RangeOutcome::Unsatisfiable => {
+            let mut headers = HeaderMap::new();
+            headers.insert(
+                header::CONTENT_RANGE,
+                HeaderValue::from_str(&format!("bytes */{total}")).unwrap(),
+            );
+            Ok((StatusCode::RANGE_NOT_SATISFIABLE, headers).into_response())
+        }
+        RangeOutcome::Ok { start, end } => {
+            let len = end - start + 1;
+            let mut file = tokio::fs::File::open(path).await?;
+            file.seek(std::io::SeekFrom::Start(start)).await?;
+            let body = Body::from_stream(ReaderStream::new(file.take(len)));
+            let mut headers = HeaderMap::new();
+            headers.insert(header::CONTENT_TYPE, HeaderValue::from_static(ctype));
+            headers.insert(header::ACCEPT_RANGES, HeaderValue::from_static("bytes"));
+            headers.insert(header::CONTENT_LENGTH, HeaderValue::from(len));
+            headers.insert(
+                header::CONTENT_RANGE,
+                HeaderValue::from_str(&format!("bytes {start}-{end}/{total}")).unwrap(),
+            );
+            Ok((StatusCode::PARTIAL_CONTENT, headers, body).into_response())
+        }
+        // Absent / Unsupported → full 200.
+        _ => {
+            let file = tokio::fs::File::open(path).await?;
+            let body = Body::from_stream(ReaderStream::new(file));
+            let mut headers = HeaderMap::new();
+            headers.insert(header::CONTENT_TYPE, HeaderValue::from_static(ctype));
+            headers.insert(header::ACCEPT_RANGES, HeaderValue::from_static("bytes"));
+            headers.insert(header::CONTENT_LENGTH, HeaderValue::from(total));
+            Ok((StatusCode::OK, headers, body).into_response())
+        }
+    }
+}
+
+/// Coarse content-type from extension — enough for the player to pick a demuxer.
+pub fn content_type(path: &Path) -> &'static str {
+    match path
+        .extension()
+        .and_then(|e| e.to_str())
+        .map(|e| e.to_ascii_lowercase())
+        .as_deref()
+    {
+        Some("mkv") => "video/x-matroska",
+        Some("mp4") | Some("m4v") => "video/mp4",
+        Some("webm") => "video/webm",
+        Some("avi") => "video/x-msvideo",
+        Some("mov") => "video/quicktime",
+        Some("ts") | Some("m2ts") => "video/mp2t",
+        _ => "application/octet-stream",
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    async fn body(r: Response) -> Vec<u8> {
+        axum::body::to_bytes(r.into_body(), 1 << 20).await.unwrap().to_vec()
+    }
+
+    #[tokio::test]
+    async fn range_full_and_unsatisfiable() {
+        let p = std::env::temp_dir().join(format!("file-range-{}.mkv", std::process::id()));
+        std::fs::write(&p, b"0123456789").unwrap();
+        let r = serve_file_range(&p, 10, Some(&HeaderValue::from_static("bytes=2-4"))).await.unwrap();
+        assert_eq!(r.status(), StatusCode::PARTIAL_CONTENT);
+        assert_eq!(r.headers()[header::CONTENT_RANGE], "bytes 2-4/10");
+        assert_eq!(r.headers()[header::CONTENT_TYPE], "video/x-matroska");
+        assert_eq!(body(r).await, b"234");
+        let r = serve_file_range(&p, 10, None).await.unwrap();
+        assert_eq!(r.status(), StatusCode::OK);
+        assert_eq!(body(r).await, b"0123456789");
+        let r = serve_file_range(&p, 10, Some(&HeaderValue::from_static("bytes=20-"))).await.unwrap();
+        assert_eq!(r.status(), StatusCode::RANGE_NOT_SATISFIABLE);
+        assert_eq!(r.headers()[header::CONTENT_RANGE], "bytes */10");
+        let _ = std::fs::remove_file(&p);
+        assert_eq!(
+            serve_file_range(&p, 10, None).await.unwrap_err().kind(),
+            std::io::ErrorKind::NotFound
+        );
+    }
+}

src/transport/focus.rs

@@ -0,0 +1,464 @@
+//! Playback focus, as a transport plugin sees it.
+//!
+//! meta-share's hull owns the focus *policy*: `POST/DELETE /api/playback/:cid`,
+//! the leases, the TTL reaper, which sibling CIDs make up a title. A plugin only
+//! needs to *obey* it: pause a torrent that isn't focused, park a background
+//! fill, floor a non-focused ingress. So the hull pushes a [`FocusSnapshot`]
+//! (`PUT /focus`) on every transition, plus periodically so a restarted plugin
+//! recovers, and the plugin answers its gates from a local [`FocusView`].
+//!
+//! The read side mirrors meta-share's `focus::FocusGate` method for method, so
+//! code moved out of meta-share keeps its call sites. Two things are
+//! deliberately different from the in-process gate:
+//!
+//! - **No leases.** A view never expires an entry on its own; only the next
+//!   snapshot changes it. The hull's reaper is the single source of truth.
+//! - **Per-process buckets.** `throttle_in`/`throttle_out` floor non-focused
+//!   traffic at `bg_rate_bytes` *per process*, where the monolith shared one
+//!   bucket across tiers. HTTP response bodies are still metered once, by the
+//!   hull, when it relays a plugin's `/raw` — plugins must not meter their own
+//!   `/raw` egress, or a peer pulling through the hull pays twice.
+
+use std::collections::HashSet;
+use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+
+use serde::{Deserialize, Serialize};
+use tokio::sync::watch;
+
+/// How often a parked background task re-checks whether the focus cleared.
+/// Same value as meta-share's `focus::PARK_POLL`.
+pub const PARK_POLL: Duration = Duration::from_millis(750);
+
+/// Burst allowance as a multiple of the per-second rate, floored at
+/// [`MIN_BURST_BYTES`]. Same values as meta-share's `focus.rs`.
+const BURST_SECS: u64 = 2;
+const MIN_BURST_BYTES: u64 = 512 * 1024;
+
+/// Which bandwidth lane a byte flow belongs to.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum Lane {
+    /// Belongs to a title a viewer is playing right now — never gated.
+    Focused,
+    /// Everything else. Parked or throttled while any title is focused;
+    /// unrestricted otherwise.
+    Background,
+}
+
+impl Lane {
+    pub fn is_focused(self) -> bool {
+        matches!(self, Lane::Focused)
+    }
+
+    /// Wire value of the `X-MetaMesh-Lane` header.
+    pub fn as_header(self) -> &'static str {
+        match self {
+            Lane::Focused => "focused",
+            Lane::Background => "background",
+        }
+    }
+
+    /// Parse the `X-MetaMesh-Lane` header. An absent or unknown value is
+    /// `Focused`: a request the hull didn't classify must never be starved.
+    pub fn from_header(v: Option<&str>) -> Lane {
+        match v {
+            Some("background") => Lane::Background,
+            _ => Lane::Focused,
+        }
+    }
+}
+
+/// One focused title: its group key plus every identity it is reachable under.
+#[derive(Clone, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
+pub struct FocusTitle {
+    /// The record's canonical cid — the stable per-title key.
+    pub group: String,
+    /// Every sibling cid of the title.
+    #[serde(default)]
+    pub cids: Vec<String>,
+    /// The torrent infohashes among those siblings, lowercase hex, pre-decoded
+    /// by the hull so a torrent plugin never re-parses cids on a transition.
+    #[serde(default)]
+    pub infohashes: Vec<String>,
+}
+
+/// The whole focus state, pushed by the hull on `PUT /focus`.
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct FocusSnapshot {
+    /// Master switch (`META_SHARE_FOCUS_ENABLED`). Off → every gate is a no-op.
+    pub enabled: bool,
+    /// Bytes/s a non-focused flow is floored at. `0` → hard park.
+    pub bg_rate_bytes: u64,
+    #[serde(default)]
+    pub titles: Vec<FocusTitle>,
+}
+
+impl Default for FocusSnapshot {
+    fn default() -> Self {
+        Self {
+            enabled: true,
+            bg_rate_bytes: 10 * 1024,
+            titles: Vec::new(),
+        }
+    }
+}
+
+struct Inner {
+    titles: Vec<FocusTitle>,
+    cids: HashSet<String>,
+    infohashes: HashSet<[u8; 20]>,
+}
+
+/// A plugin's local replica of the hull's focus gate.
+pub struct FocusView {
+    inner: Mutex<Inner>,
+    /// `titles.len()` as an atomic, so the resting state is one relaxed load.
+    n: AtomicUsize,
+    enabled: std::sync::atomic::AtomicBool,
+    rate: AtomicU64,
+    out_bucket: TokenBucket,
+    in_bucket: TokenBucket,
+    /// Bumped on every applied snapshot that changed something. A torrent
+    /// plugin watches it to re-run its pause/unpause pass.
+    changed: watch::Sender<u64>,
+}
+
+impl FocusView {
+    pub fn new() -> Arc<Self> {
+        let d = FocusSnapshot::default();
+        let (changed, _) = watch::channel(0);
+        let burst = (d.bg_rate_bytes * BURST_SECS).max(MIN_BURST_BYTES);
+        Arc::new(Self {
+            inner: Mutex::new(Inner {
+                titles: Vec::new(),
+                cids: HashSet::new(),
+                infohashes: HashSet::new(),
+            }),
+            n: AtomicUsize::new(0),
+            enabled: std::sync::atomic::AtomicBool::new(d.enabled),
+            rate: AtomicU64::new(d.bg_rate_bytes),
+            out_bucket: TokenBucket::new(d.bg_rate_bytes, burst),
+            in_bucket: TokenBucket::new(d.bg_rate_bytes, burst),
+            changed,
+        })
+    }
+
+    /// Replace the state with `snap`. Returns `true` when anything changed (and
+    /// notifies [`subscribe`](Self::subscribe)rs).
+    pub fn apply(&self, snap: FocusSnapshot) -> bool {
+        let mut cids = HashSet::new();
+        let mut infohashes = HashSet::new();
+        for t in &snap.titles {
+            cids.extend(t.cids.iter().cloned());
+            cids.insert(t.group.clone());
+            infohashes.extend(t.infohashes.iter().filter_map(|h| decode_hex20(h)));
+        }
+        let changed = {
+            let mut inner = self.inner.lock().unwrap();
+            let changed = inner.titles != snap.titles
+                || self.enabled.load(Ordering::Relaxed) != snap.enabled
+                || self.rate.load(Ordering::Relaxed) != snap.bg_rate_bytes;
+            inner.titles = snap.titles;
+            inner.cids = cids;
+            inner.infohashes = infohashes;
+            self.n.store(inner.titles.len(), Ordering::Relaxed);
+            changed
+        };
+        self.enabled.store(snap.enabled, Ordering::Relaxed);
+        if self.rate.swap(snap.bg_rate_bytes, Ordering::Relaxed) != snap.bg_rate_bytes {
+            let burst = (snap.bg_rate_bytes * BURST_SECS).max(MIN_BURST_BYTES);
+            self.out_bucket.reconfigure(snap.bg_rate_bytes, burst);
+            self.in_bucket.reconfigure(snap.bg_rate_bytes, burst);
+        }
+        if changed {
+            self.changed.send_modify(|g| *g += 1);
+        }
+        changed
+    }
+
+    /// The current state, as the hull last pushed it.
+    pub fn snapshot(&self) -> FocusSnapshot {
+        FocusSnapshot {
+            enabled: self.enabled.load(Ordering::Relaxed),
+            bg_rate_bytes: self.rate.load(Ordering::Relaxed),
+            titles: self.inner.lock().unwrap().titles.clone(),
+        }
+    }
+
+    /// Fires after every snapshot that changed something.
+    pub fn subscribe(&self) -> watch::Receiver<u64> {
+        self.changed.subscribe()
+    }
+
+    pub fn bg_rate_bytes(&self) -> u64 {
+        self.rate.load(Ordering::Relaxed)
+    }
+
+    /// Is *any* title focused?
+    pub fn any(&self) -> bool {
+        self.enabled.load(Ordering::Relaxed) && self.n.load(Ordering::Relaxed) > 0
+    }
+
+    pub fn count(&self) -> usize {
+        self.n.load(Ordering::Relaxed)
+    }
+
+    /// Which lane does a byte flow for `cid` belong to? Same rule as the hull:
+    /// nothing focused → everything is `Focused`.
+    pub fn lane(&self, cid: &str) -> Lane {
+        if !self.any() {
+            return Lane::Focused;
+        }
+        if self.inner.lock().unwrap().cids.contains(cid) {
+            Lane::Focused
+        } else {
+            Lane::Background
+        }
+    }
+
+    pub fn is_cid_focused(&self, cid: &str) -> bool {
+        self.inner.lock().unwrap().cids.contains(cid)
+    }
+
+    pub fn focused_infohashes(&self) -> HashSet<[u8; 20]> {
+        self.inner.lock().unwrap().infohashes.clone()
+    }
+
+    pub fn is_infohash_focused(&self, infohash: &[u8; 20]) -> bool {
+        self.inner.lock().unwrap().infohashes.contains(infohash)
+    }
+
+    /// Park a background task while any title is focused.
+    pub async fn wait_turn(&self, lane: Lane) {
+        if lane.is_focused() {
+            return;
+        }
+        while self.any() {
+            tokio::time::sleep(PARK_POLL).await;
+        }
+    }
+
+    /// [`wait_turn`](Self::wait_turn) with a deadline; `false` if the focus
+    /// outlasted `budget`.
+    pub async fn wait_turn_timeout(&self, lane: Lane, budget: Duration) -> bool {
+        if lane.is_focused() {
+            return true;
+        }
+        let deadline = Instant::now() + budget;
+        while self.any() {
+            if Instant::now() >= deadline {
+                return false;
+            }
+            tokio::time::sleep(PARK_POLL).await;
+        }
+        true
+    }
+
+    /// Floor a non-focused egress of `n` bytes (bitswap block service).
+    pub async fn throttle_out(&self, lane: Lane, n: u64) {
+        if lane.is_focused() || !self.any() {
+            return;
+        }
+        if self.bg_rate_bytes() == 0 {
+            self.wait_turn(lane).await;
+            return;
+        }
+        self.out_bucket.consume(n).await;
+    }
+
+    /// Floor a non-focused ingress of `n` bytes that a caller is waiting on.
+    pub async fn throttle_in(&self, lane: Lane, n: u64) {
+        if lane.is_focused() || !self.any() {
+            return;
+        }
+        if self.bg_rate_bytes() == 0 {
+            self.wait_turn(lane).await;
+            return;
+        }
+        self.in_bucket.consume(n).await;
+    }
+}
+
+/// Meter a byte stream against a view's egress floor. Same shape as
+/// meta-share's `focus::gated_stream`.
+pub fn gated_stream<S, B, E>(
+    stream: S,
+    focus: Arc<FocusView>,
+    lane: Lane,
+) -> impl futures::Stream<Item = Result<B, E>>
+where
+    S: futures::Stream<Item = Result<B, E>>,
+    B: AsRef<[u8]>,
+{
+    use futures::StreamExt;
+    stream.then(move |item| {
+        let focus = Arc::clone(&focus);
+        async move {
+            if let Ok(bytes) = &item {
+                focus.throttle_out(lane, bytes.as_ref().len() as u64).await;
+            }
+            item
+        }
+    })
+}
+
+/// Lowercase hex of a 20-byte infohash — the wire form in [`FocusTitle`].
+pub fn infohash_hex(ih: &[u8; 20]) -> String {
+    let mut s = String::with_capacity(40);
+    for b in ih {
+        s.push_str(&format!("{b:02x}"));
+    }
+    s
+}
+
+fn decode_hex20(s: &str) -> Option<[u8; 20]> {
+    if s.len() != 40 {
+        return None;
+    }
+    let mut out = [0u8; 20];
+    for (i, pair) in s.as_bytes().chunks(2).enumerate() {
+        let hi = (pair[0] as char).to_digit(16)?;
+        let lo = (pair[1] as char).to_digit(16)?;
+        out[i] = ((hi << 4) | lo) as u8;
+    }
+    Some(out)
+}
+
+/// Leaky bucket that always waits (never errors on `n > burst`). Copied from
+/// meta-share's `focus.rs`, plus `reconfigure` for a rate pushed at runtime.
+struct TokenBucket {
+    state: Mutex<BucketState>,
+}
+
+struct BucketState {
+    tokens: f64,
+    last: Instant,
+    rate: u64,
+    burst: u64,
+}
+
+impl TokenBucket {
+    fn new(rate: u64, burst: u64) -> Self {
+        Self {
+            state: Mutex::new(BucketState {
+                tokens: burst as f64,
+                last: Instant::now(),
+                rate,
+                burst,
+            }),
+        }
+    }
+
+    fn reconfigure(&self, rate: u64, burst: u64) {
+        let mut s = self.state.lock().unwrap();
+        s.rate = rate;
+        s.burst = burst;
+        s.tokens = s.tokens.min(burst as f64);
+    }
+
+    async fn consume(&self, n: u64) {
+        let wait = {
+            let mut s = self.state.lock().unwrap();
+            if s.rate == 0 || n == 0 {
+                return;
+            }
+            let now = Instant::now();
+            let elapsed = now.duration_since(s.last).as_secs_f64();
+            s.last = now;
+            s.tokens = (s.tokens + elapsed * s.rate as f64).min(s.burst as f64);
+            s.tokens -= n as f64;
+            if s.tokens >= 0.0 {
+                return;
+            }
+            Duration::from_secs_f64(-s.tokens / s.rate as f64)
+        };
+        tokio::time::sleep(wait).await;
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn snap(titles: Vec<FocusTitle>) -> FocusSnapshot {
+        FocusSnapshot {
+            titles,
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn nothing_focused_means_every_flow_is_focused() {
+        let v = FocusView::new();
+        assert!(!v.any());
+        assert_eq!(v.lane("bafyanything"), Lane::Focused);
+    }
+
+    #[test]
+    fn siblings_and_group_are_focused_others_background() {
+        let v = FocusView::new();
+        v.apply(snap(vec![FocusTitle {
+            group: "g".into(),
+            cids: vec!["a".into(), "b".into()],
+            infohashes: vec![infohash_hex(&[7u8; 20])],
+        }]));
+        assert!(v.any());
+        assert_eq!(v.lane("a"), Lane::Focused);
+        assert_eq!(v.lane("g"), Lane::Focused);
+        assert_eq!(v.lane("z"), Lane::Background);
+        assert!(v.is_infohash_focused(&[7u8; 20]));
+        assert!(!v.is_infohash_focused(&[8u8; 20]));
+    }
+
+    #[test]
+    fn disabled_gate_focuses_nothing() {
+        let v = FocusView::new();
+        v.apply(FocusSnapshot {
+            enabled: false,
+            bg_rate_bytes: 1,
+            titles: vec![FocusTitle {
+                group: "g".into(),
+                ..Default::default()
+            }],
+        });
+        assert!(!v.any());
+        assert_eq!(v.lane("z"), Lane::Focused);
+    }
+
+    #[test]
+    fn apply_reports_change_only_on_change() {
+        let v = FocusView::new();
+        let rx = v.subscribe();
+        let s = snap(vec![FocusTitle {
+            group: "g".into(),
+            ..Default::default()
+        }]);
+        assert!(v.apply(s.clone()));
+        assert!(!v.apply(s));
+        assert_eq!(*rx.borrow(), 1);
+        assert!(v.apply(snap(vec![])));
+        assert_eq!(*rx.borrow(), 2);
+    }
+
+    #[test]
+    fn lane_header_round_trip_defaults_to_focused() {
+        assert_eq!(Lane::from_header(Some("background")), Lane::Background);
+        assert_eq!(Lane::from_header(Some("focused")), Lane::Focused);
+        assert_eq!(Lane::from_header(None), Lane::Focused);
+        assert_eq!(Lane::from_header(Some("junk")), Lane::Focused);
+        assert_eq!(Lane::from_header(Some(Lane::Background.as_header())), Lane::Background);
+    }
+
+    #[test]
+    fn snapshot_json_round_trips() {
+        let s = snap(vec![FocusTitle {
+            group: "g".into(),
+            cids: vec!["a".into()],
+            infohashes: vec![infohash_hex(&[1u8; 20])],
+        }]);
+        let back: FocusSnapshot = serde_json::from_str(&serde_json::to_string(&s).unwrap()).unwrap();
+        assert_eq!(back, s);
+    }
+}

src/transport/fsutil.rs

@@ -0,0 +1,230 @@
+//! Filesystem helpers shared by the hull and the plugins that own a corner of
+//! the shared `/data` volume: free space, on-disk size, and the grace-period
+//! removal both `reclaim` passes (cache containers in the hull, `bt-temp/` in
+//! the torrent plugin) are built on. Moved verbatim from meta-share
+//! `reclaim.rs` / `eviction.rs`.
+
+use std::ffi::OsString;
+use std::path::{Path, PathBuf};
+use std::time::{Duration, SystemTime};
+
+use tracing::warn;
+
+/// Free space (bytes available to an unprivileged process) on the filesystem
+/// backing `path`; `None` when the query fails or the platform isn't supported.
+#[cfg(unix)]
+pub fn available_bytes(path: &Path) -> Option<u64> {
+    use std::os::unix::ffi::OsStrExt;
+    let c_path = std::ffi::CString::new(path.as_os_str().as_bytes()).ok()?;
+    // SAFETY: `c_path` is a valid NUL-terminated path; `statvfs` only writes
+    // into the zeroed-out struct and returns 0 on success.
+    let mut stat: libc::statvfs = unsafe { std::mem::zeroed() };
+    let rc = unsafe { libc::statvfs(c_path.as_ptr(), &mut stat) };
+    if rc != 0 {
+        return None;
+    }
+    #[allow(clippy::unnecessary_cast)]
+    Some((stat.f_bavail as u64).saturating_mul(stat.f_frsize as u64))
+}
+
+#[cfg(not(unix))]
+pub fn available_bytes(_path: &Path) -> Option<u64> {
+    None
+}
+
+/// Apparent bytes under `dir` (file lengths, subdirectories walked, symlinks
+/// skipped). `0` when absent. Blocking.
+pub fn dir_size_blocking(dir: &Path) -> u64 {
+    let mut total = 0u64;
+    let mut stack = vec![dir.to_path_buf()];
+    while let Some(d) = stack.pop() {
+        let Ok(rd) = std::fs::read_dir(&d) else { continue };
+        for entry in rd.flatten() {
+            match entry.file_type() {
+                Ok(ft) if ft.is_dir() => stack.push(entry.path()),
+                Ok(ft) if ft.is_file() => {
+                    if let Ok(meta) = entry.metadata() {
+                        total = total.saturating_add(meta.len());
+                    }
+                }
+                _ => {}
+            }
+        }
+    }
+    total
+}
+
+/// Bytes `path` actually occupies: allocated blocks, not apparent length (a
+/// streaming-only torrent's file is sparse and its length is the whole film).
+/// Blocking.
+pub fn disk_usage_blocking(path: &Path) -> u64 {
+    let mut total = 0u64;
+    let mut stack = vec![path.to_path_buf()];
+    while let Some(p) = stack.pop() {
+        let Ok(meta) = std::fs::symlink_metadata(&p) else { continue };
+        if meta.is_dir() {
+            if let Ok(rd) = std::fs::read_dir(&p) {
+                stack.extend(rd.flatten().map(|e| e.path()));
+            }
+        } else if meta.is_file() {
+            total = total.saturating_add(allocated(&meta));
+        }
+    }
+    total
+}
+
+#[cfg(unix)]
+fn allocated(meta: &std::fs::Metadata) -> u64 {
+    use std::os::unix::fs::MetadataExt;
+    meta.blocks().saturating_mul(512)
+}
+
+#[cfg(not(unix))]
+fn allocated(meta: &std::fs::Metadata) -> u64 {
+    meta.len()
+}
+
+/// Top-level entries of `dir` that `is_owned` doesn't claim.
+pub async fn unowned_entries(
+    dir: &Path,
+    is_owned: impl Fn(&std::ffi::OsStr) -> bool,
+) -> Vec<(PathBuf, OsString)> {
+    let mut out = Vec::new();
+    let Ok(mut rd) = tokio::fs::read_dir(dir).await else {
+        return out;
+    };
+    while let Ok(Some(entry)) = rd.next_entry().await {
+        let name = entry.file_name();
+        if !is_owned(&name) {
+            out.push((entry.path(), name));
+        }
+    }
+    out
+}
+
+/// Outcome of [`remove_if_stale`].
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Removal {
+    Removed(u64),
+    TooYoung,
+    Failed,
+}
+
+/// Has nothing under `path` changed for `grace`? Unreadable → `false` (leave it).
+pub async fn is_stale(path: &Path, grace: Duration, now: SystemTime) -> bool {
+    let p = path.to_path_buf();
+    let newest = tokio::task::spawn_blocking(move || newest_change_blocking(&p))
+        .await
+        .ok()
+        .flatten();
+    match newest {
+        Some(t) => now.duration_since(t).map(|age| age >= grace).unwrap_or(false),
+        None => false,
+    }
+}
+
+/// Remove a file or directory tree once nothing under it changed for `grace`.
+/// Returns the bytes it occupied on disk.
+pub async fn remove_if_stale(path: &Path, grace: Duration, now: SystemTime) -> Removal {
+    if !grace.is_zero() && !is_stale(path, grace, now).await {
+        return Removal::TooYoung;
+    }
+    let p = path.to_path_buf();
+    let res = tokio::task::spawn_blocking(move || {
+        let bytes = disk_usage_blocking(&p);
+        let r = match std::fs::symlink_metadata(&p) {
+            Ok(m) if m.is_dir() => std::fs::remove_dir_all(&p),
+            Ok(_) => std::fs::remove_file(&p),
+            Err(e) => Err(e),
+        };
+        r.map(|()| bytes)
+    })
+    .await;
+    match res {
+        Ok(Ok(b)) => Removal::Removed(b),
+        Ok(Err(e)) if e.kind() == std::io::ErrorKind::NotFound => Removal::Removed(0),
+        Ok(Err(e)) => {
+            warn!(path = %path.display(), error = %e, "reclaim: removal failed");
+            Removal::Failed
+        }
+        Err(_) => Removal::Failed,
+    }
+}
+
+/// The most recent modification or status change of `path` or anything under
+/// it. ctime matters: a container renamed from `tmp/` into `cache/` keeps its
+/// old mtime, and only ctime records the move.
+pub fn newest_change_blocking(path: &Path) -> Option<SystemTime> {
+    let mut newest: Option<SystemTime> = None;
+    let mut stack = vec![path.to_path_buf()];
+    while let Some(p) = stack.pop() {
+        let Ok(meta) = std::fs::symlink_metadata(&p) else {
+            if p == path {
+                return None;
+            }
+            continue;
+        };
+        for t in [meta.modified().ok(), change_time(&meta)].into_iter().flatten() {
+            newest = Some(newest.map_or(t, |n| n.max(t)));
+        }
+        if meta.is_dir() {
+            if let Ok(rd) = std::fs::read_dir(&p) {
+                stack.extend(rd.flatten().map(|e| e.path()));
+            }
+        }
+    }
+    newest
+}
+
+#[cfg(unix)]
+fn change_time(meta: &std::fs::Metadata) -> Option<SystemTime> {
+    use std::os::unix::fs::MetadataExt;
+    let secs = u64::try_from(meta.ctime()).ok()?;
+    Some(SystemTime::UNIX_EPOCH + Duration::from_secs(secs))
+}
+
+#[cfg(not(unix))]
+fn change_time(_meta: &std::fs::Metadata) -> Option<SystemTime> {
+    None
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn young_entries_survive_and_zero_grace_removes() {
+        let d = std::env::temp_dir().join(format!("fsutil-{}", std::process::id()));
+        std::fs::create_dir_all(d.join("a")).unwrap();
+        std::fs::write(d.join("a/f"), vec![1u8; 4096]).unwrap();
+        let now = SystemTime::now();
+        assert_eq!(
+            remove_if_stale(&d.join("a"), Duration::from_secs(3600), now).await,
+            Removal::TooYoung
+        );
+        assert!(matches!(
+            remove_if_stale(&d.join("a"), Duration::ZERO, now).await,
+            Removal::Removed(_)
+        ));
+        assert!(!d.join("a").exists());
+        assert_eq!(remove_if_stale(&d.join("gone"), Duration::ZERO, now).await, Removal::Removed(0));
+        let _ = std::fs::remove_dir_all(&d);
+    }
+
+    #[tokio::test]
+    async fn unowned_lists_only_unclaimed_names() {
+        let d = std::env::temp_dir().join(format!("fsutil-own-{}", std::process::id()));
+        std::fs::create_dir_all(d.join("keep")).unwrap();
+        std::fs::create_dir_all(d.join("drop")).unwrap();
+        let got = unowned_entries(&d, |n| n == "keep").await;
+        assert_eq!(got.len(), 1);
+        assert_eq!(got[0].1, "drop");
+        let _ = std::fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn sizes_of_a_missing_dir_are_zero() {
+        assert_eq!(dir_size_blocking(Path::new("/definitely/not/here")), 0);
+        assert_eq!(disk_usage_blocking(Path::new("/definitely/not/here")), 0);
+    }
+}

src/transport/hull.rs

@@ -0,0 +1,226 @@
+//! The plugin → hull direction: callbacks a plugin makes into meta-share.
+//!
+//! The hull serves these on a **separate internal listener**
+//! (`META_SHARE_INTERNAL_LISTEN`, default `0.0.0.0:3001`) that no deployment
+//! publishes: they reach meta-core's records and write to them, so they must
+//! never sit behind the public perimeter. A plugin finds it at
+//! `META_SHARE_HULL_URL` (default `http://metashare-app:3001`).
+//!
+//! Routes:
+//!
+//! - `POST /internal/events`            ← [`Event`] → `204` once the hull acted.
+//! - `GET  /internal/records/:cid[?cached=1]` → [`RecordInfo`] (`404` = no
+//!   record). `cached=1` answers from the hull's record cache only (no
+//!   meta-core round trip) — for display names on a hot path.
+//! - `GET  /internal/nzb/manifest/:cid` → the `.nzb` bytes the cid resolves to,
+//!   with [`HDR_MANIFEST_SOURCE`] (`record` | `redeem`) and, for `record`,
+//!   [`HDR_MANIFEST_CID`]. `404` = not a Usenet cid; any other error carries
+//!   `{"error"}` — the manifest could not be redeemed or fetched. Parsing and
+//!   the `nzb-posting` digest check stay in the plugin.
+//!
+//! Why callbacks and not plugin-side logic: these need what only the hull has —
+//! the record cache and meta-core, the gateway redeem, the IPFS tier. Plugins
+//! never call each other; the hull routes every such edge.
+
+use std::time::Duration;
+
+use serde::de::DeserializeOwned;
+use serde::{Deserialize, Serialize};
+
+use super::dto::Event;
+
+/// Default hull callback base, matching the dev and store compose service name.
+pub const DEFAULT_HULL_URL: &str = "http://metashare-app:3001";
+
+/// Where the hull got a manifest: the record's `manifest` pointer, or a
+/// gateway redeem.
+pub const HDR_MANIFEST_SOURCE: &str = "x-metamesh-manifest-source";
+/// The content cid of a `record`-sourced manifest.
+pub const HDR_MANIFEST_CID: &str = "x-metamesh-manifest-cid";
+
+/// `.nzb` bytes as the hull resolved them.
+#[derive(Clone, Debug)]
+pub struct ManifestBytes {
+    pub bytes: bytes::Bytes,
+    /// `true` when redeemed through a gateway, `false` for the record pointer.
+    pub redeemed: bool,
+    pub manifest_cid: Option<String>,
+}
+
+/// What a plugin may need to know about a record.
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct RecordInfo {
+    /// meta-core's hash id for the record (write-back key).
+    #[serde(default)]
+    pub hash_id: Option<String>,
+    /// The record's `fileName`.
+    #[serde(default)]
+    pub file_name: Option<String>,
+    /// The record's display title.
+    #[serde(default)]
+    pub title: Option<String>,
+}
+
+/// Plugin-side client for the hull's internal API. A client with no base URL
+/// (`META_SHARE_HULL_URL=` empty, or [`HullClient::disabled`]) answers every call
+/// with "not available", which is what unit tests and a standalone plugin get.
+#[derive(Clone, Debug)]
+pub struct HullClient {
+    base: Option<String>,
+    http: reqwest::Client,
+}
+
+impl HullClient {
+    pub fn new(base: Option<String>, http: reqwest::Client) -> Self {
+        let base = base
+            .map(|b| b.trim().trim_end_matches('/').to_string())
+            .filter(|b| !b.is_empty());
+        Self { base, http }
+    }
+
+    /// From `META_SHARE_HULL_URL`, defaulting to [`DEFAULT_HULL_URL`]; an
+    /// explicitly empty value disables callbacks.
+    pub fn from_env(http: reqwest::Client) -> Self {
+        let base = match std::env::var("META_SHARE_HULL_URL") {
+            Ok(v) => Some(v),
+            Err(_) => Some(DEFAULT_HULL_URL.to_string()),
+        };
+        Self::new(base, http)
+    }
+
+    pub fn disabled() -> Self {
+        Self::new(None, reqwest::Client::new())
+    }
+
+    pub fn is_enabled(&self) -> bool {
+        self.base.is_some()
+    }
+
+    fn url(&self, path: &str) -> Result<String, String> {
+        self.base
+            .as_ref()
+            .map(|b| format!("{b}{path}"))
+            .ok_or_else(|| "hull callbacks disabled (META_SHARE_HULL_URL unset)".to_string())
+    }
+
+    /// Notify the hull; returns once it has acted.
+    pub async fn emit(&self, ev: &Event) -> Result<(), String> {
+        let r = self
+            .http
+            .post(self.url("/internal/events")?)
+            .json(ev)
+            .timeout(Duration::from_secs(30))
+            .send()
+            .await
+            .map_err(|e| format!("hull unreachable: {e}"))?;
+        if r.status().is_success() {
+            Ok(())
+        } else {
+            Err(format!("hull refused event: {}", r.status()))
+        }
+    }
+
+    /// `GET` a JSON resource; `404` → `Ok(None)`.
+    pub async fn get_json<T: DeserializeOwned>(&self, path: &str, timeout: Duration) -> Result<Option<T>, String> {
+        let r = self
+            .http
+            .get(self.url(path)?)
+            .timeout(timeout)
+            .send()
+            .await
+            .map_err(|e| format!("hull unreachable: {e}"))?;
+        if r.status().as_u16() == 404 {
+            return Ok(None);
+        }
+        if !r.status().is_success() {
+            return Err(error_message(r).await);
+        }
+        r.json::<T>().await.map(Some).map_err(|e| format!("hull sent malformed JSON: {e}"))
+    }
+
+    /// `GET` raw bytes; `404` → `Ok(None)`, any other failure → the hull's
+    /// `{"error"}` message.
+    pub async fn get_bytes(&self, path: &str, timeout: Duration) -> Result<Option<bytes::Bytes>, String> {
+        let r = self
+            .http
+            .get(self.url(path)?)
+            .timeout(timeout)
+            .send()
+            .await
+            .map_err(|e| format!("hull unreachable: {e}"))?;
+        if r.status().as_u16() == 404 {
+            return Ok(None);
+        }
+        if !r.status().is_success() {
+            return Err(error_message(r).await);
+        }
+        r.bytes().await.map(Some).map_err(|e| format!("hull body: {e}"))
+    }
+
+    /// The `.nzb` behind a Usenet cid. `Ok(None)` = not a Usenet cid.
+    pub async fn nzb_manifest(&self, cid: &str, timeout: Duration) -> Result<Option<ManifestBytes>, String> {
+        let r = self
+            .http
+            .get(self.url(&format!("/internal/nzb/manifest/{cid}"))?)
+            .timeout(timeout)
+            .send()
+            .await
+            .map_err(|e| format!("hull unreachable: {e}"))?;
+        if r.status().as_u16() == 404 {
+            return Ok(None);
+        }
+        if !r.status().is_success() {
+            return Err(error_message(r).await);
+        }
+        let h = r.headers();
+        let redeemed = h
+            .get(HDR_MANIFEST_SOURCE)
+            .and_then(|v| v.to_str().ok())
+            .is_some_and(|v| v == "redeem");
+        let manifest_cid = h
+            .get(HDR_MANIFEST_CID)
+            .and_then(|v| v.to_str().ok())
+            .map(str::to_string);
+        let bytes = r.bytes().await.map_err(|e| format!("hull body: {e}"))?;
+        Ok(Some(ManifestBytes { bytes, redeemed, manifest_cid }))
+    }
+
+    /// [`RecordInfo`] for `cid` (record cache, else meta-core), or `None` when
+    /// the hull has no record.
+    pub async fn record(&self, cid: &str) -> Result<Option<RecordInfo>, String> {
+        self.get_json(&format!("/internal/records/{cid}"), Duration::from_secs(15)).await
+    }
+
+    /// [`RecordInfo`] from the hull's record cache only.
+    pub async fn cached_record(&self, cid: &str) -> Result<Option<RecordInfo>, String> {
+        self.get_json(&format!("/internal/records/{cid}?cached=1"), Duration::from_secs(5)).await
+    }
+}
+
+async fn error_message(r: reqwest::Response) -> String {
+    let status = r.status();
+    r.json::<serde_json::Value>()
+        .await
+        .ok()
+        .and_then(|v| v.get("error").and_then(|e| e.as_str()).map(str::to_string))
+        .unwrap_or_else(|| format!("hull error {status}"))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn a_disabled_client_refuses_without_network() {
+        let h = HullClient::disabled();
+        assert!(!h.is_enabled());
+        assert!(h.emit(&Event::Archive { cid: "x".into() }).await.is_err());
+        assert!(h.record("x").await.is_err());
+    }
+
+    #[test]
+    fn blank_base_disables() {
+        assert!(!HullClient::new(Some("  ".into()), reqwest::Client::new()).is_enabled());
+        assert!(HullClient::new(Some("http://h:3001/".into()), reqwest::Client::new()).is_enabled());
+    }
+}

src/transport/ipfs.rs

@@ -0,0 +1,157 @@
+//! IPFS-specific half of the contract: the ipfs plugin's extra routes.
+//!
+//! The ipfs plugin owns the libp2p swarm (identify / kad / mdns / bitswap), the
+//! blockstore (`ipfs/blocks.redb`: material blocks + leaf refs into files on the
+//! shared volume), bitswap ingress and the peer directory. These routes are the
+//! hull's facade onto that — each mirrors one blockstore / swarm call the hull
+//! made in-process, so the hull's logic (seed rows, eviction, meta-core links,
+//! probe grading) did not have to change shape.
+//!
+//! Public passthrough (the hull relays meta-share's own route verbatim):
+//!
+//! - `GET /ipfs/:cid` — the read-only IPFS gateway.
+//!
+//! Internal (all under `/ipfs-tier/`):
+//!
+//! | route | in-process equivalent |
+//! |---|---|
+//! | `GET  peers` → [`PeersInfo`] | `swarm::Command::Peers` |
+//! | `GET  directory` → [`Directory`] | `swarm::PeerDirectory` (gateway discovery) |
+//! | `GET  local/:cid` (Range) → bytes / `404` | `files::bitswap::try_local_block` |
+//! | `GET  has/:cid` → [`Present`] | `block_store.get(cid).is_some()` |
+//! | `GET  complete/:cid` → [`Complete`] | `local_dag_complete` |
+//! | `GET  cat/:cid?max=` → bytes | `ipfs_walker::get_block` (+ dag-pb assemble) |
+//! | `GET  fetch/:cid?timeout_ms=` → block bytes | `block_store.get` → `bitswap_get_block` |
+//! | `PUT  block/:cid` ← bytes | `blockstore::put_ipfs_block` |
+//! | `POST import` ← file bytes → [`Imported`] | `put_ipfs_blocks(compute_ipfs_blocks(..))` |
+//! | `POST forget/:cid` | `api::seeds::remove_ipfs_blocks` |
+//! | `POST forget-library` ← [`ForgetLibrary`] | `block_store.remove_library_dag` |
+//! | `POST provide/:cid`, `POST unprovide/:cid` | `Command::Provide` / `StopProviding` |
+//! | `POST drop-refs` ← [`DropRefs`] | `block_store.drop_refs_for_materials` |
+//! | `GET  first-leaf/:cid` → [`FirstLeaf`] | `block_store.first_leaf_backing` |
+//! | `GET  rel-path/:midhash` → [`RelPath`] | `resolver().rel_path_for` |
+//! | `POST share/library` ← [`ShareLibrary`] → [`Shared`] | `ipfs_seed::seed_file_as_refs` |
+//! | `POST share/container` ← [`ShareContainer`] → [`Shared`] | the seal / rebuild re-chunk |
+//! | `GET  stats/blockstore` | `/api/stats/blockstore` body |
+//! | `GET  debug/:cid` | `/api/debug/blockstore/:cid` body |
+//!
+//! Plugin → hull, over [`super::hull`]: [`Event::Seed`](super::Event::Seed)
+//! (a bitswap fetch now seeds), [`Event::Promoted`](super::Event::Promoted) (an
+//! ingress job was renamed into `cache/`), and `GET /internal/records/:cid` for
+//! a display name.
+
+use serde::{Deserialize, Serialize};
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct PeersInfo {
+    pub local_peer_id: String,
+    pub connected: Vec<String>,
+}
+
+/// One peer of the directory, with its freshness as *ages* (the hull rebuilds
+/// its own `Instant`s from them).
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq)]
+pub struct DirectoryPeer {
+    pub peer_id: String,
+    #[serde(default)]
+    pub base_url: Option<String>,
+    /// Milliseconds since the gateway last advertised `nzbFetch`; `None` = not.
+    #[serde(default)]
+    pub nzb_fetch_age_ms: Option<u64>,
+    /// Milliseconds since the caps were fetched; `None` = never.
+    #[serde(default)]
+    pub caps_age_ms: Option<u64>,
+    /// The gateway's redeem claims, as the gateway serves them
+    /// (`{codec, field, hosts, sources}`).
+    #[serde(default)]
+    pub redeems: Vec<serde_json::Value>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq)]
+pub struct Directory {
+    pub peers: Vec<DirectoryPeer>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct Present {
+    pub present: bool,
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct Complete {
+    pub complete: bool,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct Imported {
+    pub root: String,
+    pub size: u64,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct ForgetLibrary {
+    pub root: String,
+    pub midhash: String,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct DropRefs {
+    /// `(container, rel)` pairs.
+    pub materials: Vec<(String, String)>,
+}
+
+/// Where a dag's first leaf points (the filestore's `LeafBacking`).
+#[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq)]
+#[serde(rename_all = "snake_case")]
+pub enum LeafBacking {
+    /// A ref into a file meta-core owns, by midhash256.
+    MetaCore(String),
+    /// A ref into a file in a cache container.
+    Cache,
+    /// The leaf's bytes are stored material (swarm-fetched, no file behind it).
+    Material,
+    /// The root or a node on the way down is absent.
+    Unknown,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq)]
+pub struct FirstLeaf {
+    pub backing: LeafBacking,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct RelPath {
+    pub rel: Option<String>,
+}
+
+/// Seed a meta-core library file as refs `(midhash, offset, len)`.
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct ShareLibrary {
+    /// Path relative to the files volume (what meta-core's WebDAV serves).
+    pub rel: String,
+    pub midhash: String,
+}
+
+/// Seed a file in a cache container as refs `(container, rel, offset, len)`.
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct ShareContainer {
+    pub container: String,
+    pub rel: String,
+    /// Absolute path on the shared volume.
+    pub path: String,
+    /// Verify the resulting dag is fully held before answering (the seal does;
+    /// the rebuild doesn't).
+    #[serde(default)]
+    pub verify_complete: bool,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct Shared {
+    pub root: String,
+    pub size: u64,
+    /// Leaf refs written.
+    pub refs: u64,
+    /// `verify_complete` was asked and the dag is fully held locally.
+    #[serde(default)]
+    pub complete: bool,
+}

src/transport/magic.rs

@@ -0,0 +1,230 @@
+//! Content sniffing — what the bytes *are*, independent of what they are called.
+//!
+//! Every classifier in the usenet plugin's `archive` keys off the filename quoted in an NZB
+//! subject. A fully obfuscated posting has no filename to key off, so all of them
+//! fail **open**: `subject_is_par2` misses the parity volumes, `subject_is_rar`
+//! misses the archive, and `pick_media`'s `best_video.or(best_any)` then elects
+//! the largest arbitrary blob as "the movie". That blob is what gets served, and
+//! — far worse — what the Usenet → IPFS re-seed hashes and publishes as
+//! the record's content identity.
+//!
+//! Names are advisory. Bytes are not. This module is the byte half.
+//!
+//! The resolution order mirrors `filename-tool`'s `FileTypeConfigurable`
+//! (the source of truth named by `METADATA_KEYS.md` §`fileType`): sniff the
+//! content first, compare with the extension when there is one, and on
+//! **disagreement return `undefined` rather than guessing** — a wrong `fileType`
+//! routes a file to a handler that cannot open it.
+
+/// A container format recognised from its leading bytes.
+///
+/// Deliberately *not* exhaustive over "every format that exists" — it only needs
+/// to separate payload from wrapper from parity well enough to decide whether a
+/// posting has resolved. [`Magic::Unknown`] is a first-class answer, and callers
+/// must treat it as "cannot prove anything", never as "bad".
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum Magic {
+    /// RAR 1.5–4.x — signature `52 61 72 21 1A 07 00`.
+    Rar4,
+    /// RAR 5.0+ — signature `52 61 72 21 1A 07 01 00`.
+    Rar5,
+    /// par2 recovery data. **Never payload**, whatever its size.
+    Par2,
+    SevenZip,
+    /// A zip that is not an EPUB (includes `.cbz`, which `METADATA_KEYS.md`
+    /// §`fileType` classifies as `archive`).
+    Zip,
+    /// A zip whose first entry is the stored `mimetype` = `application/epub+zip`.
+    /// Split from [`Magic::Zip`] because an EPUB **is** the document, not a
+    /// wrapper around one — three correct ebook postings on watch.nsl.sh are
+    /// this shape and a naive "zip ⇒ archive" rule condemns all of them.
+    Epub,
+    /// Palm database holding an ebook (`BOOK`/`TEXt` type at offset 60) — the
+    /// `CR!…`-named MobiPocket/eReader files an indexer hands back for a book.
+    PalmDoc,
+    Matroska,
+    Mp4,
+    /// RIFF container (AVI, and also WAV — the extension disambiguates).
+    Riff,
+    Pdf,
+    /// Nothing recognised. Not an error: see the type-level note above.
+    Unknown,
+}
+
+impl Magic {
+    /// The `METADATA_KEYS.md` §`fileType` value these bytes imply, or `None`
+    /// when the content says nothing.
+    ///
+    /// [`Magic::Riff`] is deliberately `None`: RIFF is AVI *and* WAV, i.e. a
+    /// video and an audio type, and guessing between them is exactly the
+    /// "disagreement returns undefined" case the registry legislates.
+    pub fn file_type(self) -> Option<&'static str> {
+        match self {
+            Magic::Matroska | Magic::Mp4 => Some("video"),
+            Magic::Epub | Magic::PalmDoc | Magic::Pdf => Some("document"),
+            Magic::Rar4 | Magic::Rar5 | Magic::SevenZip | Magic::Zip => Some("archive"),
+            // Parity is not a fileType — it is not payload at all. Callers route
+            // on `is_parity()` before they ever ask for a type.
+            Magic::Par2 => None,
+            Magic::Riff | Magic::Unknown => None,
+        }
+    }
+
+    /// The extension these bytes imply, for naming a materialised file whose
+    /// posted name carried none (or carried an obfuscated one) — see
+    /// the usenet plugin's `naming`.
+    ///
+    /// Abstains wherever the bytes do not pick a single extension, on the same
+    /// discipline as [`Magic::file_type`]: [`Magic::Riff`] is AVI *and* WAV, and
+    /// [`Magic::PalmDoc`] is `.prc`/`.mobi`/`.pdb`. [`Magic::Par2`] abstains for
+    /// a stronger reason — parity is never the payload, so it never gets named.
+    pub fn extension(self) -> Option<&'static str> {
+        match self {
+            Magic::Matroska => Some("mkv"),
+            Magic::Mp4 => Some("mp4"),
+            Magic::Pdf => Some("pdf"),
+            Magic::Epub => Some("epub"),
+            Magic::Rar4 | Magic::Rar5 => Some("rar"),
+            Magic::SevenZip => Some("7z"),
+            Magic::Zip => Some("zip"),
+            Magic::Riff | Magic::PalmDoc | Magic::Par2 | Magic::Unknown => None,
+        }
+    }
+
+    /// par2 recovery data: present to repair the payload, never to *be* it.
+    ///
+    /// One posting on watch.nsl.sh (`…gk2iwj5i…`, claiming `video`/`movie`)
+    /// materialised to a single 122 MB par2 file and published a dag-pb root
+    /// over it, because `subject_is_par2` reads the `.par2` extension the
+    /// obfuscated subject did not carry.
+    pub fn is_parity(self) -> bool {
+        matches!(self, Magic::Par2)
+    }
+
+    /// A multi-volume wrapper that must be unwrapped before anything inside it
+    /// can be addressed. `Epub`/`PalmDoc` are containers in the file-format
+    /// sense but **not** in this one — they are the document.
+    pub fn is_archive(self) -> bool {
+        matches!(self, Magic::Rar4 | Magic::Rar5 | Magic::SevenZip | Magic::Zip)
+    }
+}
+
+/// Bytes needed for a confident answer. The EPUB probe is the deepest reader:
+/// local header (30) + `"mimetype"` (8) + extra field + `"application/epub+zip"`
+/// (20). 128 covers any sane extra-field length.
+pub const SNIFF_LEN: usize = 128;
+
+/// Classify a file from its leading bytes. Short input is fine — every probe
+/// bounds-checks and falls through to [`Magic::Unknown`].
+pub fn sniff(head: &[u8]) -> Magic {
+    // RAR5 before RAR4: the RAR4 signature is a prefix of neither, but the
+    // shared first 6 bytes make ordering worth being explicit about.
+    if head.starts_with(b"Rar!\x1a\x07\x01\x00") {
+        return Magic::Rar5;
+    }
+    if head.starts_with(b"Rar!\x1a\x07\x00") {
+        return Magic::Rar4;
+    }
+    if head.starts_with(b"PAR2\x00PKT") {
+        return Magic::Par2;
+    }
+    if head.starts_with(b"7z\xbc\xaf\x27\x1c") {
+        return Magic::SevenZip;
+    }
+    if head.starts_with(b"%PDF-") {
+        return Magic::Pdf;
+    }
+    // EBML — Matroska and WebM share it; both demux the same way here.
+    if head.starts_with(b"\x1a\x45\xdf\xa3") {
+        return Magic::Matroska;
+    }
+    if head.len() >= 8 && &head[4..8] == b"ftyp" {
+        return Magic::Mp4;
+    }
+    if head.starts_with(b"RIFF") {
+        return Magic::Riff;
+    }
+    if head.starts_with(b"PK\x03\x04") {
+        return if is_epub(head) { Magic::Epub } else { Magic::Zip };
+    }
+    if is_palm_ebook(head) {
+        return Magic::PalmDoc;
+    }
+    Magic::Unknown
+}
+
+/// An EPUB is a zip whose **first** entry is an uncompressed `mimetype` member
+/// holding exactly `application/epub+zip` (OCF 3, §4.2 — the rule exists so the
+/// type is readable without inflating anything). That is what separates a book
+/// from a `.cbz` or a scene `.zip`.
+fn is_epub(head: &[u8]) -> bool {
+    const NAME_AT: usize = 30;
+    const NAME: &[u8] = b"mimetype";
+    const VALUE: &[u8] = b"application/epub+zip";
+
+    // Stored (method 0) is mandatory for this member; anything else is not an
+    // OCF container even if it happens to hold the string.
+    if head.len() < NAME_AT + NAME.len() || u16::from_le_bytes([head[8], head[9]]) != 0 {
+        return false;
+    }
+    if &head[NAME_AT..NAME_AT + NAME.len()] != NAME {
+        return false;
+    }
+    let extra = u16::from_le_bytes([head[28], head[29]]) as usize;
+    let at = NAME_AT + NAME.len() + extra;
+    head.len() >= at + VALUE.len() && &head[at..at + VALUE.len()] == VALUE
+}
+
+/// A Palm database (`name[32]`, then `type` at 60 and `creator` at 64) whose
+/// type marks it an ebook. Covers MobiPocket (`BOOK`/`MOBI`), eReader
+/// (`BOOK`/`PNRd`) and PalmDOC (`TEXt`/`REAd`).
+fn is_palm_ebook(head: &[u8]) -> bool {
+    const TYPE_AT: usize = 60;
+    head.len() >= TYPE_AT + 4 && matches!(&head[TYPE_AT..TYPE_AT + 4], b"BOOK" | b"TEXt")
+}
+
+/// The `fileType` an extension implies, for the agreement check. Mirrors the
+/// table in `METADATA_KEYS.md` §`fileType`, whose own source of truth is
+/// `filename-tool`'s `ExtentionsMapping.ts`. Only the values reachable from a
+/// Usenet posting are listed; anything else answers `None` ("no opinion"),
+/// which agrees with everything.
+pub fn file_type_from_extension(name: &str) -> Option<&'static str> {
+    let ext = name.rsplit_once('.')?.1.to_ascii_lowercase();
+    Some(match ext.as_str() {
+        "mkv" | "mp4" | "webm" | "avi" | "mov" | "wmv" | "flv" | "m4v" | "mpg" | "mpeg" | "3gp"
+        | "ts" | "m2ts" | "mts" | "vob" | "ogm" | "ogv" | "divx" | "xvid" => "video",
+        "mp3" | "wav" | "flac" | "aac" | "ogg" | "oga" | "opus" | "m4a" | "m4b" | "wma" | "mka" => {
+            "audio"
+        }
+        "pdf" | "epub" | "mobi" | "azw" | "azw3" | "djvu" | "txt" | "doc" | "docx" | "prc" => {
+            "document"
+        }
+        "zip" | "rar" | "7z" | "tar" | "gz" | "bz2" | "xz" | "zst" | "cbz" | "cbr" => "archive",
+        "srt" | "sub" | "ass" | "ssa" | "vtt" | "sbv" | "smi" => "subtitle",
+        "par2" => return None, // parity, not a fileType — see `Magic::is_parity`
+        _ => return None,
+    })
+}
+
+/// Resolve one file's `fileType` from its bytes and its name together.
+///
+/// The registry's rule, verbatim: *"sniff the MIME from the first bytes; if it
+/// maps to a definite value, compare it with the extension's value — agreement
+/// wins, **disagreement returns `undefined`**"*. `None` here is that
+/// `undefined`: it means "these two sources disagree, do not route on it",
+/// which is a different and strictly weaker statement than "wrong".
+pub fn resolve_file_type(head: &[u8], name: &str) -> Option<&'static str> {
+    let sniffed = sniff(head).file_type();
+    let from_name = file_type_from_extension(name);
+    match (sniffed, from_name) {
+        (Some(a), Some(b)) if a == b => Some(a),
+        (Some(_), Some(_)) => None, // disagreement ⇒ undefined
+        (Some(a), None) => Some(a), // no usable extension: content decides
+        (None, Some(b)) => Some(b), // unrecognised bytes: the name is all we have
+        (None, None) => None,
+    }
+}
+
+#[cfg(test)]
+#[path = "magic_tests.rs"]
+mod tests;

src/transport/magic_tests.rs

@@ -0,0 +1,205 @@
+//! Tests for [`super`] — content sniffing.
+//!
+//! Every fixture below is **real**: the leading bytes of files materialised on
+//! watch.nsl.sh, captured 2026-09-16. Synthetic headers would have agreed with
+//! whatever this module happened to implement; these are the bytes that actually
+//! defeated the name-based classifiers.
+
+use super::*;
+
+/// Decode a hex fixture. Panics on bad input — these are consts, not user data.
+fn hx(s: &str) -> Vec<u8> {
+    assert!(s.len() % 2 == 0, "odd-length hex fixture");
+    (0..s.len() / 2)
+        .map(|i| u8::from_str_radix(&s[i * 2..i * 2 + 2], 16).expect("hex"))
+        .collect()
+}
+
+/// Volume 1 of 35 of the trigger posting — `[smol] Tensei Oujo … S01E05v2`,
+/// `bagcsaaa5brqxa2jonz5geltmnftglnrqztt2n23topk2bsnglvhjmty`. Extension-less
+/// name, RAR5, method 0 (store). The EBML header of the contained MKV is
+/// visible inside these same 200 bytes, at the volume's data offset.
+const RAR5_VOL1: &str = "526172211a0701002409274c1001050c010b0101bac0d7af80808080004ce244be8b0102130b94bfd7af80800004eefaa1818d8000b683021a946591800001635b736d6f6c5d2054656e736569204f756a6f20746f2054656e736169205265696a6f75206e6f204d61686f75204b616b756d6569202d205330314530357632202842442031303830702048455643204f70757329205b39394434344436465d2e6d6b760a0313184c91644af1d2231a45dfa3a34286810142f7810142f2810442f381084282886d61";
+
+/// Volume 2 — a *continuation* volume. Byte-identical signature to volume 1,
+/// which is why "does it start with `Rar!`" cannot tell you which volume you
+/// hold, and why electing "the largest file" picks one at random.
+const RAR5_VOL2: &str = "526172211a0701000d6002471101050c03010b0101bac0d7af808080800030554b778b01021b0b93bfd7af80800004eefaa1818d8000b6830236ead743800001635b736d6f6c5d2054656e736569204f756a6f20746f2054656e736169205265696a6f75206e6f204d61686f75204b616b756d6569202d205330314530357632202842442031303830702048455643204f70757329205b39394434344436465d2e6d6b760a0313184c91644af1d223da5b7aaacbc910a46f8b3a8886bc4bfaff0518a54e4a2eb791";
+
+/// RAR4 head, from `…gkigqufwwvhpkbhijvufrdcshogy` (claims `video`/`movie`).
+const RAR4: &str = "526172211a0700197a7311000d00000000000000d9427423916f0069ff3f067a";
+
+/// par2 head, from `…gk2iwj5iy3cyzafghckcab7j2fuy` — the posting that
+/// materialised to a *single* par2 file and published a dag-pb root over it.
+const PAR2: &str = "5041523200504b54083b1a00000000000226e38aa627ac7ee833f4fc92637a69";
+
+/// A genuine EPUB — `Brandon Sanderson - Mistborn`, `…gkdthedplbm5qwmtedo2ymyct2si`.
+/// Stored `mimetype` member first, exactly as OCF requires.
+const EPUB: &str = "504b0304140000000000ee523a486f61ab2c1400000014000000080000006d696d65747970656170706c69636174696f6e2f657075622b7a6970504b0304140002080800ee523a4827f0a79ab4000000";
+
+/// A MobiPocket Palm database — the `CR!…`-named ebook from `…gle4fanocf5txcdocnto6dh4au4a`.
+/// Type `BOOK` at offset 60, creator `MOBI` at 64.
+const PALM: &str = "43522158593841474857463731314e4642394a44314443314d374b4548344d000000000057d30d1e57d3113200000000000000000000000000000000424f4f4b4d4f4249000001df";
+
+/// A Matroska head from a posting that resolved correctly.
+const MKV: &str = "1a45dfa3a34286810142f7810142f281";
+
+// ---- the six shapes that mattered -------------------------------------------
+
+#[test]
+fn the_real_headers_are_each_recognised() {
+    assert_eq!(sniff(&hx(RAR5_VOL1)), Magic::Rar5);
+    assert_eq!(sniff(&hx(RAR5_VOL2)), Magic::Rar5);
+    assert_eq!(sniff(&hx(RAR4)), Magic::Rar4);
+    assert_eq!(sniff(&hx(PAR2)), Magic::Par2);
+    assert_eq!(sniff(&hx(EPUB)), Magic::Epub);
+    assert_eq!(sniff(&hx(PALM)), Magic::PalmDoc);
+    assert_eq!(sniff(&hx(MKV)), Magic::Matroska);
+}
+
+/// The regression that started this: three correct ebook postings on the box are
+/// zips, and a "zip ⇒ archive" rule condemns all of them. An EPUB is the
+/// document, not a wrapper around one.
+#[test]
+fn an_epub_is_a_document_not_an_archive() {
+    let m = sniff(&hx(EPUB));
+    assert_eq!(m.file_type(), Some("document"));
+    assert!(!m.is_archive(), "an epub must not route into the unwrapper");
+}
+
+/// A `.cbz` and a scene `.zip` share the EPUB signature for four bytes and
+/// diverge at the first member. `METADATA_KEYS.md` §`fileType` puts `cbz` in
+/// `archive`, so plain-zip ⇒ archive is the correct fallback.
+#[test]
+fn a_zip_without_the_ocf_mimetype_member_is_an_archive() {
+    let mut z = hx(EPUB);
+    // Corrupt just the mimetype *value*; everything else stays a valid zip.
+    let at = 38;
+    z[at..at + 3].copy_from_slice(b"XXX");
+    assert_eq!(sniff(&z), Magic::Zip);
+    assert_eq!(sniff(&z).file_type(), Some("archive"));
+}
+
+/// OCF requires the `mimetype` member to be *stored*. A deflated one is not an
+/// OCF container even if the bytes happen to spell the string.
+#[test]
+fn a_deflated_mimetype_member_is_not_an_epub() {
+    let mut z = hx(EPUB);
+    z[8] = 8; // method = deflate
+    assert_eq!(sniff(&z), Magic::Zip);
+}
+
+#[test]
+fn parity_is_not_a_file_type_at_all() {
+    let m = sniff(&hx(PAR2));
+    assert!(m.is_parity());
+    assert_eq!(m.file_type(), None, "par2 must never present as payload");
+}
+
+#[test]
+fn every_rar_generation_reads_as_an_archive() {
+    for f in [RAR4, RAR5_VOL1, RAR5_VOL2] {
+        let m = sniff(&hx(f));
+        assert!(m.is_archive());
+        assert_eq!(m.file_type(), Some("archive"));
+        assert!(!m.is_parity());
+    }
+}
+
+// ---- Unknown is an answer, not an error -------------------------------------
+
+#[test]
+fn unrecognised_and_truncated_input_is_unknown_not_a_panic() {
+    assert_eq!(sniff(b""), Magic::Unknown);
+    assert_eq!(sniff(b"Rar"), Magic::Unknown);
+    assert_eq!(sniff(b"PK\x03"), Magic::Unknown);
+    assert_eq!(sniff(&[0u8; 8]), Magic::Unknown);
+    assert_eq!(sniff(b"not a container at all"), Magic::Unknown);
+    assert_eq!(sniff(&hx(MKV)[..2]), Magic::Unknown);
+}
+
+/// The holed posting on the box (`…gkisayoy…`, 837 MB of zeros) sniffs Unknown.
+/// It is already refused upstream by `disk::first_hole`; this only pins that
+/// Unknown carries no `fileType` claim either way.
+#[test]
+fn a_hole_claims_nothing() {
+    let zeros = vec![0u8; SNIFF_LEN];
+    assert_eq!(sniff(&zeros), Magic::Unknown);
+    assert_eq!(sniff(&zeros).file_type(), None);
+}
+
+/// RIFF is AVI *and* WAV — a video type and an audio type. Guessing between
+/// them is precisely the disagreement case the registry legislates away.
+#[test]
+fn riff_is_recognised_but_declines_to_name_a_type() {
+    let m = sniff(b"RIFF\x00\x00\x00\x00AVI LIST");
+    assert_eq!(m, Magic::Riff);
+    assert_eq!(m.file_type(), None);
+}
+
+// ---- name + bytes together ---------------------------------------------------
+
+#[test]
+fn content_decides_when_the_name_says_nothing() {
+    // The obfuscated case: no extension anywhere.
+    assert_eq!(
+        resolve_file_type(&hx(RAR5_VOL1), "IuYf3jBKJOukfUesqNCw0LoUc1Gqxxmilo0ICJJMeddjB6"),
+        Some("archive")
+    );
+    assert_eq!(resolve_file_type(&hx(MKV), "56GqgfEUMp7RAqF94AGp"), Some("video"));
+}
+
+#[test]
+fn agreement_between_name_and_bytes_wins() {
+    assert_eq!(resolve_file_type(&hx(MKV), "The.Movie.2019.1080p.mkv"), Some("video"));
+    assert_eq!(resolve_file_type(&hx(EPUB), "book.epub"), Some("document"));
+}
+
+/// `METADATA_KEYS.md` §`fileType`: *"disagreement returns `undefined`"*. A file
+/// named `.mkv` whose bytes are a RAR is not a video and is not reliably an
+/// archive either — it is a thing we refuse to route on.
+#[test]
+fn disagreement_returns_undefined_rather_than_guessing() {
+    assert_eq!(resolve_file_type(&hx(RAR5_VOL1), "episode.mkv"), None);
+    assert_eq!(resolve_file_type(&hx(MKV), "release.rar"), None);
+}
+
+#[test]
+fn an_unreadable_header_falls_back_to_the_extension() {
+    assert_eq!(resolve_file_type(b"\x00\x01\x02\x03", "movie.mkv"), Some("video"));
+    assert_eq!(resolve_file_type(b"", "chapter.cbz"), Some("archive"));
+}
+
+#[test]
+fn neither_source_has_an_opinion() {
+    assert_eq!(resolve_file_type(b"\x00\x01\x02\x03", "blob"), None);
+    // par2 is deliberately absent from the extension table: it is parity, and
+    // `Magic::is_parity` is how callers route it.
+    assert_eq!(file_type_from_extension("recovery.par2"), None);
+}
+
+#[test]
+fn the_extension_table_matches_the_registry_for_the_shapes_we_see() {
+    assert_eq!(file_type_from_extension("a.mkv"), Some("video"));
+    assert_eq!(file_type_from_extension("a.MKV"), Some("video"));
+    assert_eq!(file_type_from_extension("a.epub"), Some("document"));
+    assert_eq!(file_type_from_extension("a.cbz"), Some("archive"));
+    assert_eq!(file_type_from_extension("a.rar"), Some("archive"));
+    assert_eq!(file_type_from_extension("a.srt"), Some("subtitle"));
+    assert_eq!(file_type_from_extension("no-extension-here"), None);
+}
+
+/// The MKV inside the trigger posting is *stored*, so its EBML header sits in
+/// plaintext a few hundred bytes into volume 1. That is what makes the
+/// offset-map unwrap possible at all, and it is worth pinning independently of
+/// the RAR parser.
+#[test]
+fn the_stored_mkv_header_is_visible_inside_volume_one() {
+    let v1 = hx(RAR5_VOL1);
+    let ebml = v1
+        .windows(4)
+        .position(|w| w == [0x1a, 0x45, 0xdf, 0xa3])
+        .expect("EBML header should be plainly visible inside a stored RAR volume");
+    assert_eq!(ebml, 174, "volume 1's data offset, per the RAR5 header chain");
+    assert_eq!(sniff(&v1[ebml..]), Magic::Matroska);
+}

src/transport/mod.rs

@@ -0,0 +1,57 @@
+//! **Transport plugins** — the contract between meta-share's protocol-agnostic
+//! hull and the processes that actually move bytes (`meta-transport-torrent`,
+//! `meta-transport-nzb`, `meta-transport-ipfs`, or a competing implementation
+//! of any of them).
+//!
+//! Same split as gateway ⇄ feeders: the plugin finds and moves bytes for one
+//! protocol; the hull keeps everything that is policy and bookkeeping — the
+//! public HTTP API, the material index, seed rows, eviction, focus, meta-core.
+//! Plugins never call each other and never talk to meta-core's record API;
+//! every cross-tier edge goes through the hull.
+//!
+//! - [`plugin::TransportPlugin`] + [`serve::serve_transport`]: the plugin side.
+//! - [`client::RemoteTransport`]: the hull side.
+//! - [`hull::HullClient`]: plugin → hull callbacks (internal port only).
+//! - [`focus::FocusView`]: the plugin's replica of the hull's playback focus.
+//! - [`dto`]: common wire types; [`error::ApiError`] / [`range`]: shared so a
+//!   relayed plugin response is byte-identical to what the monolith answered.
+//! - [`cid`]: CID parsing + cid-shape decoders, shared so every process
+//!   classifies a cid the same way.
+//! - [`torrent`]: the torrent plugin's extra routes + magnet helpers.
+//! - [`ipfs`]: the ipfs plugin's extra routes (the hull's blockstore/swarm facade).
+//! - [`nzb`]: Usenet cid shapes, the usenet plugin's extra routes + settings.
+//! - [`magic`]: content sniffing; [`file`]: range-serving a file off `/data`.
+//! - `testkit` (feature `testkit`): conformance checks for any implementation.
+//!
+//! Bytes on disk: every plugin mounts the same `/data` volume as the hull, so a
+//! plugin-written file is opened by the hull directly (local-first reads) and
+//! `rename(2)` between `tmp/` and `cache/` stays atomic.
+
+pub mod cid;
+pub mod client;
+pub mod dto;
+pub mod error;
+pub mod file;
+pub mod focus;
+pub mod fsutil;
+pub mod hull;
+pub mod ipfs;
+pub mod magic;
+pub mod nzb;
+pub mod plugin;
+pub mod range;
+pub mod serve;
+pub mod torrent;
+#[cfg(any(test, feature = "testkit"))]
+pub mod testkit;
+
+pub use client::RemoteTransport;
+pub use dto::{
+    Capabilities, Deleted, Event, Health, Job, Jobs, Manifest, PromotedFile, ReconcileReport,
+    ReconcileRequest, CONTRACT_VERSION,
+};
+pub use error::ApiError;
+pub use focus::{FocusSnapshot, FocusTitle, FocusView, Lane};
+pub use hull::{HullClient, RecordInfo};
+pub use plugin::TransportPlugin;
+pub use serve::serve_transport;

src/transport/nzb.rs

@@ -0,0 +1,560 @@
+//! Usenet cid shapes, shared by the hull and the usenet transport plugin, plus
+//! the plugin's extra-route DTOs and its settings.
+//!
+//! The codec helpers are moved verbatim from meta-share `nzb/mod.rs`; the hull
+//! dispatches and grades with them, the plugin decodes with them, so they must
+//! be one copy.
+//!
+//! Extra routes of the usenet plugin (all under `/nzb/`):
+//!
+//! - `GET    /nzb/status/:cid`      → the `kind:"nzb"` status JSON (never starts a job)
+//! - `GET    /nzb/jobs`             → `{"jobs":[…]}` in-flight jobs (`NzbActivity` JSON)
+//! - `DELETE /nzb/jobs/:cid`        → `{"cancelled":bool}`
+//! - `GET    /nzb/capability`       → [`Capability`]
+//! - `GET    /nzb/verdict/:cid`     → `Option<String>` remembered verdict
+//! - `POST   /nzb/verify/:cid`      → `Option<String>` verdict, verified if free
+//! - `POST   /nzb/test-nntp`        ← [`NntpTest`] → `{"ok","message"}`
+//! - `POST   /nzb/restart`          → `204`, then the process exits (settings changed)
+//! - `POST   /nzb/seal/:cid/claim`  → [`SealClaim`] / `409` (not `Ready`, or already claimed)
+//! - `POST   /nzb/seal/:cid/promote`→ [`SealPrepared`] (publish into `cache/` + the hole/magic gates)
+//! - `POST   /nzb/seal/:cid/done`   ← [`SealDone`]
+//! - `POST   /nzb/seal/:cid/abort`  → `204`
+//!
+//! `/raw/:cid` may answer `409` with header [`HDR_REDIRECT_CID`] — "these bytes
+//! now live under that cid (an IPFS root) and my copy is gone": the hull
+//! re-dispatches to it, exactly where the monolith fell back to the blockstore.
+
+use anyhow::{Context, Result};
+use serde::{Deserialize, Serialize};
+
+use super::cid::MsCid;
+
+/// `409` + this header on `/raw`: re-dispatch the request to this cid.
+pub const HDR_REDIRECT_CID: &str = "x-metamesh-redirect-cid";
+
+/// `GET /nzb/capability`.
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct Capability {
+    /// The plugin holds an NNTP pool (host+user+pass configured).
+    pub nntp: bool,
+    /// The NNTP provider host (never the account), when credentialed.
+    #[serde(default)]
+    pub provider: Option<String>,
+}
+
+/// `POST /nzb/test-nntp` — a candidate connection, already merged with the
+/// stored secret by the hull (which owns `settings.json`).
+#[derive(Serialize, Deserialize, Clone, Debug, Default)]
+pub struct NntpTest {
+    pub host: String,
+    pub port: u16,
+    pub user: String,
+    pub pass: String,
+    pub tls: bool,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq)]
+pub struct SealClaim {
+    pub path: String,
+    pub size: u64,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq)]
+pub struct SealPrepared {
+    /// The published file, under `cache/<cid>/`.
+    pub path: String,
+    /// Its name inside the container.
+    pub rel: String,
+    pub size: u64,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq)]
+pub struct SealDone {
+    pub root: String,
+    pub size: u64,
+    pub path: String,
+}
+
+/// Usenet (NZB) retrieval settings — the `usenet` section of the hull's
+/// `settings.json`. The plugin reads the file read-only at boot (the hull owns
+/// it and restarts the plugin when it changes).
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+#[serde(default)]
+pub struct UsenetSettings {
+    pub nntp_host: Option<String>,
+    pub nntp_port: Option<u16>,
+    pub nntp_user: Option<String>,
+    /// Secret.
+    pub nntp_pass: Option<String>,
+    pub nntp_tls: bool,
+    pub nntp_connections: u32,
+}
+
+impl Default for UsenetSettings {
+    fn default() -> Self {
+        Self {
+            nntp_host: None,
+            nntp_port: None,
+            nntp_user: None,
+            nntp_pass: None,
+            nntp_tls: true,
+            nntp_connections: 20,
+        }
+    }
+}
+
+fn env_nonempty(key: &str) -> Option<String> {
+    std::env::var(key)
+        .ok()
+        .map(|s| s.trim().to_string())
+        .filter(|s| !s.is_empty())
+}
+
+impl UsenetSettings {
+    /// DEPRECATED seed path: the legacy `META_SHARE_NNTP_*` env vars. Only
+    /// consulted when `settings.json` doesn't exist yet.
+    pub fn from_env() -> Self {
+        let mut s = Self::default();
+        if let Some(v) = env_nonempty("META_SHARE_NNTP_HOST") {
+            s.nntp_host = Some(v);
+        }
+        if let Some(v) = env_nonempty("META_SHARE_NNTP_PORT").and_then(|v| v.parse().ok()) {
+            s.nntp_port = Some(v);
+        }
+        if let Some(v) = env_nonempty("META_SHARE_NNTP_USER") {
+            s.nntp_user = Some(v);
+        }
+        if let Some(v) = env_nonempty("META_SHARE_NNTP_PASS") {
+            s.nntp_pass = Some(v);
+        }
+        if let Ok(v) = std::env::var("META_SHARE_NNTP_TLS") {
+            s.nntp_tls = !matches!(v.trim(), "0" | "false" | "no");
+        }
+        if let Some(v) = env_nonempty("META_SHARE_NNTP_CONNECTIONS").and_then(|v| v.parse().ok()) {
+            s.nntp_connections = v;
+        }
+        s.nntp_connections = s.nntp_connections.clamp(1, 100);
+        s
+    }
+
+    /// The `usenet` section of `<config_dir>/settings.json`, or the legacy env
+    /// seed when the file is absent or unparseable.
+    pub fn load(config_dir: &std::path::Path) -> Self {
+        #[derive(Deserialize, Default)]
+        #[serde(default)]
+        struct File {
+            usenet: UsenetSettings,
+        }
+        match std::fs::read(config_dir.join("settings.json")) {
+            Ok(bytes) => match serde_json::from_slice::<File>(&bytes) {
+                Ok(f) => f.usenet,
+                Err(_) => Self::from_env(),
+            },
+            Err(_) => Self::from_env(),
+        }
+    }
+}
+
+/// MetaMesh-private multicodec for a self-describing Newznab **release** locator
+/// (`nzb-release`). Mirrors `meta-feeder-sdk`'s `hash::NZB_RELEASE_CODEC`. The
+/// cid's identity multihash embeds `{host, id}` directly, so the feeder emits it
+/// with **no `.nzb` download** and the credentialed peer decodes the locator
+/// straight from the cid at playback — no KV side-table.
+pub const NZB_RELEASE_CODEC: u64 = 0x1005;
+
+/// MetaMesh-private multicodec for a **self-scanned Usenet posting**
+/// (`nzb-posting`). Mirrors `meta-feeder-sdk`'s `hash::NZB_POSTING_CODEC`.
+///
+/// The crucial difference from [`NZB_RELEASE_CODEC`]: this is a **digest over
+/// the article Message-ID set**, not a locator. It embeds no indexer host, so
+/// **any peer with a plain NNTP provider can redeem it — no indexer credential**.
+/// That portability is the entire point of scanning Usenet ourselves.
+///
+/// Two consequences the code must respect:
+/// - It is **not reversible**. There is no `decode_*` for it. The `.nzb`
+///   manifest arrives out of band, as an ordinary sha2-256 cid named by the
+///   record's `manifest` field (see the hull's manifest resolution).
+/// - It is **not bitswap-fetchable under itself**. Bitswap derives a block's
+///   cid by hashing the block; this digest is over the id set, so a want for it
+///   could never be satisfied. Never offer it to the swarm.
+///
+/// See `meta-gateway/docs/others/self-hosted-usenet-indexer-study.md` §5.
+pub const NZB_POSTING_CODEC: u64 = 0x1003;
+
+/// Does this cid address a Usenet **release locator** (codec `0x1005`) — the
+/// self-describing `{host, id}` form that must be redeemed against a specific
+/// indexer with a matching credential?
+///
+/// Narrower than [`is_nzb_cid`] on purpose: only the locator carries a host, so
+/// only the locator gates on an indexer key (held by a gateway's feeder). Use this where
+/// the *host* matters; use [`is_nzb_cid`] where "should this go to the Usenet
+/// transport" matters.
+pub fn is_nzb_locator(mscid: &MsCid) -> bool {
+    mscid.codec() == NZB_RELEASE_CODEC
+}
+
+/// Does this cid address a **self-scanned Usenet posting** (codec `0x1003`)?
+pub fn is_nzb_posting(mscid: &MsCid) -> bool {
+    mscid.codec() == NZB_POSTING_CODEC
+}
+
+/// Does this cid go to the **Usenet transport** at all — either form?
+///
+/// This is the `/raw` dispatch predicate: both shapes end at
+/// the usenet plugin's `/raw`, which resolves a manifest (differently per
+/// form — see the hull's manifest resolution), NNTP-fetches the articles, and
+/// materialises/streams the media. Everything downstream of manifest
+/// resolution is identical for the two.
+pub fn is_nzb_cid(mscid: &MsCid) -> bool {
+    is_nzb_locator(mscid) || is_nzb_posting(mscid)
+}
+
+/// String convenience over [`is_nzb_cid`] — either Usenet form. Used by the
+/// playback-focus endpoints, which see the bare CID string. There is no
+/// locator-only counterpart: every string-typed call site that used to ask
+/// "is this specifically a `0x1005` locator" now needs "is this Usenet at
+/// all" — see the six dispatch sites widened alongside `nzb-posting`.
+pub fn cid_is_nzb(cid: &str) -> bool {
+    super::cid::parse_record_cid(cid)
+        .map(|c| is_nzb_cid(&c))
+        .unwrap_or(false)
+}
+
+/// Re-derive an `nzb-posting` digest from a manifest's Message-IDs.
+///
+/// ⚠ **MIRRORED LOGIC** — byte-for-byte equivalent to
+/// `meta_feeder_sdk::hash::compute_nzb_posting_cid` (the gateway/feeder side
+/// mints with it; we verify with it). The two crates cannot share a path-dep,
+/// so the normalisation rule is duplicated and must be changed on both sides
+/// together. A drift here does not error — it silently makes every self-scanned
+/// posting fail verification and become unplayable.
+///
+/// The rule: trim, strip a surrounding `<…>`, drop empties, dedupe, sort
+/// lexicographically, join with `\n`, SHA-256. Case is preserved (a Message-ID
+/// local part is case-sensitive per RFC 5322).
+pub fn nzb_posting_digest<S: AsRef<str>>(message_ids: &[S]) -> [u8; 32] {
+    use sha2::{Digest, Sha256};
+
+    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());
+    }
+    hasher.finalize().into()
+}
+
+/// A decoded self-describing `nzb-release` locator: the Newznab **API base**,
+/// recovered from the cid's identity-multihash digest. The release id stays in
+/// the cid — only the feeder that redeems it reads it.
+pub struct NzbLocator {
+    /// The indexer's Newznab API base, **scheme-less, authority + optional
+    /// path** (`api.nzb.life`, `api.nzbgeek.info/api`). Stamped by the feeder
+    /// from the indexer aggregator's own configured `baseUrl` — i.e. the URL
+    /// that aggregator itself queries, so it is redeemable by construction.
+    ///
+    /// ⚠ This used to be a bare `host` taken from the release's `<guid>`
+    /// **permalink**, which is only the API host by coincidence. nzb.life
+    /// serves both from `api.nzb.life` so it worked; nzbgeek's permalinks are on
+    /// `nzbgeek.info` while its API is on `api.nzbgeek.info/api`, so every
+    /// nzbgeek release was unredeemable — `https://nzbgeek.info/api?t=get`
+    /// answers 302→HTML and the grab died in `nzb xml parse`.
+    ///
+    /// Because it may carry a path it is **not** a matching key: use
+    /// [`authority()`](Self::authority) for that.
+    pub api_base: String,
+}
+
+impl NzbLocator {
+    /// The host to match against the gateways' redeem claims — the authority
+    /// alone, path stripped.
+    ///
+    /// The base may carry a path (`api.nzbgeek.info/api`) that a claim never
+    /// does: operators configure, and feeders advertise, plain hosts.
+    pub fn authority(&self) -> &str {
+        match self.api_base.split_once('/') {
+            Some((authority, _path)) => authority,
+            None => &self.api_base,
+        }
+    }
+}
+
+/// Normalise an indexer host for matching: drop any `http(s)://` scheme, trim a
+/// trailing slash and surrounding whitespace, lowercase. A locator's
+/// [`NzbLocator::authority`] passes through this before it is compared with the
+/// hosts a gateway advertises, so a claim for `https://api.example.com/` and one
+/// for `api.example.com` match the same releases.
+pub fn normalize_indexer_host(raw: &str) -> String {
+    raw.trim()
+        .trim_start_matches("https://")
+        .trim_start_matches("http://")
+        .trim_end_matches('/')
+        .to_ascii_lowercase()
+}
+
+/// Decode a self-describing `nzb-release` cid (codec `0x1005`) into its
+/// API base. The identity-multihash digest is
+/// `varint(base_len) ‖ api_base ‖ id_bytes` — the exact inverse of
+/// meta-feeder-sdk's `hash::compute_nzb_release_cid`; `id_bytes` is the
+/// hex-decoded release id, which only the redeeming feeder reads. Errors
+/// (mapped to 4xx by the caller) on a wrong codec or a malformed digest.
+///
+/// The wire shape is unchanged from when this field held a bare host — only the
+/// *meaning* of those bytes widened to "authority + optional path", so old cids
+/// still decode (they just name an unredeemable host; see [`NzbLocator`]).
+pub fn decode_nzb_release_cid(mscid: &MsCid) -> Result<NzbLocator> {
+    anyhow::ensure!(
+        mscid.codec() == NZB_RELEASE_CODEC,
+        "not an nzb-release cid (codec 0x{:x}, want 0x{:x})",
+        mscid.codec(),
+        NZB_RELEASE_CODEC
+    );
+    let digest = mscid.hash().digest();
+    let (base_len, consumed) =
+        super::cid::read_uvarint(digest).context("nzb-release digest: base-length varint")?;
+    let base_len = usize::try_from(base_len)
+        .map_err(|_| anyhow::anyhow!("nzb-release base length {base_len} overflows usize"))?;
+    let rest = &digest[consumed..];
+    anyhow::ensure!(
+        rest.len() > base_len,
+        "nzb-release digest truncated: base_len={base_len}, only {} byte(s) left (need base + id)",
+        rest.len()
+    );
+    let base_bytes = &rest[..base_len];
+    let api_base = std::str::from_utf8(base_bytes)
+        .context("nzb-release api base is not valid utf-8")?
+        .to_string();
+    Ok(NzbLocator { api_base })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::transport::cid::parse_record_cid;
+
+    #[test]
+    fn settings_load_reads_the_usenet_section_and_ignores_the_rest() {
+        let d = std::env::temp_dir().join(format!("nzb-settings-{}", std::process::id()));
+        std::fs::create_dir_all(&d).unwrap();
+        std::fs::write(
+            d.join("settings.json"),
+            r#"{"network":{"meta_core_url":"x"},"usenet":{"nntp_host":"news.example","nntp_user":"u","nntp_pass":"p","nntp_tls":true,"nntp_connections":7,"indexers":[1]}}"#,
+        )
+        .unwrap();
+        let s = UsenetSettings::load(&d);
+        assert_eq!(s.nntp_host.as_deref(), Some("news.example"));
+        assert_eq!(s.nntp_connections, 7);
+        let _ = std::fs::remove_dir_all(&d);
+    }
+
+
+    /// RFC 4648 base32 lowercase, no padding (multibase `b` body).
+    fn base32_lower(input: &[u8]) -> String {
+        const ALPHABET: &[u8; 32] = b"abcdefghijklmnopqrstuvwxyz234567";
+        let mut out = String::new();
+        let (mut buffer, mut bits): (u64, u32) = (0, 0);
+        for &b in input {
+            buffer = (buffer << 8) | b as u64;
+            bits += 8;
+            while bits >= 5 {
+                bits -= 5;
+                out.push(ALPHABET[((buffer >> bits) & 0x1F) as usize] as char);
+            }
+        }
+        if bits > 0 {
+            out.push(ALPHABET[((buffer << (5 - bits)) & 0x1F) as usize] as char);
+        }
+        out
+    }
+
+    fn varint(mut v: u64, out: &mut Vec<u8>) {
+        while v >= 0x80 {
+            out.push((v as u8) | 0x80);
+            v >>= 7;
+        }
+        out.push(v as u8);
+    }
+
+    /// Build a self-describing nzb-release cid exactly as meta-feeder-sdk's
+    /// `compute_nzb_release_cid` does: identity multihash over
+    /// `varint(host_len) ‖ host ‖ id_bytes`.
+    fn make_nzb_cid(host: &str, id_bytes: &[u8]) -> String {
+        let mut digest = Vec::new();
+        varint(host.len() as u64, &mut digest);
+        digest.extend_from_slice(host.as_bytes());
+        digest.extend_from_slice(id_bytes);
+        let mut wire = vec![0x01u8];
+        varint(NZB_RELEASE_CODEC, &mut wire);
+        wire.push(0x00); // identity multihash
+        varint(digest.len() as u64, &mut wire);
+        wire.extend_from_slice(&digest);
+        format!("b{}", base32_lower(&wire))
+    }
+
+    #[test]
+    fn decode_round_trips_the_base() {
+        let cid = make_nzb_cid("api.example.com", &[0x48, 0x5c, 0x52, 0xf0]);
+        let mscid = parse_record_cid(&cid).expect("parse");
+        assert!(is_nzb_locator(&mscid));
+        let loc = decode_nzb_release_cid(&mscid).expect("decode");
+        assert_eq!(loc.api_base, "api.example.com");
+        // No path ⇒ authority is the whole thing.
+        assert_eq!(loc.authority(), "api.example.com");
+    }
+
+    /// The nzbgeek shape, and the whole reason `authority()` exists: its API
+    /// lives at `https://api.nzbgeek.info/api` while a feeder claims the plain
+    /// host. The base keeps the path; every match must drop it.
+    #[test]
+    fn a_base_with_a_path_matches_on_its_authority() {
+        let cid = make_nzb_cid("api.nzbgeek.info/api", &[0x19, 0xfe, 0x38, 0xe4]);
+        let mscid = parse_record_cid(&cid).expect("parse");
+        let loc = decode_nzb_release_cid(&mscid).expect("decode");
+        assert_eq!(loc.api_base, "api.nzbgeek.info/api");
+        assert_eq!(
+            loc.authority(),
+            "api.nzbgeek.info",
+            "can_redeem / probe / redeem claims all compare on the authority"
+        );
+    }
+
+    /// A port is part of the authority, not a path — it must survive the split.
+    #[test]
+    fn a_port_stays_in_the_authority() {
+        let cid = make_nzb_cid("indexer.local:5076/api", &[0x01, 0x02]);
+        let mscid = parse_record_cid(&cid).expect("parse");
+        let loc = decode_nzb_release_cid(&mscid).expect("decode");
+        assert_eq!(loc.authority(), "indexer.local:5076");
+    }
+
+    #[test]
+    fn decode_rejects_wrong_codec() {
+        // codec 0x1001 (btih-v1-file), not an nzb-release locator.
+        let mut wire = vec![0x01u8];
+        varint(0x1001, &mut wire);
+        wire.extend_from_slice(&[0x00, 0x02, 0xAB, 0xCD]); // identity mh, 2-byte digest
+        let cid = format!("b{}", base32_lower(&wire));
+        let mscid = parse_record_cid(&cid).expect("parse");
+        assert!(!is_nzb_locator(&mscid));
+        assert!(decode_nzb_release_cid(&mscid).is_err());
+    }
+
+    #[test]
+    fn decode_rejects_truncated_digest() {
+        // host_len claims the whole digest, leaving no id bytes.
+        let cid = make_nzb_cid("api.example.com", &[]);
+        let mscid = parse_record_cid(&cid).expect("parse");
+        assert!(decode_nzb_release_cid(&mscid).is_err());
+    }
+
+    /// Build an nzb-posting cid exactly as meta-feeder-sdk's
+    /// `compute_nzb_posting_cid` does: codec 0x1003 + sha2-256 over the
+    /// normalised, sorted, newline-joined Message-ID set.
+    fn make_posting_cid(message_ids: &[&str]) -> String {
+        let digest = nzb_posting_digest(message_ids);
+        let mut wire = vec![0x01u8];
+        varint(NZB_POSTING_CODEC, &mut wire);
+        wire.push(0x12); // sha2-256
+        wire.push(0x20); // 32 bytes
+        wire.extend_from_slice(&digest);
+        format!("b{}", base32_lower(&wire))
+    }
+
+    /// ⚠ THE MIRROR TEST. `nzb_posting_digest` here must agree byte-for-byte
+    /// with `meta_feeder_sdk::hash::compute_nzb_posting_cid`'s digest — the
+    /// gateway mints with one, we verify with the other, and a drift silently
+    /// makes every self-scanned posting unplayable rather than erroring.
+    ///
+    /// This vector is the same one pinned on the SDK side
+    /// (`hash::tests::nzb_posting_cid_wire_shape`): a single id `[email protected]`
+    /// digests to plain `sha256("[email protected]")`.
+    #[test]
+    fn posting_digest_matches_the_sdk_mint() {
+        use sha2::{Digest, Sha256};
+        let expect: [u8; 32] = Sha256::digest(b"[email protected]").into();
+        assert_eq!(nzb_posting_digest(&["[email protected]"]), expect);
+    }
+
+    /// The normalisation rule, mirrored from the SDK: order-independent,
+    /// bracket/blank/dupe-insensitive, case-SENSITIVE.
+    #[test]
+    fn posting_digest_normalisation_matches_the_mint() {
+        // Order does not matter — this is why we hash a sorted set and never
+        // the raw `.nzb` bytes (the retired 0x1004's mistake).
+        assert_eq!(
+            nzb_posting_digest(&["[email protected]", "[email protected]", "[email protected]"]),
+            nzb_posting_digest(&["[email protected]", "[email protected]", "[email protected]"])
+        );
+        // Angle brackets, whitespace, empties and duplicates all normalise away.
+        assert_eq!(
+            nzb_posting_digest(&["[email protected]", "[email protected]"]),
+            nzb_posting_digest(&[" <[email protected]> ", "[email protected]", "", "  ", "[email protected]"])
+        );
+        // Case is preserved (RFC 5322: the local part is case-sensitive).
+        assert_ne!(
+            nzb_posting_digest(&["[email protected]"]),
+            nzb_posting_digest(&["[email protected]"])
+        );
+        // A different article set is a different posting.
+        assert_ne!(
+            nzb_posting_digest(&["[email protected]", "[email protected]"]),
+            nzb_posting_digest(&["[email protected]", "[email protected]"])
+        );
+    }
+
+    /// Both Usenet forms route to the Usenet transport, but only the `0x1005`
+    /// locator is host-bound. Keeping the two predicates distinct is what lets
+    /// `can_redeem` skip the credential check for a posting.
+    #[test]
+    fn posting_and_locator_are_both_nzb_but_only_one_is_a_locator() {
+        let posting = parse_record_cid(&make_posting_cid(&["[email protected]"])).expect("parse posting");
+        assert!(is_nzb_posting(&posting));
+        assert!(is_nzb_cid(&posting), "postings go to the Usenet transport");
+        assert!(
+            !is_nzb_locator(&posting),
+            "a posting names no indexer host, so it must NOT be treated as a locator              (that would send it down the credential-gated grab path)"
+        );
+
+        let locator = parse_record_cid(&make_nzb_cid("api.example.com", &[0xab, 0xcd]))
+            .expect("parse locator");
+        assert!(is_nzb_locator(&locator));
+        assert!(is_nzb_cid(&locator));
+        assert!(!is_nzb_posting(&locator));
+    }
+
+    /// A posting cid must never decode as a release locator — the digest is not
+    /// a `{host, id}` payload, and misreading it would produce a garbage host.
+    #[test]
+    fn posting_cid_does_not_decode_as_a_locator() {
+        let posting = parse_record_cid(&make_posting_cid(&["[email protected]", "[email protected]"]))
+            .expect("parse posting");
+        assert!(decode_nzb_release_cid(&posting).is_err());
+    }
+
+    /// The string convenience wrappers agree with the parsed forms, including
+    /// on garbage input (which must be `false`, never a panic).
+    #[test]
+    fn cid_string_helpers_agree() {
+        let posting = make_posting_cid(&["[email protected]"]);
+        assert!(cid_is_nzb(&posting));
+
+        let locator = make_nzb_cid("api.example.com", &[0x01]);
+        assert!(cid_is_nzb(&locator));
+
+        assert!(!cid_is_nzb("not-a-cid"));
+    }
+}

src/transport/plugin.rs

@@ -0,0 +1,50 @@
+//! The plugin side of the contract: what a transport implementation provides.
+
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use axum::http::HeaderMap;
+use axum::response::Response;
+use axum::Router;
+
+use super::dto::{Deleted, Health, Job, Manifest, ReconcileReport, ReconcileRequest};
+use super::error::ApiError;
+use super::focus::FocusView;
+
+/// One transport implementation (BitTorrent, Usenet, IPFS, or a competing
+/// implementation of one of them). [`serve_transport`](super::serve::serve_transport)
+/// wraps it in the common routes; [`extra_routes`](Self::extra_routes) adds its
+/// protocol-specific ones.
+#[async_trait]
+pub trait TransportPlugin: Send + Sync + 'static {
+    fn manifest(&self) -> Manifest;
+
+    async fn health(&self) -> Health {
+        Health::ok()
+    }
+
+    /// `GET /raw/:cid`. Must answer exactly what the in-process tier answered
+    /// in the monolith (status, `Content-*`, `Retry-After`, error body) — the
+    /// hull relays it to meta-share's public API untouched. Must **not** meter
+    /// its own egress: the hull does, on relay.
+    async fn raw(&self, cid: String, headers: HeaderMap) -> Response;
+
+    /// `GET /jobs`.
+    async fn jobs(&self) -> Result<Vec<Job>, ApiError>;
+
+    /// `DELETE /jobs/:cid?keep=…`. `keep` lists sibling cids of the same unit
+    /// the hull still wants.
+    async fn delete(&self, cid: String, keep: Vec<String>) -> Result<Deleted, ApiError>;
+
+    /// `POST /reconcile`.
+    async fn reconcile(&self, req: ReconcileRequest) -> Result<ReconcileReport, ApiError>;
+
+    /// The view `PUT /focus` writes into.
+    fn focus(&self) -> &Arc<FocusView>;
+
+    /// Protocol-specific routes, merged into the common router. Paths must not
+    /// collide with the common ones.
+    fn extra_routes(self: Arc<Self>) -> Router {
+        Router::new()
+    }
+}

src/transport/range.rs

@@ -0,0 +1,277 @@
+//! HTTP `Range:` request-header parser.
+//!
+//! Pure helper: takes an inbound `Range:` header value plus the total
+//! resource size in bytes, returns a [`RangeOutcome`] describing how the
+//! caller should respond.
+//!
+//! Shared by the meta-share hull and every transport plugin so the
+//! `<video>` element can seek and resume downloads identically whichever
+//! process serves the bytes. Moved verbatim from meta-share
+//! `crates/meta-share/src/api/range.rs`.
+//!
+//! ## Supported syntax
+//!
+//! RFC 7233 §3.1 `bytes=`-prefixed single-range only. Multi-range
+//! requests (`bytes=0-99,200-299`) are surfaced as
+//! [`RangeOutcome::Unsupported`] so the caller can fall back to `200 OK`
+//! with the full body — the RFC permits ignoring multi-range, and
+//! every common client (browser video element, curl, wget) accepts the
+//! degraded response.
+//!
+//! ### Returned shape
+//!
+//! - `bytes=N-M`   → `Ok((N, min(M, total-1)))`
+//! - `bytes=N-`    → `Ok((N, total-1))`
+//! - `bytes=-M`    → `Ok((max(0, total-M), total-1))` (suffix)
+//! - `bytes=N-`    with `N >= total` → `Unsatisfiable`
+//! - `bytes=N-M`   with `N > M`      → `Unsatisfiable`
+//! - missing / blank / malformed     → `Unsatisfiable` for non-`Absent`,
+//!   `Absent` only for the literal "no header" case (callers pass
+//!   `Option<&HeaderValue>`).
+//!
+//! The function is total: it never panics, never allocates, and always
+//! returns one of the four variants.
+
+use axum::http::HeaderValue;
+
+/// Outcome of parsing an inbound `Range:` header.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum RangeOutcome {
+    /// Header absent on the request — caller serves `200 OK` with the
+    /// full body.
+    Absent,
+    /// Single-range request parsed cleanly. Inclusive `[start, end]`
+    /// clamped to `[0, total-1]`.
+    Ok { start: u64, end: u64 },
+    /// Multi-range or unsupported syntax — caller serves `200 OK` with
+    /// the full body (the RFC permits ignoring; browser video and curl
+    /// both fall through to a full GET on a follow-up request).
+    Unsupported,
+    /// Malformed range, reversed range, or `start >= total` — caller
+    /// serves `416 Range Not Satisfiable` with
+    /// `Content-Range: bytes */{total}`.
+    Unsatisfiable,
+}
+
+/// Parse a `Range:` header into a [`RangeOutcome`]. `header` is the raw
+/// header value (or `None` if the request didn't carry one). `total` is
+/// the resource's full size in bytes — used to clamp open / suffix
+/// ranges and to detect out-of-bounds requests.
+pub fn parse_range_header(header: Option<&HeaderValue>, total: u64) -> RangeOutcome {
+    let Some(value) = header else {
+        return RangeOutcome::Absent;
+    };
+    let Ok(s) = value.to_str() else {
+        return RangeOutcome::Unsatisfiable;
+    };
+    let s = s.trim();
+    // Strip the `bytes=` unit prefix. Anything else is unsupported
+    // (per RFC the unit is extensible, but in practice `bytes` is the
+    // only one any browser sends).
+    let Some(rest) = s.strip_prefix("bytes=") else {
+        return RangeOutcome::Unsatisfiable;
+    };
+    let rest = rest.trim();
+    if rest.is_empty() {
+        return RangeOutcome::Unsatisfiable;
+    }
+    // Multi-range: caller serves 200 with the full body.
+    if rest.contains(',') {
+        return RangeOutcome::Unsupported;
+    }
+    // Must contain exactly one '-' separator.
+    let (lhs, rhs) = match rest.split_once('-') {
+        Some(parts) => parts,
+        None => return RangeOutcome::Unsatisfiable,
+    };
+    let lhs = lhs.trim();
+    let rhs = rhs.trim();
+    // Three valid shapes:
+    //   - "N-M"  bounded range
+    //   - "N-"   open-ended (start known, end = total-1)
+    //   - "-M"   suffix (last M bytes)
+    if lhs.is_empty() && rhs.is_empty() {
+        return RangeOutcome::Unsatisfiable;
+    }
+    // Edge case: file is empty. Any range against a 0-byte resource is
+    // unsatisfiable (RFC 7233 §4.4).
+    if total == 0 {
+        return RangeOutcome::Unsatisfiable;
+    }
+    if lhs.is_empty() {
+        // Suffix range. `rhs` is the suffix length.
+        let Ok(suffix_len) = rhs.parse::<u64>() else {
+            return RangeOutcome::Unsatisfiable;
+        };
+        if suffix_len == 0 {
+            return RangeOutcome::Unsatisfiable;
+        }
+        let start = total.saturating_sub(suffix_len);
+        return RangeOutcome::Ok {
+            start,
+            end: total - 1,
+        };
+    }
+    let Ok(start) = lhs.parse::<u64>() else {
+        return RangeOutcome::Unsatisfiable;
+    };
+    if start >= total {
+        return RangeOutcome::Unsatisfiable;
+    }
+    let end = if rhs.is_empty() {
+        total - 1
+    } else {
+        let Ok(parsed_end) = rhs.parse::<u64>() else {
+            return RangeOutcome::Unsatisfiable;
+        };
+        if parsed_end < start {
+            return RangeOutcome::Unsatisfiable;
+        }
+        // Clamp end to total-1 (RFC: "the last byte position is at most
+        // total-1").
+        parsed_end.min(total - 1)
+    };
+    RangeOutcome::Ok { start, end }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn hv(s: &str) -> HeaderValue {
+        HeaderValue::from_str(s).expect("static header value")
+    }
+
+    #[test]
+    fn absent_when_no_header() {
+        assert_eq!(parse_range_header(None, 1000), RangeOutcome::Absent);
+    }
+
+    #[test]
+    fn closed_range_in_bounds() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=0-99")), 1000),
+            RangeOutcome::Ok { start: 0, end: 99 }
+        );
+    }
+
+    #[test]
+    fn open_end_clamped_to_total_minus_one() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=100-")), 1000),
+            RangeOutcome::Ok { start: 100, end: 999 }
+        );
+    }
+
+    #[test]
+    fn suffix_range() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=-512")), 1000),
+            RangeOutcome::Ok { start: 488, end: 999 }
+        );
+    }
+
+    #[test]
+    fn single_byte_range() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=0-0")), 1000),
+            RangeOutcome::Ok { start: 0, end: 0 }
+        );
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=999-999")), 1000),
+            RangeOutcome::Ok { start: 999, end: 999 }
+        );
+    }
+
+    #[test]
+    fn suffix_clamps_to_whole_file_when_larger_than_total() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=-99999")), 1000),
+            RangeOutcome::Ok { start: 0, end: 999 }
+        );
+    }
+
+    #[test]
+    fn end_clamped_when_above_total() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=0-1000")), 1000),
+            RangeOutcome::Ok { start: 0, end: 999 }
+        );
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=0-999999")), 1000),
+            RangeOutcome::Ok { start: 0, end: 999 }
+        );
+    }
+
+    #[test]
+    fn start_at_or_past_total_is_unsatisfiable() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=1000-")), 1000),
+            RangeOutcome::Unsatisfiable
+        );
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=2000-3000")), 1000),
+            RangeOutcome::Unsatisfiable
+        );
+    }
+
+    #[test]
+    fn reversed_range_is_unsatisfiable() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=10-9")), 1000),
+            RangeOutcome::Unsatisfiable
+        );
+    }
+
+    #[test]
+    fn garbage_input_is_unsatisfiable() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=abc")), 1000),
+            RangeOutcome::Unsatisfiable
+        );
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=")), 1000),
+            RangeOutcome::Unsatisfiable
+        );
+        assert_eq!(
+            parse_range_header(Some(&hv("blocks=0-99")), 1000),
+            RangeOutcome::Unsatisfiable
+        );
+        assert_eq!(
+            parse_range_header(Some(&hv("0-99")), 1000),
+            RangeOutcome::Unsatisfiable
+        );
+    }
+
+    #[test]
+    fn multi_range_is_unsupported() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=0-50,100-150")), 1000),
+            RangeOutcome::Unsupported
+        );
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=0-50, 100-150")), 1000),
+            RangeOutcome::Unsupported
+        );
+    }
+
+    #[test]
+    fn empty_resource_is_always_unsatisfiable() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=0-99")), 0),
+            RangeOutcome::Unsatisfiable
+        );
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=-10")), 0),
+            RangeOutcome::Unsatisfiable
+        );
+    }
+
+    #[test]
+    fn zero_length_suffix_is_unsatisfiable() {
+        assert_eq!(
+            parse_range_header(Some(&hv("bytes=-0")), 1000),
+            RangeOutcome::Unsatisfiable
+        );
+    }
+}

src/transport/serve.rs

@@ -0,0 +1,130 @@
+//! `serve_transport` — the axum harness every transport plugin binary runs.
+//!
+//! # Common routes
+//!
+//! - `GET    /manifest`        → [`Manifest`](super::dto::Manifest)
+//! - `GET    /health`          → [`Health`](super::dto::Health)
+//! - `GET    /raw/:cid`        → the tier's byte response, verbatim
+//! - `GET    /jobs`            → [`Jobs`](super::dto::Jobs)
+//! - `DELETE /jobs/:cid?keep=` → [`Deleted`](super::dto::Deleted)
+//! - `POST   /reconcile`       → [`ReconcileReport`](super::dto::ReconcileReport)
+//! - `PUT    /focus`           ← [`FocusSnapshot`](super::focus::FocusSnapshot), `204`
+//! - `GET    /focus`           → the snapshot the plugin currently obeys
+//!
+//! plus whatever [`TransportPlugin::extra_routes`] adds.
+
+use std::net::SocketAddr;
+use std::sync::Arc;
+
+use axum::extract::{Path, Query, State};
+use axum::http::{HeaderMap, StatusCode};
+use axum::response::{IntoResponse, Response};
+use axum::routing::{delete, get, post};
+use axum::{Json, Router};
+use tracing::info;
+
+use super::dto::{DeleteQuery, Jobs, ReconcileRequest};
+use super::focus::FocusSnapshot;
+use super::plugin::TransportPlugin;
+
+type Shared = Arc<dyn TransportPlugin>;
+
+/// The common router for `plugin`, merged with its extra routes.
+pub fn router<P: TransportPlugin>(plugin: Arc<P>) -> Router {
+    let extra = Arc::clone(&plugin).extra_routes();
+    let shared: Shared = plugin;
+    Router::new()
+        .route("/manifest", get(manifest))
+        .route("/health", get(health))
+        .route("/raw/:cid", get(raw))
+        .route("/jobs", get(jobs))
+        .route("/jobs/:cid", delete(delete_job))
+        .route("/reconcile", post(reconcile))
+        .route("/focus", get(get_focus).put(put_focus))
+        .with_state(shared)
+        .merge(extra)
+}
+
+/// Bind `listen` and serve `plugin` until SIGTERM / Ctrl-C.
+pub async fn serve_transport<P: TransportPlugin>(
+    plugin: Arc<P>,
+    listen: SocketAddr,
+) -> anyhow::Result<()> {
+    let id = plugin.manifest().id;
+    let app = router(plugin);
+    let listener = tokio::net::TcpListener::bind(listen).await?;
+    info!(%listen, plugin = %id, "transport plugin listening");
+    axum::serve(listener, app)
+        .with_graceful_shutdown(shutdown_signal())
+        .await?;
+    Ok(())
+}
+
+/// Resolves on SIGTERM (docker stop) or Ctrl-C.
+pub async fn shutdown_signal() {
+    let ctrl_c = async {
+        let _ = tokio::signal::ctrl_c().await;
+    };
+    #[cfg(unix)]
+    let term = async {
+        if let Ok(mut s) =
+            tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate())
+        {
+            s.recv().await;
+        }
+    };
+    #[cfg(not(unix))]
+    let term = std::future::pending::<()>();
+    tokio::select! {
+        _ = ctrl_c => {},
+        _ = term => {},
+    }
+}
+
+async fn manifest(State(p): State<Shared>) -> Response {
+    Json(p.manifest()).into_response()
+}
+
+async fn health(State(p): State<Shared>) -> Response {
+    let h = p.health().await;
+    let code = if h.ok { StatusCode::OK } else { StatusCode::SERVICE_UNAVAILABLE };
+    (code, Json(h)).into_response()
+}
+
+async fn raw(State(p): State<Shared>, Path(cid): Path<String>, headers: HeaderMap) -> Response {
+    p.raw(cid, headers).await
+}
+
+async fn jobs(State(p): State<Shared>) -> Response {
+    match p.jobs().await {
+        Ok(jobs) => Json(Jobs { jobs }).into_response(),
+        Err(e) => e.into_response(),
+    }
+}
+
+async fn delete_job(
+    State(p): State<Shared>,
+    Path(cid): Path<String>,
+    Query(q): Query<DeleteQuery>,
+) -> Response {
+    match p.delete(cid, q.keep_list()).await {
+        Ok(d) => Json(d).into_response(),
+        Err(e) => e.into_response(),
+    }
+}
+
+async fn reconcile(State(p): State<Shared>, Json(req): Json<ReconcileRequest>) -> Response {
+    match p.reconcile(req).await {
+        Ok(r) => Json(r).into_response(),
+        Err(e) => e.into_response(),
+    }
+}
+
+async fn get_focus(State(p): State<Shared>) -> Response {
+    Json(p.focus().snapshot()).into_response()
+}
+
+async fn put_focus(State(p): State<Shared>, Json(snap): Json<FocusSnapshot>) -> StatusCode {
+    p.focus().apply(snap);
+    StatusCode::NO_CONTENT
+}

src/transport/testkit.rs

@@ -0,0 +1,168 @@
+//! Conformance checks any transport plugin must pass — ours or a competing
+//! implementation (a qBittorrent-backed torrent plugin, say). Run them against a
+//! live plugin URL, or in a unit test via [`spawn_local`].
+//!
+//! The checks cover the *common* contract only. A plugin's byte behaviour is
+//! verified by its own tests and by meta-share's contract suite against the
+//! hull, which is where "identical to the monolith" is decided.
+
+use std::net::SocketAddr;
+use std::sync::Arc;
+use std::time::Duration;
+
+use super::client::RemoteTransport;
+use super::dto::{Deleted, CONTRACT_VERSION};
+use super::focus::{FocusSnapshot, FocusTitle};
+use super::plugin::TransportPlugin;
+
+/// A well-formed cid no plugin holds a job for: the raw/sha2-256 CID of the
+/// empty block.
+pub const UNKNOWN_CID: &str = "bafkreihdwdcefgh4dqkjv67uzcmw7ojee6xedzdetojuzjevtenxquvyku";
+
+/// Serve `plugin` on an ephemeral localhost port. Returns its address; the
+/// server lives until the runtime shuts down.
+pub async fn spawn_local<P: TransportPlugin>(plugin: Arc<P>) -> SocketAddr {
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.expect("bind");
+    let addr = listener.local_addr().expect("addr");
+    let app = super::serve::router(plugin);
+    tokio::spawn(async move {
+        let _ = axum::serve(listener, app).await;
+    });
+    addr
+}
+
+/// Run every check against the plugin at `base`. `Ok(())` or the list of
+/// failures, each a one-line explanation.
+pub async fn conformance(base: &str) -> Result<(), Vec<String>> {
+    let http = reqwest::Client::new();
+    let t = RemoteTransport::new("under-test", base, http.clone())
+        .with_control_timeout(Duration::from_secs(10));
+    let mut fails = Vec::new();
+
+    match t.manifest().await {
+        Ok(m) => {
+            if m.id.is_empty() {
+                fails.push("manifest.id is empty".into());
+            }
+            if m.contract != CONTRACT_VERSION {
+                fails.push(format!("manifest.contract {} != {CONTRACT_VERSION}", m.contract));
+            }
+            if !m.capabilities.fetch {
+                fails.push("manifest.capabilities.fetch is false — a transport must fetch".into());
+            }
+        }
+        Err(e) => fails.push(format!("GET /manifest: {e}")),
+    }
+
+    match http.get(t.url("/health")).send().await {
+        Ok(r) if r.status().is_success() || r.status().as_u16() == 503 => {
+            if r.json::<super::dto::Health>().await.is_err() {
+                fails.push("GET /health body is not a Health".into());
+            }
+        }
+        Ok(r) => fails.push(format!("GET /health: status {}", r.status())),
+        Err(e) => fails.push(format!("GET /health: {e}")),
+    }
+
+    if let Err(e) = t.jobs().await {
+        fails.push(format!("GET /jobs: {e}"));
+    }
+
+    let snap = FocusSnapshot {
+        enabled: true,
+        bg_rate_bytes: 4321,
+        titles: vec![FocusTitle {
+            group: "conformance-group".into(),
+            cids: vec!["conformance-cid".into()],
+            infohashes: vec![],
+        }],
+    };
+    match t.push_focus(&snap).await {
+        Ok(()) => match t.get_json::<FocusSnapshot>("/focus").await {
+            Ok(back) if back == snap => {}
+            Ok(back) => fails.push(format!("GET /focus returned {back:?}, pushed {snap:?}")),
+            Err(e) => fails.push(format!("GET /focus: {e}")),
+        },
+        Err(e) => fails.push(format!("PUT /focus: {e}")),
+    }
+    // Leave the plugin unfocused.
+    let _ = t.push_focus(&FocusSnapshot::default()).await;
+
+    match http.delete(t.url(&format!("/jobs/{UNKNOWN_CID}"))).send().await {
+        Ok(r) if r.status().as_u16() == 404 => {}
+        Ok(r) if r.status().is_success() => match r.json::<Deleted>().await {
+            Ok(d) if d.bytes_freed == 0 => {}
+            Ok(d) => fails.push(format!("DELETE unknown job freed {} bytes", d.bytes_freed)),
+            Err(e) => fails.push(format!("DELETE unknown job: body: {e}")),
+        },
+        Ok(r) => fails.push(format!("DELETE unknown job: status {}", r.status())),
+        Err(e) => fails.push(format!("DELETE unknown job: {e}")),
+    }
+
+    match http.get(t.url("/raw/not-a-cid")).send().await {
+        Ok(r) if r.status().is_client_error() => {
+            let v: Option<serde_json::Value> = r.json().await.ok();
+            if !v.as_ref().is_some_and(|v| v.get("error").is_some()) {
+                fails.push("GET /raw/not-a-cid: 4xx without an {\"error\"} body".into());
+            }
+        }
+        Ok(r) => fails.push(format!("GET /raw/not-a-cid: expected 4xx, got {}", r.status())),
+        Err(e) => fails.push(format!("GET /raw/not-a-cid: {e}")),
+    }
+
+    if fails.is_empty() {
+        Ok(())
+    } else {
+        Err(fails)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::transport::dto::{Capabilities, Job, Manifest, ReconcileReport, ReconcileRequest};
+    use crate::transport::error::ApiError;
+    use crate::transport::focus::FocusView;
+    use async_trait::async_trait;
+    use axum::http::HeaderMap;
+    use axum::response::{IntoResponse, Response};
+
+    /// The smallest plugin that should pass: holds nothing, fetches nothing.
+    struct Empty {
+        focus: Arc<FocusView>,
+    }
+
+    #[async_trait]
+    impl TransportPlugin for Empty {
+        fn manifest(&self) -> Manifest {
+            Manifest {
+                id: "empty".into(),
+                implementation: "testkit-stub".into(),
+                version: "0".into(),
+                contract: CONTRACT_VERSION,
+                capabilities: Capabilities { fetch: true, share: false },
+            }
+        }
+        async fn raw(&self, _cid: String, _h: HeaderMap) -> Response {
+            ApiError::BadRequest("unsupported cid".into()).into_response()
+        }
+        async fn jobs(&self) -> Result<Vec<Job>, ApiError> {
+            Ok(vec![])
+        }
+        async fn delete(&self, _cid: String, _keep: Vec<String>) -> Result<Deleted, ApiError> {
+            Ok(Deleted::default())
+        }
+        async fn reconcile(&self, _r: ReconcileRequest) -> Result<ReconcileReport, ApiError> {
+            Ok(ReconcileReport::default())
+        }
+        fn focus(&self) -> &Arc<FocusView> {
+            &self.focus
+        }
+    }
+
+    #[tokio::test]
+    async fn the_empty_plugin_conforms() {
+        let addr = spawn_local(Arc::new(Empty { focus: FocusView::new() })).await;
+        conformance(&format!("http://{addr}")).await.unwrap();
+    }
+}

src/transport/torrent.rs

@@ -0,0 +1,263 @@
+//! BitTorrent-specific half of the contract: the torrent plugin's extra routes,
+//! and the magnet helpers both sides use.
+//!
+//! Extra routes (all under `/torrent/`):
+//!
+//! - `GET  /torrent/inventory`         → [`Inventory`] — every managed torrent,
+//!   with have-bytes, per-file progress and live stats. One call per eviction
+//!   pass / dashboard poll instead of one per row.
+//! - `GET  /torrent/status/:cid`       → `Option<`[`LiveStats`]`>` — peeks only,
+//!   never starts the session (meta-watch polls it while connecting).
+//! - `GET  /torrent/peers/:cid?budget_ms=&cap=` → [`PeerCount`] — query-only
+//!   mainline-DHT population for the availability probe.
+//! - `POST /torrent/confirm`           ← [`ConfirmRequest`] → [`ConfirmResponse`]
+//!   — metadata-only reachability probe (no torrent kept, no seed row).
+//! - `PUT  /torrent/commits`           ← [`Commits`] — the watch-commit intents.
+//!   The hull owns the intent (`seeds.json`); the plugin's supervisor owns the
+//!   fill (disk floor, concurrency cap, stall hysteresis).
+
+use serde::{Deserialize, Serialize};
+
+/// Live transfer stats for one torrent. Serialises exactly like meta-share's
+/// former `bt::SeedLiveStats`, which `/api/file/:cid/status` and `/api/seeds`
+/// still emit.
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq)]
+pub struct LiveStats {
+    /// librqbit torrent state: `live` | `paused` | `initializing` | `error`.
+    pub state: String,
+    pub down_bps: u64,
+    pub up_bps: u64,
+    pub uploaded_bytes: u64,
+    pub downloaded_bytes: u64,
+    pub total_bytes: u64,
+    pub ratio: f64,
+    pub peers_live: usize,
+    pub peers_seen: usize,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq)]
+pub struct FileProgress {
+    pub index: usize,
+    /// File length from the torrent metadata.
+    pub len: u64,
+    /// Bytes of it on disk.
+    pub have: u64,
+}
+
+impl FileProgress {
+    pub fn complete(&self) -> bool {
+        self.have >= self.len
+    }
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq)]
+pub struct TorrentInfo {
+    /// Lowercase hex v1 infohash.
+    pub infohash: String,
+    #[serde(default)]
+    pub name: Option<String>,
+    /// librqbit have-bytes — the eviction size of the unit.
+    pub have_bytes: u64,
+    pub streaming_only: bool,
+    /// Metadata resolved (`files` is empty until it is).
+    pub metadata: bool,
+    #[serde(default)]
+    pub files: Vec<FileProgress>,
+    pub stats: LiveStats,
+}
+
+impl TorrentInfo {
+    /// Is file `index` fully downloaded? `None` when unknowable yet.
+    pub fn file_complete(&self, index: usize) -> Option<bool> {
+        self.files.iter().find(|f| f.index == index).map(FileProgress::complete)
+    }
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq)]
+pub struct Inventory {
+    /// Is the librqbit session up? When it isn't, `torrents` is empty and the
+    /// hull falls back to row lengths (the session is lazy on purpose).
+    pub session_up: bool,
+    /// Does the session directory hold persisted torrents?
+    pub persisted: bool,
+    pub torrents: Vec<TorrentInfo>,
+}
+
+impl Inventory {
+    pub fn get(&self, infohash_hex: &str) -> Option<&TorrentInfo> {
+        self.torrents.iter().find(|t| t.infohash == infohash_hex)
+    }
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub struct PeerCount {
+    /// Distinct DHT peers seen; `None` when the DHT couldn't be queried.
+    pub peers: Option<usize>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default)]
+pub struct ConfirmRequest {
+    pub magnet: String,
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default)]
+pub struct ConfirmResponse {
+    /// Length of file 0.
+    pub len: u64,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct CommitIntent {
+    pub cid: String,
+    /// The seed row's `added_at` — the supervisor's recency order.
+    pub added_at: u64,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq, Eq)]
+pub struct Commits {
+    pub intents: Vec<CommitIntent>,
+}
+
+// ---------------------------------------------------------------------------
+// Magnet helpers (moved verbatim from meta-share `bt.rs`)
+// ---------------------------------------------------------------------------
+
+/// A bare-infohash magnet: DHT resolves peers from the infohash alone.
+pub fn build_magnet(infohash20: &[u8; 20]) -> String {
+    format!("magnet:?xt=urn:btih:{}", super::cid::to_hex(infohash20))
+}
+
+/// High-uptime open trackers appended to every magnet before librqbit sees it,
+/// so a trackerless indexer magnet doesn't leave peer discovery to DHT alone.
+pub const DEFAULT_TRACKERS: &[&str] = &[
+    "udp://tracker.opentrackr.org:1337/announce",
+    "udp://open.stealth.si:80/announce",
+    "udp://tracker.openbittorrent.com:6969/announce",
+    "udp://exodus.desync.com:6969/announce",
+    "udp://tracker.torrent.eu.org:451/announce",
+    "udp://open.demonii.com:1337/announce",
+    "http://nyaa.tracker.wf:7777/announce",
+];
+
+fn encode_tr(url: &str) -> String {
+    let mut out = String::with_capacity(url.len() * 3);
+    for b in url.bytes() {
+        match b {
+            b'A'..=b'Z' | b'a'..=b'z' | b'0'..=b'9' | b'-' | b'_' | b'.' | b'~' => {
+                out.push(b as char)
+            }
+            _ => out.push_str(&format!("%{b:02X}")),
+        }
+    }
+    out
+}
+
+/// Append [`DEFAULT_TRACKERS`] to `magnet`, skipping any already present. A
+/// non-`magnet:` string is returned unchanged.
+pub fn with_default_trackers(mut magnet: String) -> String {
+    if !magnet.starts_with("magnet:") {
+        return magnet;
+    }
+    for tr in DEFAULT_TRACKERS {
+        let enc = encode_tr(tr);
+        if magnet.contains(&enc) || magnet.contains(tr) {
+            continue;
+        }
+        magnet.push_str("&tr=");
+        magnet.push_str(&enc);
+    }
+    magnet
+}
+
+fn decode_tr(s: &str) -> String {
+    let bytes = s.as_bytes();
+    let mut out: Vec<u8> = Vec::with_capacity(bytes.len());
+    let mut i = 0;
+    while i < bytes.len() {
+        if bytes[i] == b'%' && i + 2 < bytes.len() {
+            let hi = (bytes[i + 1] as char).to_digit(16);
+            let lo = (bytes[i + 2] as char).to_digit(16);
+            if let (Some(hi), Some(lo)) = (hi, lo) {
+                out.push((hi * 16 + lo) as u8);
+                i += 3;
+                continue;
+            }
+        }
+        out.push(bytes[i]);
+        i += 1;
+    }
+    String::from_utf8_lossy(&out).into_owned()
+}
+
+/// The decoded, de-duplicated `tr=` announce URLs of a magnet.
+pub fn parse_trackers(magnet: &str) -> Vec<String> {
+    if !magnet.starts_with("magnet:") {
+        return Vec::new();
+    }
+    let mut seen = std::collections::HashSet::new();
+    let mut out = Vec::new();
+    let query = magnet.splitn(2, '?').nth(1).unwrap_or("");
+    for pair in query.split('&') {
+        if let Some(val) = pair.strip_prefix("tr=") {
+            let url = decode_tr(val);
+            if !url.is_empty() && seen.insert(url.clone()) {
+                out.push(url);
+            }
+        }
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn build_magnet_includes_infohash() {
+        assert_eq!(
+            build_magnet(&[0xab; 20]),
+            "magnet:?xt=urn:btih:abababababababababababababababababababab"
+        );
+    }
+
+    #[test]
+    fn with_default_trackers_appends_to_trackerless_magnet() {
+        let bare = "magnet:?xt=urn:btih:abababababababababababababababababababab&dn=Foo".to_string();
+        let out = with_default_trackers(bare);
+        assert!(out.contains("&tr=udp%3A%2F%2Ftracker.opentrackr.org%3A1337%2Fannounce"));
+        assert!(out.contains("&tr=http%3A%2F%2Fnyaa.tracker.wf%3A7777%2Fannounce"));
+        assert!(out.contains("&dn=Foo"));
+    }
+
+    #[test]
+    fn with_default_trackers_dedups_existing() {
+        let pre = "magnet:?xt=urn:btih:abababababababababababababababababababab\
+                   &tr=udp%3A%2F%2Ftracker.opentrackr.org%3A1337%2Fannounce"
+            .to_string();
+        let out = with_default_trackers(pre);
+        assert_eq!(out.matches("tracker.opentrackr.org%3A1337").count(), 1);
+    }
+
+    #[test]
+    fn with_default_trackers_leaves_non_magnet_untouched() {
+        assert_eq!(with_default_trackers("not-a-magnet".to_string()), "not-a-magnet");
+    }
+
+    #[test]
+    fn parse_trackers_round_trips_the_defaults() {
+        let m = with_default_trackers(build_magnet(&[1; 20]));
+        assert_eq!(parse_trackers(&m), DEFAULT_TRACKERS.iter().map(|s| s.to_string()).collect::<Vec<_>>());
+        assert!(parse_trackers("not-a-magnet").is_empty());
+    }
+
+    #[test]
+    fn live_stats_serialise_with_the_legacy_field_names() {
+        let v = serde_json::to_value(LiveStats::default()).unwrap();
+        for k in [
+            "state", "down_bps", "up_bps", "uploaded_bytes", "downloaded_bytes",
+            "total_bytes", "ratio", "peers_live", "peers_seen",
+        ] {
+            assert!(v.get(k).is_some(), "missing {k}");
+        }
+    }
+}