feat: import Chinese-localized Buzz source snapshot

Signed-off-by: cls_宁波本机 <908705107@qq.com>

crates/buzz-media/Cargo.toml

@@ -0,0 +1,37 @@
+[package]
+name = "buzz-media"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+description = "Media storage, validation, and thumbnail generation for Buzz"
+
+[dependencies]
+buzz-core = { workspace = true }
+nostr = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+tokio = { workspace = true }
+tracing = { workspace = true }
+thiserror = { workspace = true }
+sha2 = { workspace = true }
+hex = { workspace = true }
+chrono = { workspace = true }
+ulid = "1"
+uuid = { workspace = true }
+axum = { workspace = true }
+s3 = { version = "0.37", package = "rust-s3", default-features = false, features = ["tokio-rustls-tls", "fail-on-err", "tags"] }
+infer = "0.19"
+image = { version = "0.25", default-features = false, features = ["jpeg", "png", "gif", "webp"] }
+blurhash = "0.2"
+imagesize = "0.14"
+bytes = "1"
+mp4 = "0.14"
+tempfile = "3"
+tokio-util = { version = "0.7", features = ["io"] }
+futures-util = "0.3"
+futures-core = "0.3"
+
+[dev-dependencies]
+tokio = { workspace = true, features = ["test-util"] }

crates/buzz-media/src/auth.rs

@@ -0,0 +1,552 @@
+//! Blossom kind:24242 auth verification (BUD-11 compliant).
+
+use crate::error::MediaError;
+
+/// Blossom kind:24242 verbs Buzz currently accepts.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum BlossomVerb {
+    Upload,
+    Get,
+}
+
+impl BlossomVerb {
+    fn as_str(self) -> &'static str {
+        match self {
+            Self::Upload => "upload",
+            Self::Get => "get",
+        }
+    }
+}
+
+/// Verify common kind:24242 Blossom auth event validity:
+///   1. Schnorr signature
+///   2. kind == 24242
+///   3. `t` tag matches `verb`
+///   4. `expiration` tag in the future
+///   5. `created_at` in the past (with 5s clock-skew tolerance)
+///   6. If `server` tags present, our domain must appear in at least one
+///
+/// Does NOT check verb-specific scope tags (`x` for upload, `x` OR `server`
+/// for get). Call this BEFORE trusting the event's pubkey for scope resolution.
+pub fn verify_blossom_auth_event_for_verb(
+    auth_event: &nostr::Event,
+    verb: BlossomVerb,
+    server_domain: Option<&str>,
+    max_age_secs: u64,
+) -> Result<(), MediaError> {
+    // 1. Verify Schnorr signature
+    auth_event
+        .verify()
+        .map_err(|_| MediaError::InvalidSignature)?;
+
+    // 2. Kind must be 24242
+    if auth_event.kind.as_u16() != 24242 {
+        return Err(MediaError::InvalidAuthKind);
+    }
+
+    // 2b. Content must be non-empty (BUD-11: "human readable string")
+    if auth_event.content.trim().is_empty() {
+        return Err(MediaError::InvalidAuthEvent);
+    }
+
+    let mut found_t = false;
+    let mut found_exp = false;
+    let mut server_tags: Vec<&str> = Vec::new();
+    let mut exp_value: u64 = 0;
+
+    for tag in auth_event.tags.iter() {
+        let kind = tag.kind().to_string();
+        match kind.as_str() {
+            "t" => {
+                if let Some(v) = tag.content() {
+                    if v != verb.as_str() {
+                        return Err(MediaError::InvalidAuthVerb);
+                    }
+                    found_t = true;
+                }
+            }
+            "expiration" => {
+                if let Some(v) = tag.content() {
+                    exp_value = v.parse().unwrap_or(0);
+                    found_exp = true;
+                }
+            }
+            "server" => {
+                if let Some(v) = tag.content() {
+                    server_tags.push(v);
+                }
+            }
+            _ => {}
+        }
+    }
+
+    // 3. t tag required
+    if !found_t {
+        return Err(MediaError::MissingTag("t"));
+    }
+
+    // 4. Expiration must exist and be in the future
+    if !found_exp {
+        return Err(MediaError::MissingTag("expiration"));
+    }
+    let now = nostr::Timestamp::now().as_secs();
+    if exp_value <= now {
+        return Err(MediaError::TokenExpired);
+    }
+
+    // 5. created_at must be recent: not in the future (5s tolerance) and not
+    //    older than 10 minutes. This bounds the replay window — even if the
+    //    expiration tag allows a longer lifetime, the token must have been
+    //    freshly minted.
+    let created = auth_event.created_at.as_secs();
+    if created > now + 5 {
+        return Err(MediaError::TimestampOutOfWindow);
+    }
+    if now > created + max_age_secs {
+        return Err(MediaError::TimestampOutOfWindow);
+    }
+
+    // 6. Server tag enforcement (BUD-11 §5): if server tags present, our host must appear.
+    //
+    // `server_domain` is the host this request was bound to — the per-request
+    // tenant host (`TenantContext::host()`), NOT a single process-global domain.
+    // A relay process serves many tenant hosts; validating against one global
+    // host would 401 every non-primary tenant's server-tagged client (the stock
+    // CLI always tags its configured relay host). Comparison is done under the
+    // shared [`normalize_host`] rule so a tag and the bound host agree by
+    // construction across case, trailing dot, default ports, and an optional
+    // URL scheme/path — exactly as every other host seam resolves tenants.
+    //
+    // Fail closed: if the bound host is unknown, reject tokens that carry server
+    // tags rather than silently accepting them.
+    if !server_tags.is_empty() {
+        match server_domain {
+            Some(domain) => {
+                let want = normalize_server_host(domain);
+                let matches = server_tags
+                    .iter()
+                    .any(|tag| normalize_server_host(tag) == want);
+                if !matches {
+                    return Err(MediaError::ServerMismatch);
+                }
+            }
+            None => {
+                // Server tags present but we don't know our own host — reject.
+                return Err(MediaError::ServerMismatch);
+            }
+        }
+    }
+
+    Ok(())
+}
+
+/// Verify common upload auth event validity.
+///
+/// Kept as the upload-shaped public wrapper for existing callers; new verb-aware
+/// code should prefer [`verify_blossom_auth_event_for_verb`].
+pub fn verify_blossom_auth_event(
+    auth_event: &nostr::Event,
+    server_domain: Option<&str>,
+    max_age_secs: u64,
+) -> Result<(), MediaError> {
+    verify_blossom_auth_event_for_verb(auth_event, BlossomVerb::Upload, server_domain, max_age_secs)
+}
+
+/// Normalize a Blossom `server` tag value (or a bound tenant host) into the
+/// canonical host form used as the community lookup key.
+///
+/// A `server` tag may be a bare authority (`relay.example:3100`, what the stock
+/// CLI emits) or a full URL (`https://relay.example/`). We strip an optional
+/// scheme and path down to the authority, then apply the one shared
+/// [`buzz_core::tenant::normalize_host`] rule so the comparison agrees with how
+/// the WS/HTTP/git doors resolve tenants.
+fn normalize_server_host(value: &str) -> String {
+    let authority = match value.split_once("://") {
+        Some((_scheme, rest)) => rest.split('/').next().unwrap_or(rest),
+        None => value.split('/').next().unwrap_or(value),
+    };
+    buzz_core::tenant::normalize_host(authority)
+}
+
+/// Verify a kind:24242 Blossom upload auth event, including the x tag hash check.
+///
+/// Calls [`verify_blossom_auth_event`] first, then verifies that at least one
+/// `x` tag matches `sha256` (BUD-11 §6: "at least one x tag matches").
+pub fn verify_blossom_upload_auth(
+    auth_event: &nostr::Event,
+    sha256: &str,
+    server_domain: Option<&str>,
+    max_age_secs: u64,
+) -> Result<(), MediaError> {
+    verify_blossom_auth_event_for_verb(
+        auth_event,
+        BlossomVerb::Upload,
+        server_domain,
+        max_age_secs,
+    )?;
+
+    // At least one x tag must match the body sha256 (BUD-11 §6)
+    let has_matching_x = auth_event
+        .tags
+        .iter()
+        .any(|tag| tag.kind().to_string() == "x" && (tag.content() == Some(sha256)));
+
+    if !has_matching_x {
+        return Err(MediaError::HashMismatch);
+    }
+
+    Ok(())
+}
+
+/// Verify a kind:24242 Blossom get auth event for one requested blob.
+///
+/// BUD-01 permits either blob-scoped authorization (`x` tag matches `sha256`)
+/// or server-scoped authorization (`server` tag matches this relay host). The
+/// latter intentionally grants reads for all blobs on the host until expiration;
+/// callers must still apply relay membership after this verifier returns.
+pub fn verify_blossom_get_auth(
+    auth_event: &nostr::Event,
+    sha256: &str,
+    server_domain: Option<&str>,
+    max_age_secs: u64,
+) -> Result<(), MediaError> {
+    verify_blossom_auth_event_for_verb(auth_event, BlossomVerb::Get, server_domain, max_age_secs)?;
+
+    let has_matching_x = auth_event
+        .tags
+        .iter()
+        .any(|tag| tag.kind().to_string() == "x" && (tag.content() == Some(sha256)));
+
+    let has_matching_server = match server_domain {
+        Some(domain) => {
+            let want = normalize_server_host(domain);
+            auth_event.tags.iter().any(|tag| {
+                tag.kind().to_string() == "server"
+                    && tag
+                        .content()
+                        .map(|value| normalize_server_host(value) == want)
+                        .unwrap_or(false)
+            })
+        }
+        None => false,
+    };
+
+    if !has_matching_x && !has_matching_server {
+        return Err(MediaError::InsufficientScope);
+    }
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use nostr::{EventBuilder, Keys, Kind, Tag, Timestamp};
+
+    fn build_valid_auth(keys: &Keys, sha256: &str) -> nostr::Event {
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let tags = vec![
+            Tag::parse(["t", "upload"]).unwrap(),
+            Tag::parse(["x", sha256]).unwrap(),
+            Tag::parse(["expiration", &exp_str]).unwrap(),
+        ];
+        EventBuilder::new(Kind::from(24242), "Upload buzz-media")
+            .tags(tags)
+            .sign_with_keys(keys)
+            .unwrap()
+    }
+
+    #[test]
+    fn test_verify_valid() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let event = build_valid_auth(&keys, &sha256);
+        assert!(verify_blossom_upload_auth(&event, &sha256, None, 600).is_ok());
+    }
+
+    #[test]
+    fn test_verify_auth_event_valid() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let event = build_valid_auth(&keys, &sha256);
+        assert!(verify_blossom_auth_event(&event, None, 600).is_ok());
+    }
+
+    fn build_get_auth(keys: &Keys, tags: Vec<Tag>) -> nostr::Event {
+        EventBuilder::new(Kind::from(24242), "Get buzz-media")
+            .tags(tags)
+            .sign_with_keys(keys)
+            .unwrap()
+    }
+
+    #[test]
+    fn test_verify_get_accepts_matching_x_without_server_tag() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let event = build_get_auth(
+            &keys,
+            vec![
+                Tag::parse(["t", "get"]).unwrap(),
+                Tag::parse(["x", &sha256]).unwrap(),
+                Tag::parse(["expiration", &exp_str]).unwrap(),
+            ],
+        );
+
+        assert!(verify_blossom_get_auth(&event, &sha256, Some("relay.example"), 600).is_ok());
+    }
+
+    #[test]
+    fn test_verify_get_accepts_matching_server_without_x_tag() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let event = build_get_auth(
+            &keys,
+            vec![
+                Tag::parse(["t", "get"]).unwrap(),
+                Tag::parse(["server", "https://Relay.Example./media/ignored"]).unwrap(),
+                Tag::parse(["expiration", &exp_str]).unwrap(),
+            ],
+        );
+
+        assert!(verify_blossom_get_auth(&event, &sha256, Some("relay.example"), 600).is_ok());
+    }
+
+    #[test]
+    fn test_verify_get_rejects_upload_verb() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let event = build_valid_auth(&keys, &sha256);
+
+        assert!(matches!(
+            verify_blossom_get_auth(&event, &sha256, Some("relay.example"), 600),
+            Err(MediaError::InvalidAuthVerb)
+        ));
+    }
+
+    #[test]
+    fn test_verify_get_requires_x_or_server_scope() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let other_hash = "b".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let event = build_get_auth(
+            &keys,
+            vec![
+                Tag::parse(["t", "get"]).unwrap(),
+                Tag::parse(["x", &other_hash]).unwrap(),
+                Tag::parse(["expiration", &exp_str]).unwrap(),
+            ],
+        );
+
+        assert!(matches!(
+            verify_blossom_get_auth(&event, &sha256, Some("relay.example"), 600),
+            Err(MediaError::InsufficientScope)
+        ));
+    }
+
+    #[test]
+    fn test_verify_get_rejects_wrong_server_scope() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let event = build_get_auth(
+            &keys,
+            vec![
+                Tag::parse(["t", "get"]).unwrap(),
+                Tag::parse(["server", "other.example"]).unwrap(),
+                Tag::parse(["expiration", &exp_str]).unwrap(),
+            ],
+        );
+
+        assert!(matches!(
+            verify_blossom_get_auth(&event, &sha256, Some("relay.example"), 600),
+            Err(MediaError::ServerMismatch)
+        ));
+    }
+
+    #[test]
+    fn test_verify_hash_mismatch() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let event = build_valid_auth(&keys, &sha256);
+        let wrong_hash = "b".repeat(64);
+        assert!(matches!(
+            verify_blossom_upload_auth(&event, &wrong_hash, None, 600),
+            Err(MediaError::HashMismatch)
+        ));
+    }
+
+    #[test]
+    fn test_verify_wrong_kind() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let tags = vec![
+            Tag::parse(["t", "upload"]).unwrap(),
+            Tag::parse(["x", &sha256]).unwrap(),
+            Tag::parse(["expiration", &exp_str]).unwrap(),
+        ];
+        let event = EventBuilder::new(Kind::from(27235), "wrong kind")
+            .tags(tags)
+            .sign_with_keys(&keys)
+            .unwrap();
+        assert!(matches!(
+            verify_blossom_upload_auth(&event, &sha256, None, 600),
+            Err(MediaError::InvalidAuthKind)
+        ));
+    }
+
+    #[test]
+    fn test_verify_multi_x_tags() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let other_hash = "b".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let tags = vec![
+            Tag::parse(["t", "upload"]).unwrap(),
+            Tag::parse(["x", &other_hash]).unwrap(),
+            Tag::parse(["x", &sha256]).unwrap(),
+            Tag::parse(["expiration", &exp_str]).unwrap(),
+        ];
+        let event = EventBuilder::new(Kind::from(24242), "Upload multi-x")
+            .tags(tags)
+            .sign_with_keys(&keys)
+            .unwrap();
+        // Should pass because at least one x tag matches
+        assert!(verify_blossom_upload_auth(&event, &sha256, None, 600).is_ok());
+    }
+
+    #[test]
+    fn test_server_tag_enforcement() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let tags = vec![
+            Tag::parse(["t", "upload"]).unwrap(),
+            Tag::parse(["x", &sha256]).unwrap(),
+            Tag::parse(["expiration", &exp_str]).unwrap(),
+            Tag::parse(["server", "other.example.com"]).unwrap(),
+        ];
+        let event = EventBuilder::new(Kind::from(24242), "Upload scoped")
+            .tags(tags)
+            .sign_with_keys(&keys)
+            .unwrap();
+        // Should fail — server tag present but doesn't match our domain
+        assert!(matches!(
+            verify_blossom_upload_auth(&event, &sha256, Some("buzz.example.com"), 600),
+            Err(MediaError::ServerMismatch)
+        ));
+        // Should pass when our domain matches
+        assert!(
+            verify_blossom_upload_auth(&event, &sha256, Some("other.example.com"), 600).is_ok()
+        );
+        // Should fail when server_domain is None — fail closed
+        assert!(matches!(
+            verify_blossom_upload_auth(&event, &sha256, None, 600),
+            Err(MediaError::ServerMismatch)
+        ));
+    }
+
+    #[test]
+    fn test_no_server_tags_always_passes() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let event = build_valid_auth(&keys, &sha256);
+        // No server tags → passes regardless of our domain
+        assert!(verify_blossom_upload_auth(&event, &sha256, Some("any.domain.com"), 600).is_ok());
+    }
+
+    /// A `server` tag is matched against the *bound tenant host* under the
+    /// shared `normalize_host` rule, so equivalent host spellings agree — the
+    /// stock CLI's bare `host:port`, an explicit default port, a trailing dot,
+    /// mixed case, and a full URL all match the same bound host. This is the
+    /// regression guard for the multi-tenant media blocker: a non-primary
+    /// tenant must accept its own server-tagged client.
+    #[test]
+    fn test_server_tag_normalized_against_bound_host() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let build = |server: &str| {
+            let tags = vec![
+                Tag::parse(["t", "upload"]).unwrap(),
+                Tag::parse(["x", &sha256]).unwrap(),
+                Tag::parse(["expiration", &exp_str]).unwrap(),
+                Tag::parse(["server", server]).unwrap(),
+            ];
+            EventBuilder::new(Kind::from(24242), "Upload scoped")
+                .tags(tags)
+                .sign_with_keys(&keys)
+                .unwrap()
+        };
+
+        // Non-primary tenant host with explicit non-default port (the live
+        // repro: tenant B on 127.0.0.1:3100). Stock CLI tags `host:port`.
+        assert!(verify_blossom_upload_auth(
+            &build("127.0.0.1:3100"),
+            &sha256,
+            Some("127.0.0.1:3100"),
+            600
+        )
+        .is_ok());
+
+        // Equivalence under normalize_host: explicit default port, trailing
+        // dot, mixed case, and a full URL all collapse to the bound host.
+        for tag in [
+            "Relay.Example:443",
+            "relay.example.",
+            "RELAY.EXAMPLE",
+            "https://relay.example/",
+        ] {
+            assert!(
+                verify_blossom_upload_auth(&build(tag), &sha256, Some("relay.example"), 600)
+                    .is_ok(),
+                "server tag {tag:?} should match bound host relay.example"
+            );
+        }
+
+        // A different tenant host still fails closed.
+        assert!(matches!(
+            verify_blossom_upload_auth(
+                &build("127.0.0.1:3100"),
+                &sha256,
+                Some("127.0.0.1:3200"),
+                600
+            ),
+            Err(MediaError::ServerMismatch)
+        ));
+    }
+
+    #[test]
+    fn test_empty_content_rejected() {
+        let keys = Keys::generate();
+        let sha256 = "a".repeat(64);
+        let now = Timestamp::now().as_secs();
+        let exp_str = (now + 300).to_string();
+        let tags = vec![
+            Tag::parse(["t", "upload"]).unwrap(),
+            Tag::parse(["x", &sha256]).unwrap(),
+            Tag::parse(["expiration", &exp_str]).unwrap(),
+        ];
+        // Empty content — BUD-11 requires a human-readable string
+        let event = EventBuilder::new(Kind::from(24242), "")
+            .tags(tags)
+            .sign_with_keys(&keys)
+            .unwrap();
+        assert!(matches!(
+            verify_blossom_auth_event(&event, None, 600),
+            Err(MediaError::InvalidAuthEvent)
+        ));
+    }
+}

crates/buzz-media/src/bucket_index.rs

@@ -0,0 +1,755 @@
+//! Bucket key taxonomy classifier and pure aggregation fold for the S3-truth
+//! storage sweep.
+//!
+//! This module has **zero S3 I/O** — [`classify_key`] and [`BucketAggregate`]
+//! operate on plain `(key, size)` pairs, and [`fold_bucket_listing`] takes a
+//! caller-supplied page-fetching closure so the pagination/cap logic is
+//! testable against synthetic listings. The relay wires a real
+//! [`crate::storage::MediaStorage::list_page`] closure at the call site (see
+//! `buzz-relay`'s storage sweep task).
+//!
+//! Five key classes (thumb matched first, everything unrecognized is
+//! `Unknown` — never silently folded into another class):
+//!
+//! | Class | Shape |
+//! |---|---|
+//! | thumb | `{sha256}.thumb.jpg` |
+//! | blob | `{sha256}.{ext}` (ext: 1-8 mixed-case alphanumeric) |
+//! | sidecar | `_meta/{community-uuid}/{sha256}.json` |
+//! | auxiliary | `_uploads/{community-uuid}/{sha256}/{ulid}.json` |
+//! | unknown | everything else |
+
+use std::collections::HashMap;
+use std::future::Future;
+
+use uuid::Uuid;
+
+use crate::error::MediaError;
+
+/// The classification of one bucket key. `Unknown` is the deliberate
+/// catch-all — a malformed variant of a known prefix (e.g. a truncated
+/// `_uploads/` key) falls to `Unknown` rather than being coerced into
+/// `Auxiliary`, so visibility gauges stay loud instead of silently wrong.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum KeyClass {
+    /// `{sha256}.thumb.jpg` — attributed to the blob's sha.
+    Thumb { sha256: String },
+    /// `{sha256}.{ext}` — physical bytes, logical join key.
+    Blob { sha256: String, ext: String },
+    /// `_meta/{community}/{sha256}.json` — the (community, sha) binding.
+    Sidecar { community: Uuid, sha256: String },
+    /// `_uploads/{community}/{sha256}/{event_id}.json` — fleet physical only.
+    Auxiliary {
+        community: Uuid,
+        sha256: String,
+        event_id: String,
+    },
+    /// Anything that doesn't match one of the four strict shapes above.
+    Unknown,
+}
+
+/// Classify one bucket key. Matches `thumb` first (its suffix is a superset
+/// shape of the blob pattern's segment count), then blob, sidecar,
+/// auxiliary, and finally unknown. See module docs for the exact shapes.
+pub fn classify_key(key: &str) -> KeyClass {
+    if let Some(sha256) = parse_thumb_key(key) {
+        return KeyClass::Thumb { sha256 };
+    }
+    if let Some((sha256, ext)) = parse_blob_key(key) {
+        return KeyClass::Blob { sha256, ext };
+    }
+    if let Some((community, sha256)) = parse_sidecar_key(key) {
+        return KeyClass::Sidecar { community, sha256 };
+    }
+    if let Some((community, sha256, event_id)) = parse_auxiliary_key(key) {
+        return KeyClass::Auxiliary {
+            community,
+            sha256,
+            event_id,
+        };
+    }
+    KeyClass::Unknown
+}
+
+/// A 64-char lowercase-hex SHA-256 digest, strictly.
+fn is_sha256(s: &str) -> bool {
+    s.len() == 64
+        && s.bytes()
+            .all(|b| b.is_ascii_digit() || (b'a'..=b'f').contains(&b))
+}
+
+/// Blob extension charset: 1-8 mixed-case alphanumeric chars (F4-bis — infer
+/// 0.19 emits uppercase `Z` for `application/x-compress`, which is
+/// legitimate writer output that must classify as a blob, not unknown).
+fn is_blob_ext(s: &str) -> bool {
+    !s.is_empty() && s.len() <= 8 && s.bytes().all(|b| b.is_ascii_alphanumeric())
+}
+
+/// Strict Crockford-base32 ULID charset check, uppercase only (matches the
+/// ulid crate's `Display` output — see `upload_record.rs`'s writer). Not
+/// `ulid::Ulid::from_string`, which is deliberately case-insensitive on
+/// decode and would accept lowercase variants the writer never produces —
+/// looser than the plan's anchored regex.
+fn is_ulid_charset(s: &str) -> bool {
+    s.len() == 26
+        && s.bytes().all(|b| {
+            b.is_ascii_digit()
+                || (b'A'..=b'H').contains(&b)
+                || b == b'J'
+                || b == b'K'
+                || b == b'M'
+                || b == b'N'
+                || (b'P'..=b'T').contains(&b)
+                || (b'V'..=b'Z').contains(&b)
+        })
+}
+
+/// Strict canonical UUID: exactly 36 chars, lowercase hex + hyphens at
+/// positions 8/13/18/23. Rejects the braced/urn/no-hyphen forms
+/// `Uuid::parse_str` alone would otherwise accept — every UUID this server
+/// writes into a key is `Display`-formatted canonically lowercase, so
+/// anything else is not a UUID we wrote and must not silently parse as one.
+fn parse_canonical_uuid(s: &str) -> Option<Uuid> {
+    if s.len() != 36 {
+        return None;
+    }
+    for (i, b) in s.bytes().enumerate() {
+        let ok = match i {
+            8 | 13 | 18 | 23 => b == b'-',
+            _ => b.is_ascii_digit() || (b'a'..=b'f').contains(&b),
+        };
+        if !ok {
+            return None;
+        }
+    }
+    Uuid::parse_str(s).ok()
+}
+
+/// `{sha256}.thumb.jpg`
+fn parse_thumb_key(key: &str) -> Option<String> {
+    let mut parts = key.split('.');
+    let sha256 = parts.next()?;
+    let thumb = parts.next()?;
+    let jpg = parts.next()?;
+    if parts.next().is_some() || thumb != "thumb" || jpg != "jpg" || !is_sha256(sha256) {
+        return None;
+    }
+    Some(sha256.to_string())
+}
+
+/// `{sha256}.{ext}` — exactly two dot-separated segments.
+fn parse_blob_key(key: &str) -> Option<(String, String)> {
+    let mut parts = key.split('.');
+    let sha256 = parts.next()?;
+    let ext = parts.next()?;
+    if parts.next().is_some() || !is_sha256(sha256) || !is_blob_ext(ext) {
+        return None;
+    }
+    Some((sha256.to_string(), ext.to_string()))
+}
+
+/// `_meta/{community}/{sha256}.json`
+fn parse_sidecar_key(key: &str) -> Option<(Uuid, String)> {
+    let mut segments = key.split('/');
+    if segments.next()? != "_meta" {
+        return None;
+    }
+    let community = parse_canonical_uuid(segments.next()?)?;
+    let last = segments.next()?;
+    if segments.next().is_some() {
+        return None;
+    }
+    let mut last_parts = last.split('.');
+    let sha256 = last_parts.next()?;
+    let json = last_parts.next()?;
+    if last_parts.next().is_some() || json != "json" || !is_sha256(sha256) {
+        return None;
+    }
+    Some((community, sha256.to_string()))
+}
+
+/// `_uploads/{community}/{sha256}/{event_id}.json`, `event_id` a ULID.
+fn parse_auxiliary_key(key: &str) -> Option<(Uuid, String, String)> {
+    let mut segments = key.split('/');
+    if segments.next()? != "_uploads" {
+        return None;
+    }
+    let community = parse_canonical_uuid(segments.next()?)?;
+    let sha256 = segments.next()?;
+    if !is_sha256(sha256) {
+        return None;
+    }
+    let last = segments.next()?;
+    if segments.next().is_some() {
+        return None;
+    }
+    let mut last_parts = last.split('.');
+    let event_id = last_parts.next()?;
+    let json = last_parts.next()?;
+    if last_parts.next().is_some() || json != "json" || !is_ulid_charset(event_id) {
+        return None;
+    }
+    Some((community, sha256.to_string(), event_id.to_string()))
+}
+
+/// Per-community logical storage: bytes and object count of bound shas.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct CommunityStorage {
+    pub bytes: u64,
+    pub objects: u64,
+}
+
+/// The full computed sweep result: fleet physical/logical totals,
+/// per-community logical breakdown, and anomaly/visibility gauges. Pure
+/// data — no I/O, cheap to clone into a cached snapshot.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct BucketSnapshot {
+    /// Every listed object, every class (kind=physical).
+    pub physical_bytes: u64,
+    pub physical_objects: u64,
+    /// Sum of per-community logical bytes/objects (kind=logical). Bills a
+    /// blob bound to N communities N times — intentional (D-EXT/D-COUNT).
+    pub logical_bytes: u64,
+    pub logical_objects: u64,
+    pub per_community: HashMap<Uuid, CommunityStorage>,
+    /// Blob shas with zero sidecar binding in any community.
+    pub orphan_blob_bytes: u64,
+    pub orphan_blob_count: u64,
+    /// Sidecar bindings whose sha has no blob bytes at all.
+    pub orphan_sidecar_count: u64,
+    /// Shas with more than one blob variant (anomaly — see plan D-EXT).
+    pub multi_variant_shas: u64,
+    /// Total bytes of ALL blob variants belonging to anomalous shas.
+    pub multi_variant_bytes: u64,
+    pub unknown_key_bytes: u64,
+    pub unknown_key_objects: u64,
+}
+
+/// Pure, incremental fold over classified bucket keys. Never retains a full
+/// object listing — only per-sha/per-binding running totals, bounded by the
+/// number of distinct shas and sidecar bindings actually present.
+#[derive(Debug, Default)]
+pub struct BucketAggregate {
+    /// sha -> bytes of every blob variant seen for that sha (D-EXT: multiple
+    /// entries is the multi-variant anomaly).
+    blob_variant_bytes: HashMap<String, Vec<u64>>,
+    /// sha -> thumb bytes. At most one thumb key per sha, so a plain insert
+    /// is correct (no accumulation needed).
+    thumb_bytes: HashMap<String, u64>,
+    /// (community, sha) -> sidecar object's own byte size (informational;
+    /// not part of logical bytes).
+    sidecar_bindings: HashMap<(Uuid, String), u64>,
+    physical_bytes: u64,
+    physical_objects: u64,
+    unknown_bytes: u64,
+    unknown_objects: u64,
+}
+
+impl BucketAggregate {
+    /// Fold one classified `(key, size)` pair into the running aggregate.
+    pub fn fold(&mut self, key: &str, size: u64) {
+        self.physical_objects += 1;
+        self.physical_bytes += size;
+        match classify_key(key) {
+            KeyClass::Thumb { sha256 } => {
+                self.thumb_bytes.insert(sha256, size);
+            }
+            KeyClass::Blob { sha256, .. } => {
+                self.blob_variant_bytes
+                    .entry(sha256)
+                    .or_default()
+                    .push(size);
+            }
+            KeyClass::Sidecar { community, sha256 } => {
+                self.sidecar_bindings.insert((community, sha256), size);
+            }
+            KeyClass::Auxiliary { .. } => {
+                // Fleet physical only — never enters logical/orphan math (F4).
+            }
+            KeyClass::Unknown => {
+                self.unknown_objects += 1;
+                self.unknown_bytes += size;
+            }
+        }
+    }
+
+    /// Compute the final snapshot from everything folded so far.
+    pub fn finish(self) -> BucketSnapshot {
+        let bound_shas: std::collections::HashSet<&str> = self
+            .sidecar_bindings
+            .keys()
+            .map(|(_, sha256)| sha256.as_str())
+            .collect();
+
+        let mut multi_variant_shas = 0u64;
+        let mut multi_variant_bytes = 0u64;
+        let mut orphan_blob_count = 0u64;
+        let mut orphan_blob_bytes = 0u64;
+        for (sha256, variants) in &self.blob_variant_bytes {
+            let variant_bytes: u64 = variants.iter().sum();
+            if variants.len() > 1 {
+                multi_variant_shas += 1;
+                multi_variant_bytes += variant_bytes;
+            }
+            if !bound_shas.contains(sha256.as_str()) {
+                orphan_blob_count += 1;
+                orphan_blob_bytes += variant_bytes;
+            }
+        }
+
+        let orphan_sidecar_count = self
+            .sidecar_bindings
+            .keys()
+            .filter(|(_, sha256)| !self.blob_variant_bytes.contains_key(sha256))
+            .count() as u64;
+
+        let mut per_community: HashMap<Uuid, CommunityStorage> = HashMap::new();
+        for (community, sha256) in self.sidecar_bindings.keys() {
+            let blob_bytes: u64 = self
+                .blob_variant_bytes
+                .get(sha256)
+                .map(|v| v.iter().sum())
+                .unwrap_or(0);
+            let thumb_bytes = self.thumb_bytes.get(sha256).copied().unwrap_or(0);
+            let entry = per_community.entry(*community).or_default();
+            entry.bytes += blob_bytes + thumb_bytes;
+            entry.objects += 1;
+        }
+        let logical_bytes = per_community.values().map(|c| c.bytes).sum();
+        let logical_objects = per_community.values().map(|c| c.objects).sum();
+
+        BucketSnapshot {
+            physical_bytes: self.physical_bytes,
+            physical_objects: self.physical_objects,
+            logical_bytes,
+            logical_objects,
+            per_community,
+            orphan_blob_bytes,
+            orphan_blob_count,
+            orphan_sidecar_count,
+            multi_variant_shas,
+            multi_variant_bytes,
+            unknown_key_bytes: self.unknown_bytes,
+            unknown_key_objects: self.unknown_objects,
+        }
+    }
+}
+
+/// Failure modes for the paginated listing fold. All variants mean "failed
+/// sweep, keep the old snapshot" to the caller — never a partial one.
+#[derive(Debug, thiserror::Error)]
+pub enum SweepError {
+    /// Cumulative listed-object count exceeded `cap` mid-listing.
+    #[error("object cap exceeded: {seen} listed objects > cap {cap}")]
+    CapExceeded { seen: u64, cap: u64 },
+    /// The page source (S3, or a test double) failed.
+    #[error("storage error during listing: {0}")]
+    Storage(#[from] MediaError),
+    /// The whole sweep (this fold plus any caller-side wrapping) exceeded its
+    /// deadline. Constructed by the relay's sweep task, which wraps this
+    /// entire function in `tokio::time::timeout` — kept here (not a
+    /// relay-local error type) so `SweepError` stays the single failure
+    /// currency the whole sweep pipeline reasons about.
+    #[error("sweep timed out after {0:?}")]
+    Timeout(std::time::Duration),
+    /// A listing page reported `is_truncated=true` but supplied no
+    /// continuation token — a malformed S3 response that cannot be resumed.
+    #[error("truncated listing page with no continuation token")]
+    MalformedPage,
+}
+
+/// One page of a bucket listing, decoupled from any S3 crate type so the
+/// fold below can be driven by a synthetic closure in tests.
+#[derive(Debug, Clone, Default)]
+pub struct Page {
+    pub objects: Vec<(String, u64)>,
+    pub next_continuation_token: Option<String>,
+    pub is_truncated: bool,
+}
+
+/// Fold an entire paginated bucket listing, checking the object cap BEFORE
+/// folding each page and never retaining the full listing — only the
+/// bounded per-sha/per-binding aggregate state.
+///
+/// `fetch_page` is called with `None` for the first page and the previous
+/// page's continuation token thereafter; production callers close over a
+/// [`crate::storage::MediaStorage`], tests close over canned [`Page`]s.
+pub async fn fold_bucket_listing<F, Fut>(
+    cap: u64,
+    mut fetch_page: F,
+) -> Result<BucketSnapshot, SweepError>
+where
+    F: FnMut(Option<String>) -> Fut,
+    Fut: Future<Output = Result<Page, MediaError>>,
+{
+    let mut aggregate = BucketAggregate::default();
+    let mut continuation_token = None;
+    let mut seen: u64 = 0;
+
+    loop {
+        let page = fetch_page(continuation_token.take()).await?;
+
+        seen += page.objects.len() as u64;
+        if seen > cap {
+            return Err(SweepError::CapExceeded { seen, cap });
+        }
+
+        for (key, size) in &page.objects {
+            aggregate.fold(key, *size);
+        }
+
+        if !page.is_truncated {
+            break;
+        }
+        match page.next_continuation_token {
+            Some(token) => continuation_token = Some(token),
+            None => return Err(SweepError::MalformedPage),
+        }
+    }
+
+    Ok(aggregate.finish())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn sha(byte: u8) -> String {
+        hex::encode([byte; 32])
+    }
+
+    fn community(n: u128) -> Uuid {
+        Uuid::from_u128(n)
+    }
+
+    // --- classify_key ---
+
+    #[test]
+    fn classifies_thumb_key() {
+        let s = sha(0xaa);
+        assert_eq!(
+            classify_key(&format!("{s}.thumb.jpg")),
+            KeyClass::Thumb { sha256: s }
+        );
+    }
+
+    #[test]
+    fn classifies_blob_key_lowercase_ext() {
+        let s = sha(0xbb);
+        assert_eq!(
+            classify_key(&format!("{s}.png")),
+            KeyClass::Blob {
+                sha256: s,
+                ext: "png".to_string()
+            }
+        );
+    }
+
+    /// F4-bis: infer 0.19 emits uppercase `Z` for `application/x-compress`,
+    /// legitimate writer output — must classify as blob, not unknown.
+    #[test]
+    fn classifies_blob_key_uppercase_z_extension() {
+        let s = sha(0xcc);
+        assert_eq!(
+            classify_key(&format!("{s}.Z")),
+            KeyClass::Blob {
+                sha256: s,
+                ext: "Z".to_string()
+            }
+        );
+    }
+
+    #[test]
+    fn classifies_sidecar_key() {
+        let s = sha(0xdd);
+        let c = community(1);
+        assert_eq!(
+            classify_key(&format!("_meta/{c}/{s}.json")),
+            KeyClass::Sidecar {
+                community: c,
+                sha256: s
+            }
+        );
+    }
+
+    #[test]
+    fn classifies_auxiliary_key() {
+        let s = sha(0xee);
+        let c = community(2);
+        let ulid = "01ARZ3NDEKTSV4RRFFQ69G5FAV";
+        assert_eq!(
+            classify_key(&format!("_uploads/{c}/{s}/{ulid}.json")),
+            KeyClass::Auxiliary {
+                community: c,
+                sha256: s,
+                event_id: ulid.to_string(),
+            }
+        );
+    }
+
+    #[test]
+    fn malformed_uploads_key_is_unknown_not_auxiliary() {
+        let s = sha(0xff);
+        let c = community(3);
+        // Lowercase ULID: the writer never emits this — must not silently
+        // pass as auxiliary; visibility (unknown) beats a wrong guess.
+        assert_eq!(
+            classify_key(&format!("_uploads/{c}/{s}/01arz3ndektsv4rrffq69g5fav.json")),
+            KeyClass::Unknown
+        );
+        // Wrong-length event id.
+        assert_eq!(
+            classify_key(&format!("_uploads/{c}/{s}/TOOSHORT.json")),
+            KeyClass::Unknown
+        );
+    }
+
+    #[test]
+    fn malformed_sidecar_non_uuid_community_is_unknown() {
+        let s = sha(0x11);
+        assert_eq!(
+            classify_key(&format!("_meta/not-a-uuid/{s}.json")),
+            KeyClass::Unknown
+        );
+    }
+
+    #[test]
+    fn malformed_keys_fall_to_unknown() {
+        let s = sha(0xab);
+        // Uppercase hex in the sha segment.
+        assert_eq!(
+            classify_key(&format!("{}.png", s.to_uppercase())),
+            KeyClass::Unknown
+        );
+        // Wrong sha length.
+        assert_eq!(classify_key("abc123.png"), KeyClass::Unknown);
+        // Extra segment.
+        assert_eq!(classify_key(&format!("{s}.png.bak")), KeyClass::Unknown);
+        // Extension too long (9 chars).
+        assert_eq!(classify_key(&format!("{s}.123456789")), KeyClass::Unknown);
+        // No extension at all.
+        assert_eq!(classify_key(&s), KeyClass::Unknown);
+        // Totally unrelated key.
+        assert_eq!(classify_key("README.md"), KeyClass::Unknown);
+    }
+
+    // --- BucketAggregate / finish ---
+
+    #[test]
+    fn empty_listing_yields_zero_snapshot() {
+        let snapshot = BucketAggregate::default().finish();
+        assert_eq!(snapshot, BucketSnapshot::default());
+    }
+
+    #[test]
+    fn multi_variant_sha_is_anomalous_and_bills_the_sum() {
+        let s = sha(0x33);
+        let c = community(4);
+        let mut agg = BucketAggregate::default();
+        agg.fold(&format!("{s}.jpg"), 100);
+        agg.fold(&format!("{s}.png"), 200); // same sha, second variant
+        agg.fold(&format!("_meta/{c}/{s}.json"), 10);
+
+        let snap = agg.finish();
+        assert_eq!(snap.multi_variant_shas, 1);
+        assert_eq!(snap.multi_variant_bytes, 300);
+        assert_eq!(snap.orphan_blob_count, 0);
+        // Logical bytes bill the sum of both variants (D-EXT).
+        assert_eq!(snap.per_community[&c].bytes, 300);
+        assert_eq!(snap.per_community[&c].objects, 1);
+    }
+
+    #[test]
+    fn orphan_blob_has_no_sidecar_binding() {
+        let s = sha(0x44);
+        let mut agg = BucketAggregate::default();
+        agg.fold(&format!("{s}.jpg"), 500);
+
+        let snap = agg.finish();
+        assert_eq!(snap.orphan_blob_count, 1);
+        assert_eq!(snap.orphan_blob_bytes, 500);
+        assert!(snap.per_community.is_empty());
+        assert_eq!(snap.logical_bytes, 0);
+    }
+
+    #[test]
+    fn orphan_sidecar_has_no_blob_bytes() {
+        let s = sha(0x55);
+        let c = community(5);
+        let mut agg = BucketAggregate::default();
+        agg.fold(&format!("_meta/{c}/{s}.json"), 20);
+
+        let snap = agg.finish();
+        assert_eq!(snap.orphan_sidecar_count, 1);
+        // The binding still counts as a logical object per D-COUNT, but
+        // contributes zero bytes since there's no blob to bill.
+        assert_eq!(snap.per_community[&c].objects, 1);
+        assert_eq!(snap.per_community[&c].bytes, 0);
+    }
+
+    #[test]
+    fn unmapped_community_binding_still_aggregates_under_its_uuid() {
+        // bucket_index has no DB access — "unmapped" (no matching community
+        // row) is a join the caller performs against per_community's keys.
+        // Here we only assert the raw UUID is preserved for that join.
+        let s = sha(0x66);
+        let c = community(6);
+        let mut agg = BucketAggregate::default();
+        agg.fold(&format!("{s}.jpg"), 40);
+        agg.fold(&format!("_meta/{c}/{s}.json"), 5);
+
+        let snap = agg.finish();
+        assert!(snap.per_community.contains_key(&c));
+    }
+
+    #[test]
+    fn thumb_bytes_attribute_to_the_blobs_sha() {
+        let s = sha(0x77);
+        let c = community(7);
+        let mut agg = BucketAggregate::default();
+        agg.fold(&format!("{s}.jpg"), 1000);
+        agg.fold(&format!("{s}.thumb.jpg"), 50);
+        agg.fold(&format!("_meta/{c}/{s}.json"), 5);
+
+        let snap = agg.finish();
+        assert_eq!(snap.per_community[&c].bytes, 1050);
+        assert_eq!(snap.per_community[&c].objects, 1);
+        // Fleet physical totals count every object separately.
+        assert_eq!(snap.physical_objects, 3);
+        assert_eq!(snap.physical_bytes, 1055);
+    }
+
+    #[test]
+    fn unknown_keys_are_counted_but_excluded_from_logical_math() {
+        let mut agg = BucketAggregate::default();
+        agg.fold("garbage-key", 999);
+        agg.fold("_meta/not-a-uuid/x.json", 1);
+
+        let snap = agg.finish();
+        assert_eq!(snap.unknown_key_objects, 2);
+        assert_eq!(snap.unknown_key_bytes, 1000);
+        assert_eq!(snap.physical_objects, 2);
+        assert_eq!(snap.logical_bytes, 0);
+        assert!(snap.per_community.is_empty());
+    }
+
+    #[test]
+    fn auxiliary_keys_are_physical_only() {
+        let s = sha(0x88);
+        let c = community(8);
+        let ulid = "01ARZ3NDEKTSV4RRFFQ69G5FAV";
+        let mut agg = BucketAggregate::default();
+        agg.fold(&format!("_uploads/{c}/{s}/{ulid}.json"), 30);
+
+        let snap = agg.finish();
+        assert_eq!(snap.physical_objects, 1);
+        assert_eq!(snap.physical_bytes, 30);
+        assert_eq!(snap.logical_bytes, 0);
+        assert!(snap.per_community.is_empty());
+        assert_eq!(snap.unknown_key_objects, 0);
+    }
+
+    // --- fold_bucket_listing (pagination + cap) ---
+
+    #[tokio::test]
+    async fn empty_bucket_listing_yields_zero_snapshot() {
+        let snapshot = fold_bucket_listing(100, |_token| async {
+            Ok(Page {
+                objects: vec![],
+                next_continuation_token: None,
+                is_truncated: false,
+            })
+        })
+        .await
+        .expect("empty listing must not fail");
+        assert_eq!(snapshot, BucketSnapshot::default());
+    }
+
+    #[tokio::test]
+    async fn pagination_follows_continuation_tokens_across_pages() {
+        let s1 = sha(0x99);
+        let s2 = sha(0xa1);
+        let pages = std::sync::Arc::new(std::sync::Mutex::new(vec![
+            Page {
+                objects: vec![(format!("{s1}.jpg"), 10)],
+                next_continuation_token: Some("page-2".to_string()),
+                is_truncated: true,
+            },
+            Page {
+                objects: vec![(format!("{s2}.jpg"), 20)],
+                next_continuation_token: None,
+                is_truncated: false,
+            },
+        ]));
+
+        let snapshot = fold_bucket_listing(100, {
+            let pages = std::sync::Arc::clone(&pages);
+            move |token| {
+                let pages = std::sync::Arc::clone(&pages);
+                async move {
+                    let mut pages = pages.lock().unwrap();
+                    assert!(
+                        (token.is_none() && pages.len() == 2)
+                            || (token.as_deref() == Some("page-2") && pages.len() == 1)
+                    );
+                    Ok(pages.remove(0))
+                }
+            }
+        })
+        .await
+        .expect("two-page listing must succeed");
+
+        assert_eq!(snapshot.physical_objects, 2);
+        assert_eq!(snapshot.physical_bytes, 30);
+        assert_eq!(snapshot.orphan_blob_count, 2);
+    }
+
+    #[tokio::test]
+    async fn cap_breach_mid_listing_fails_the_sweep_before_folding_the_page() {
+        let objects: Vec<(String, u64)> = (0..5).map(|i| (format!("obj-{i}"), 1)).collect();
+        let result = fold_bucket_listing(3, move |_token| {
+            let objects = objects.clone();
+            async move {
+                Ok(Page {
+                    objects,
+                    next_continuation_token: None,
+                    is_truncated: false,
+                })
+            }
+        })
+        .await;
+
+        match result {
+            Err(SweepError::CapExceeded { seen, cap }) => {
+                assert_eq!(seen, 5);
+                assert_eq!(cap, 3);
+            }
+            other => panic!("expected CapExceeded, got {other:?}"),
+        }
+    }
+
+    #[tokio::test]
+    async fn storage_error_propagates_from_page_source() {
+        let result: Result<BucketSnapshot, SweepError> = fold_bucket_listing(10, |_token| async {
+            Err(MediaError::StorageError("boom".to_string()))
+        })
+        .await;
+        assert!(matches!(result, Err(SweepError::Storage(_))));
+    }
+
+    #[tokio::test]
+    async fn truncated_page_with_no_continuation_token_fails_the_sweep() {
+        let result = fold_bucket_listing(100, |_token| async {
+            Ok(Page {
+                objects: vec![("some-key".to_string(), 1)],
+                next_continuation_token: None,
+                is_truncated: true,
+            })
+        })
+        .await;
+        assert!(
+            matches!(result, Err(SweepError::MalformedPage)),
+            "truncated page without a continuation token must fail, not return partial data"
+        );
+    }
+}

crates/buzz-media/src/config.rs

@@ -0,0 +1,254 @@
+//! Media storage configuration.
+
+use std::str::FromStr;
+
+/// S3 URL addressing style shared by media and Git/CAS storage.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, serde::Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum S3AddressingStyle {
+    /// Put the bucket in the request path (`https://endpoint/bucket/key`).
+    ///
+    /// This preserves compatibility with the bundled MinIO deployments, whose
+    /// internal DNS only resolves the endpoint hostname.
+    #[default]
+    Path,
+    /// Put the bucket in the hostname (`https://bucket.endpoint/key`).
+    ///
+    /// This is the standard S3 form and is required by providers such as new
+    /// Railway Storage Buckets.
+    Virtual,
+}
+
+impl FromStr for S3AddressingStyle {
+    type Err = String;
+
+    fn from_str(value: &str) -> Result<Self, Self::Err> {
+        match value {
+            "path" => Ok(Self::Path),
+            "virtual" => Ok(Self::Virtual),
+            _ => Err(format!(
+                "BUZZ_S3_ADDRESSING_STYLE must be 'path' or 'virtual', got {value:?}"
+            )),
+        }
+    }
+}
+
+fn default_max_video_bytes() -> u64 {
+    524_288_000 // 500 MB
+}
+
+fn default_max_file_bytes() -> u64 {
+    104_857_600 // 100 MB
+}
+
+fn default_s3_region() -> String {
+    "us-east-1".to_string()
+}
+
+/// Configuration for media storage (S3/MinIO).
+#[derive(Debug, Clone, serde::Deserialize)]
+pub struct MediaConfig {
+    /// S3-compatible endpoint URL (e.g. "http://localhost:9000").
+    pub s3_endpoint: String,
+    /// S3 access key.
+    pub s3_access_key: String,
+    /// S3 secret key.
+    pub s3_secret_key: String,
+    /// S3 bucket name.
+    pub s3_bucket: String,
+    /// AWS region for SigV4 request signing (e.g. "us-west-2").
+    ///
+    /// Must match the region of `s3_endpoint` for real AWS S3, otherwise
+    /// requests are signed with the wrong credential scope and AWS rejects
+    /// them. Defaults to "us-east-1" to preserve MinIO/local behavior, where
+    /// the value is not meaningfully checked.
+    #[serde(default = "default_s3_region")]
+    pub s3_region: String,
+    /// S3 URL addressing style. Defaults to path style for MinIO compatibility.
+    #[serde(default)]
+    pub s3_addressing_style: S3AddressingStyle,
+    /// Maximum upload size for images (bytes). Default: 50 MB.
+    pub max_image_bytes: u64,
+    /// Maximum upload size for animated GIFs (bytes). Default: 10 MB.
+    pub max_gif_bytes: u64,
+    /// Maximum upload size for video files (bytes). Default: 500 MB.
+    #[serde(default = "default_max_video_bytes")]
+    pub max_video_bytes: u64,
+    /// Maximum upload size for generic (non-image, non-video) files (bytes). Default: 100 MB.
+    #[serde(default = "default_max_file_bytes")]
+    pub max_file_bytes: u64,
+    /// Public base URL for media URLs in BlobDescriptor (must include `/media` path).
+    pub public_base_url: String,
+    /// Whether to write per-upload-event records under `_uploads/`
+    /// (moderation side channel). Off by default; set via
+    /// `BUZZ_MEDIA_UPLOAD_RECORDS=true`.
+    #[serde(default)]
+    pub upload_records_enabled: bool,
+    /// Trusted edge header to read the uploader's public IP from (e.g.
+    /// `cf-connecting-ip`). Unset (default) → no IP is read or recorded.
+    /// Only consulted when `upload_records_enabled` is true; the value is
+    /// validated as a public IP and dropped otherwise (fail-empty).
+    #[serde(default)]
+    pub upload_ip_header: Option<String>,
+    /// Trusted edge header to read the uploader's source port from. Standard
+    /// edges don't emit one, so this is usually unset; a port is only
+    /// recorded alongside a valid IP.
+    #[serde(default)]
+    pub upload_port_header: Option<String>,
+}
+
+impl MediaConfig {
+    /// Validate configuration at startup. Returns an error on invalid config.
+    pub fn validate(&self) -> Result<(), String> {
+        if !self.public_base_url.ends_with("/media") {
+            return Err(format!(
+                "public_base_url must end with /media: got '{}'",
+                self.public_base_url
+            ));
+        }
+        if self.public_base_url.ends_with('/') {
+            return Err(format!(
+                "public_base_url must not end with /: got '{}'",
+                self.public_base_url
+            ));
+        }
+        if self.max_image_bytes == 0 {
+            return Err("max_image_bytes must be > 0".to_string());
+        }
+        if self.max_gif_bytes == 0 || self.max_gif_bytes > self.max_image_bytes {
+            return Err("max_gif_bytes must be > 0 and <= max_image_bytes".to_string());
+        }
+        if self.max_video_bytes == 0 {
+            return Err("max_video_bytes must be > 0".to_string());
+        }
+        if self.max_file_bytes == 0 {
+            return Err("max_file_bytes must be > 0".to_string());
+        }
+        // Fail startup on incoherent collection config instead of silently
+        // recording nothing — an operator who set an IP header believes they
+        // are meeting a reporting obligation.
+        if self.upload_ip_header.is_some() && !self.upload_records_enabled {
+            return Err(
+                "BUZZ_MEDIA_UPLOAD_IP_HEADER is set but BUZZ_MEDIA_UPLOAD_RECORDS is not \
+                 enabled — the IP would never be recorded. Enable upload records or unset \
+                 the header."
+                    .to_string(),
+            );
+        }
+        if self.upload_port_header.is_some() && self.upload_ip_header.is_none() {
+            return Err(
+                "BUZZ_MEDIA_UPLOAD_PORT_HEADER is set without BUZZ_MEDIA_UPLOAD_IP_HEADER — \
+                 a port is only recorded alongside an IP. Set the IP header or unset the \
+                 port header."
+                    .to_string(),
+            );
+        }
+        for (name, value) in [
+            ("BUZZ_MEDIA_UPLOAD_IP_HEADER", &self.upload_ip_header),
+            ("BUZZ_MEDIA_UPLOAD_PORT_HEADER", &self.upload_port_header),
+        ] {
+            if let Some(h) = value {
+                if axum::http::HeaderName::from_bytes(h.as_bytes()).is_err() {
+                    return Err(format!("{name} is not a valid header name: {h:?}"));
+                }
+            }
+        }
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{MediaConfig, S3AddressingStyle};
+    use std::str::FromStr;
+
+    fn valid_config() -> MediaConfig {
+        MediaConfig {
+            s3_endpoint: "http://localhost:9000".to_string(),
+            s3_access_key: "k".to_string(),
+            s3_secret_key: "s".to_string(),
+            s3_bucket: "buzz-media".to_string(),
+            s3_region: "us-east-1".to_string(),
+            s3_addressing_style: S3AddressingStyle::Path,
+            max_image_bytes: 1,
+            max_gif_bytes: 1,
+            max_video_bytes: 1,
+            max_file_bytes: 1,
+            public_base_url: "http://localhost:3000/media".to_string(),
+            upload_records_enabled: false,
+            upload_ip_header: None,
+            upload_port_header: None,
+        }
+    }
+
+    #[test]
+    fn addressing_style_parses_supported_values() {
+        assert_eq!(
+            S3AddressingStyle::from_str("path"),
+            Ok(S3AddressingStyle::Path)
+        );
+        assert_eq!(
+            S3AddressingStyle::from_str("virtual"),
+            Ok(S3AddressingStyle::Virtual)
+        );
+    }
+
+    #[test]
+    fn addressing_style_defaults_to_path() {
+        assert_eq!(S3AddressingStyle::default(), S3AddressingStyle::Path);
+    }
+
+    #[test]
+    fn addressing_style_rejects_unknown_or_ambiguous_values() {
+        for invalid in ["", "auto", "PATH", "virtual-hosted"] {
+            let error =
+                S3AddressingStyle::from_str(invalid).expect_err("must reject invalid style");
+            assert!(
+                error.contains("BUZZ_S3_ADDRESSING_STYLE must be 'path' or 'virtual'"),
+                "unexpected error for {invalid:?}: {error}"
+            );
+        }
+    }
+
+    #[test]
+    fn upload_record_knobs_default_off_and_validate() {
+        assert!(valid_config().validate().is_ok());
+
+        let mut on = valid_config();
+        on.upload_records_enabled = true;
+        assert!(on.validate().is_ok());
+
+        on.upload_ip_header = Some("cf-connecting-ip".to_string());
+        assert!(on.validate().is_ok());
+
+        on.upload_port_header = Some("x-client-port".to_string());
+        assert!(on.validate().is_ok());
+    }
+
+    #[test]
+    fn ip_header_without_records_fails_startup() {
+        // An operator who set the header believes IPs are being recorded —
+        // fail loudly instead of silently collecting nothing.
+        let mut cfg = valid_config();
+        cfg.upload_ip_header = Some("cf-connecting-ip".to_string());
+        assert!(cfg.validate().is_err());
+    }
+
+    #[test]
+    fn port_header_without_ip_header_fails_startup() {
+        let mut cfg = valid_config();
+        cfg.upload_records_enabled = true;
+        cfg.upload_port_header = Some("x-client-port".to_string());
+        assert!(cfg.validate().is_err());
+    }
+
+    #[test]
+    fn malformed_header_names_fail_startup() {
+        let mut cfg = valid_config();
+        cfg.upload_records_enabled = true;
+        for bad in ["with space", "colon:name", "bad/header", "bad,header", ""] {
+            cfg.upload_ip_header = Some(bad.to_string());
+            assert!(cfg.validate().is_err(), "should reject header name {bad:?}");
+        }
+    }
+}

crates/buzz-media/src/error.rs

@@ -0,0 +1,198 @@
+//! Media error types.
+
+use axum::http::StatusCode;
+use axum::response::{IntoResponse, Response};
+
+/// Errors from media operations.
+#[derive(Debug, thiserror::Error)]
+pub enum MediaError {
+    #[error("unknown content type")]
+    UnknownContentType,
+    #[error("disallowed content type: {0}")]
+    DisallowedContentType(String),
+    #[error("file too large: {size} bytes (max {max})")]
+    FileTooLarge { size: u64, max: u64 },
+    #[error("image dimensions too large")]
+    ImageTooLarge,
+    #[error("invalid image data")]
+    InvalidImage,
+    #[error("media contains metadata or a non-canonical metadata channel")]
+    MetadataForbidden,
+    #[error("invalid signature")]
+    InvalidSignature,
+    #[error("invalid auth event kind")]
+    InvalidAuthKind,
+    #[error("invalid auth verb")]
+    InvalidAuthVerb,
+    #[error("missing required tag: {0}")]
+    MissingTag(&'static str),
+    #[error("hash mismatch")]
+    HashMismatch,
+    #[error("server mismatch")]
+    ServerMismatch,
+    #[error("token expired")]
+    TokenExpired,
+    #[error("timestamp out of window")]
+    TimestampOutOfWindow,
+    #[error("storage error: {0}")]
+    StorageError(String),
+    #[error("internal error")]
+    Internal,
+    #[error("not found")]
+    NotFound,
+    #[error("missing authorization header")]
+    MissingAuth,
+    #[error("invalid authorization scheme")]
+    InvalidAuthScheme,
+    #[error("invalid base64 encoding")]
+    InvalidBase64,
+    #[error("invalid auth event")]
+    InvalidAuthEvent,
+    #[error("unauthorized")]
+    Unauthorized,
+    #[error("insufficient scope")]
+    InsufficientScope,
+    #[error("relay membership required")]
+    RelayMembershipRequired,
+    #[error("token revoked")]
+    TokenRevoked,
+    #[error("pubkey mismatch")]
+    PubkeyMismatch,
+    #[error("upload rate limit exceeded")]
+    UploadRateLimitExceeded,
+    #[error("upload concurrency limit reached")]
+    UploadConcurrencyLimitReached,
+    /// A video/audio track does not use the canonical H.264/AAC codecs.
+    #[error("unsupported media codec: only H.264 video and AAC audio are accepted")]
+    WrongCodec,
+    /// Video duration exceeds the 600-second limit.
+    #[error("video too long: duration exceeds 600 seconds")]
+    DurationTooLong,
+    /// Video resolution exceeds 3840×2160.
+    #[error("video resolution too high: maximum is 3840x2160")]
+    ResolutionTooHigh,
+    /// MP4 moov atom appears after mdat — not fast-start.
+    #[error("moov atom not at front of file (not fast-start)")]
+    MoovNotAtFront,
+    /// Container is not MP4 (e.g. MOV, MKV).
+    #[error("unsupported container: only MP4 is accepted")]
+    UnsupportedContainer,
+    /// MP4 metadata could not be parsed.
+    #[error("invalid video data")]
+    InvalidVideo,
+    /// I/O error during streaming upload.
+    #[error("io error: {0}")]
+    Io(String),
+}
+
+impl From<image::ImageError> for MediaError {
+    fn from(_: image::ImageError) -> Self {
+        Self::InvalidImage
+    }
+}
+
+impl From<s3::error::S3Error> for MediaError {
+    fn from(e: s3::error::S3Error) -> Self {
+        Self::StorageError(e.to_string())
+    }
+}
+
+impl From<serde_json::Error> for MediaError {
+    fn from(e: serde_json::Error) -> Self {
+        Self::StorageError(e.to_string())
+    }
+}
+
+impl IntoResponse for MediaError {
+    fn into_response(self) -> Response {
+        let (status, msg) = match &self {
+            Self::NotFound => (StatusCode::NOT_FOUND, self.to_string()),
+            Self::DisallowedContentType(_) => {
+                (StatusCode::UNSUPPORTED_MEDIA_TYPE, self.to_string())
+            }
+            Self::FileTooLarge { .. } | Self::ImageTooLarge => {
+                (StatusCode::PAYLOAD_TOO_LARGE, self.to_string())
+            }
+            // All authentication failures return the same generic 401 to prevent oracle enumeration.
+            // InsufficientScope is intentionally 403 — it's an authorization (not authentication)
+            // failure and is safe to distinguish since it requires a valid identity first.
+            Self::MissingAuth
+            | Self::InvalidAuthScheme
+            | Self::InvalidBase64
+            | Self::InvalidAuthEvent
+            | Self::InvalidSignature
+            | Self::InvalidAuthKind
+            | Self::InvalidAuthVerb
+            | Self::TokenExpired
+            | Self::TimestampOutOfWindow
+            | Self::Unauthorized
+            | Self::TokenRevoked
+            | Self::PubkeyMismatch
+            | Self::HashMismatch
+            | Self::ServerMismatch
+            | Self::MissingTag(_) => {
+                tracing::warn!(error = %self, "authentication failed");
+                (
+                    StatusCode::UNAUTHORIZED,
+                    "authentication failed".to_string(),
+                )
+            }
+            Self::InsufficientScope => (StatusCode::FORBIDDEN, self.to_string()),
+            Self::RelayMembershipRequired => (StatusCode::FORBIDDEN, self.to_string()),
+            Self::UploadRateLimitExceeded | Self::UploadConcurrencyLimitReached => {
+                (StatusCode::TOO_MANY_REQUESTS, self.to_string())
+            }
+            Self::UnknownContentType | Self::UnsupportedContainer | Self::WrongCodec => {
+                (StatusCode::UNSUPPORTED_MEDIA_TYPE, self.to_string())
+            }
+            Self::DurationTooLong
+            | Self::ResolutionTooHigh
+            | Self::MoovNotAtFront
+            | Self::InvalidVideo
+            | Self::InvalidImage
+            | Self::MetadataForbidden => (StatusCode::UNPROCESSABLE_ENTITY, self.to_string()),
+            Self::Io(_) | Self::StorageError(_) | Self::Internal => {
+                tracing::error!(error = %self, "media storage error");
+                (StatusCode::INTERNAL_SERVER_ERROR, "internal error".into())
+            }
+        };
+        (status, axum::Json(serde_json::json!({"error": msg}))).into_response()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn unsupported_media_maps_to_415() {
+        for error in [
+            MediaError::UnknownContentType,
+            MediaError::DisallowedContentType("audio/mpeg".to_string()),
+            MediaError::UnsupportedContainer,
+            MediaError::WrongCodec,
+        ] {
+            assert_eq!(
+                error.into_response().status(),
+                StatusCode::UNSUPPORTED_MEDIA_TYPE
+            );
+        }
+    }
+
+    #[test]
+    fn invalid_or_noncanonical_media_maps_to_422() {
+        for error in [
+            MediaError::InvalidImage,
+            MediaError::InvalidVideo,
+            MediaError::MetadataForbidden,
+            MediaError::MoovNotAtFront,
+            MediaError::DurationTooLong,
+            MediaError::ResolutionTooHigh,
+        ] {
+            assert_eq!(
+                error.into_response().status(),
+                StatusCode::UNPROCESSABLE_ENTITY
+            );
+        }
+    }
+}

crates/buzz-media/src/lib.rs

@@ -0,0 +1,29 @@
+//! Media storage, validation, and thumbnail generation for Buzz.
+//!
+//! Library crate — no Axum dependency for handlers. Axum handlers live in `buzz-relay`.
+
+pub mod auth;
+pub mod bucket_index;
+pub mod config;
+pub mod error;
+pub mod storage;
+pub mod thumbnail;
+pub mod types;
+pub mod upload;
+pub mod upload_record;
+pub mod validation;
+
+pub use bucket_index::{
+    classify_key, fold_bucket_listing, BucketAggregate, BucketSnapshot, CommunityStorage, KeyClass,
+    Page, SweepError,
+};
+pub use config::{MediaConfig, S3AddressingStyle};
+pub use error::MediaError;
+pub use storage::{BlobHeadMeta, BlobMeta, ByteStream, MediaStorage};
+pub use types::BlobDescriptor;
+pub use upload::{process_file_upload, process_upload, process_video_upload};
+pub use upload_record::{
+    parse_port, parse_public_ip, upload_record_key, UploadAttribution, UploadNetworkInfo,
+    UploadRecord, UPLOAD_RECORD_VERSION,
+};
+pub use validation::{looks_like_iso_bmff, serve_inline, validate_video_file, VideoMeta};

crates/buzz-media/src/storage.rs

@@ -0,0 +1,425 @@
+//! S3/MinIO storage client.
+
+use std::path::Path;
+use std::pin::Pin;
+
+use buzz_core::tenant::{CommunityId, TenantContext};
+
+use crate::config::{MediaConfig, S3AddressingStyle};
+use crate::error::MediaError;
+use bytes::Bytes;
+use s3::creds::Credentials;
+use s3::{Bucket, Region};
+use serde::{Deserialize, Serialize};
+
+/// A stream of byte chunks from S3, usable with `axum::body::Body::from_stream()`.
+pub type ByteStream = Pin<Box<dyn futures_core::Stream<Item = Result<Bytes, MediaError>> + Send>>;
+
+/// S3-compatible object storage client.
+pub struct MediaStorage {
+    bucket: Box<Bucket>,
+}
+
+impl MediaStorage {
+    /// Create a new storage client from config.
+    ///
+    /// Credential selection:
+    /// - If both `s3_access_key` and `s3_secret_key` are non-empty, use them as
+    ///   static credentials (MinIO/local/dev, or any static-key deployment).
+    /// - Otherwise, fall back to the AWS default credential chain via
+    ///   [`Credentials::default`]: environment, shared profile, web-identity
+    ///   token (IRSA on EKS — `AssumeRoleWithWebIdentity`), container, and
+    ///   instance-metadata providers, in that order. This lets the relay use
+    ///   the pod's IAM role without long-lived static keys.
+    pub fn new(config: &MediaConfig) -> Result<Self, MediaError> {
+        let region = Region::Custom {
+            region: config.s3_region.clone(),
+            endpoint: config.s3_endpoint.clone(),
+        };
+        let creds = match (
+            config.s3_access_key.is_empty(),
+            config.s3_secret_key.is_empty(),
+        ) {
+            (false, false) => Credentials::new(
+                Some(&config.s3_access_key),
+                Some(&config.s3_secret_key),
+                None,
+                None,
+                None,
+            ),
+            (true, true) => {
+                // No static keys configured: resolve from the AWS credential chain
+                // (IRSA web-identity, env, profile, instance metadata).
+                Credentials::default()
+            }
+            _ => {
+                return Err(MediaError::StorageError(
+                    "s3_access_key and s3_secret_key must be configured together, or both empty to use the AWS credential chain"
+                        .to_string(),
+                ));
+            }
+        }
+        .map_err(|e| MediaError::StorageError(e.to_string()))?;
+        let bucket = Bucket::new(&config.s3_bucket, region, creds)
+            .map_err(|e| MediaError::StorageError(e.to_string()))?;
+        let bucket = match config.s3_addressing_style {
+            S3AddressingStyle::Path => bucket.with_path_style(),
+            S3AddressingStyle::Virtual => bucket,
+        };
+        Ok(Self { bucket })
+    }
+
+    /// Store an object from a byte slice.
+    ///
+    /// Used for images, sidecars, and thumbnails. For large video files use
+    /// [`put_file`] to avoid loading the entire blob into RAM.
+    pub async fn put(&self, key: &str, bytes: &[u8], content_type: &str) -> Result<(), MediaError> {
+        self.bucket
+            .put_object_with_content_type(key, bytes, content_type)
+            .await?;
+        Ok(())
+    }
+
+    /// Stream a file from disk into S3 without loading it into RAM.
+    ///
+    /// Uses rust-s3's `put_object_stream_with_content_type` which reads from
+    /// the file incrementally via an 8 MiB `BufReader`. The full file is never
+    /// held in memory simultaneously. Intended for video blobs (up to 500 MB).
+    pub async fn put_file(
+        &self,
+        key: &str,
+        path: &Path,
+        content_type: &str,
+    ) -> Result<(), MediaError> {
+        const BUF: usize = 8 * 1024 * 1024; // 8 MiB read buffer
+
+        let file = tokio::fs::File::open(path)
+            .await
+            .map_err(|e| MediaError::Io(e.to_string()))?;
+        let mut reader = tokio::io::BufReader::with_capacity(BUF, file);
+
+        self.bucket
+            .put_object_stream_with_content_type(&mut reader, key, content_type)
+            .await?;
+        Ok(())
+    }
+
+    /// Retrieve an object's bytes.
+    pub async fn get(&self, key: &str) -> Result<Vec<u8>, MediaError> {
+        match self.bucket.get_object(key).await {
+            Ok(response) => Ok(response.to_vec()),
+            Err(s3::error::S3Error::HttpFailWithBody(404, _)) => Err(MediaError::NotFound),
+            Err(e) => Err(MediaError::StorageError(e.to_string())),
+        }
+    }
+
+    /// Retrieve a byte range from an object via S3-native `Range` GET.
+    ///
+    /// `start` and `end` are inclusive byte offsets. Only the requested slice
+    /// is transferred from S3 — the full object is never loaded into RAM.
+    /// Intended for HTTP 206 range responses on large video blobs.
+    pub async fn get_range(&self, key: &str, start: u64, end: u64) -> Result<Vec<u8>, MediaError> {
+        match self.bucket.get_object_range(key, start, Some(end)).await {
+            Ok(response) => Ok(response.to_vec()),
+            Err(s3::error::S3Error::HttpFailWithBody(404, _)) => Err(MediaError::NotFound),
+            Err(e) => Err(MediaError::StorageError(e.to_string())),
+        }
+    }
+
+    /// Stream an object's bytes from S3 without loading into RAM.
+    ///
+    /// Returns a pinned stream of `Result<Bytes, MediaError>` chunks.
+    /// The full object is never buffered — intended for streaming large
+    /// blobs (video) directly into HTTP responses via `Body::from_stream()`.
+    pub async fn get_stream(&self, key: &str) -> Result<ByteStream, MediaError> {
+        let response = self
+            .bucket
+            .get_object_stream(key)
+            .await
+            .map_err(|e| MediaError::StorageError(e.to_string()))?;
+
+        if response.status_code == 404 {
+            return Err(MediaError::NotFound);
+        }
+
+        let stream = futures_util::StreamExt::map(response.bytes, |chunk| {
+            chunk.map_err(|e| MediaError::StorageError(e.to_string()))
+        });
+        Ok(Box::pin(stream))
+    }
+
+    /// Check if an object exists. Returns false on 404.
+    pub async fn head(&self, key: &str) -> Result<bool, MediaError> {
+        match self.bucket.head_object(key).await {
+            Ok(_) => Ok(true),
+            Err(s3::error::S3Error::HttpFailWithBody(404, _)) => Ok(false),
+            Err(e) => Err(MediaError::StorageError(e.to_string())),
+        }
+    }
+
+    /// Delete an object. Returns an error on failure — callers decide whether to propagate.
+    pub async fn delete(&self, key: &str) -> Result<(), MediaError> {
+        self.bucket
+            .delete_object(key)
+            .await
+            .map_err(|e| MediaError::StorageError(e.to_string()))?;
+        Ok(())
+    }
+
+    /// HEAD with metadata — returns Content-Length (size).
+    pub async fn head_with_metadata(&self, key: &str) -> Result<Option<BlobHeadMeta>, MediaError> {
+        match self.bucket.head_object(key).await {
+            Ok((result, _)) => Ok(Some(BlobHeadMeta {
+                size: result.content_length.unwrap_or(0) as u64,
+            })),
+            Err(s3::error::S3Error::HttpFailWithBody(404, _)) => Ok(None),
+            Err(e) => Err(MediaError::StorageError(e.to_string())),
+        }
+    }
+
+    /// Build the community-scoped sidecar key for a given sha256 (bare hash).
+    ///
+    /// Raw media bytes remain shared content-addressed CAS (`{sha}.{ext}`), but
+    /// the metadata sidecar is the tenant read gate. A blob in another
+    /// community must never be observable through a global `_meta/{sha}.json`
+    /// lookup.
+    pub fn sidecar_key(community: CommunityId, sha256: &str) -> String {
+        format!("_meta/{community}/{sha256}.json")
+    }
+
+    /// Build the community-scoped sidecar key from the resolved request tenant.
+    pub fn ctx_sidecar_key(ctx: &TenantContext, sha256: &str) -> String {
+        Self::sidecar_key(ctx.community(), sha256)
+    }
+
+    /// Read community-scoped sidecar JSON for a given sha256 (bare hash).
+    pub async fn get_sidecar(
+        &self,
+        ctx: &TenantContext,
+        sha256: &str,
+    ) -> Result<BlobMeta, MediaError> {
+        let key = Self::ctx_sidecar_key(ctx, sha256);
+        let resp = self.bucket.get_object(&key).await?;
+        let meta: BlobMeta = serde_json::from_slice(&resp.to_vec())?;
+        Ok(meta)
+    }
+
+    /// Write community-scoped sidecar JSON for a given sha256 (bare hash).
+    ///
+    /// `ctx` must be the server-resolved request tenant. Callers must never
+    /// derive the community from client-supplied blob metadata, URLs, or event
+    /// tags; this sidecar key is the tenant read gate for otherwise shared CAS
+    /// bytes.
+    pub async fn put_sidecar(
+        &self,
+        ctx: &TenantContext,
+        sha256: &str,
+        meta: &BlobMeta,
+    ) -> Result<(), MediaError> {
+        let key = Self::ctx_sidecar_key(ctx, sha256);
+        let meta_json = serde_json::to_vec(meta)?;
+        self.put(&key, &meta_json, "application/json").await
+    }
+
+    /// Convenience: read just the MIME type from the community sidecar.
+    ///
+    /// Returns `None` for both absent sidecars and storage read failures. Public
+    /// read handlers intentionally collapse that distinction to 404 so an
+    /// A-bound request cannot distinguish a B-only blob from a missing blob.
+    pub async fn read_sidecar_mime(&self, ctx: &TenantContext, sha256_ext: &str) -> Option<String> {
+        let sha256 = sha256_ext.split('.').next().unwrap_or(sha256_ext);
+        self.get_sidecar(ctx, sha256)
+            .await
+            .ok()
+            .map(|m| m.mime_type)
+    }
+
+    /// One page of a full-bucket listing, for the storage sweep. Wraps
+    /// rust-s3's manual `list_page` (NOT the auto-paginating `list`, which
+    /// has no cap) and converts the result into the storage-agnostic
+    /// [`crate::bucket_index::Page`] shape the pure fold consumes.
+    ///
+    /// `max_keys` bounds one HTTP response, not the sweep's total object
+    /// cap — the caller (`fold_bucket_listing`) enforces the cumulative cap
+    /// across pages.
+    pub async fn list_page(
+        &self,
+        continuation_token: Option<String>,
+        max_keys: usize,
+    ) -> Result<crate::bucket_index::Page, MediaError> {
+        let (result, _status) = self
+            .bucket
+            .list_page(
+                String::new(),
+                None,
+                continuation_token,
+                None,
+                Some(max_keys),
+            )
+            .await?;
+        Ok(crate::bucket_index::Page {
+            objects: result
+                .contents
+                .into_iter()
+                .map(|obj| (obj.key, obj.size))
+                .collect(),
+            next_continuation_token: result.next_continuation_token,
+            is_truncated: result.is_truncated,
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::collections::HashMap;
+
+    fn tenant(n: u128) -> TenantContext {
+        TenantContext::resolved(
+            CommunityId::from_uuid(uuid::Uuid::from_u128(n)),
+            "media.example",
+        )
+    }
+
+    fn storage_config(access: &str, secret: &str) -> crate::config::MediaConfig {
+        crate::config::MediaConfig {
+            s3_endpoint: "http://localhost:9000".to_string(),
+            s3_access_key: access.to_string(),
+            s3_secret_key: secret.to_string(),
+            s3_bucket: "buzz-media".to_string(),
+            s3_region: "us-west-2".to_string(),
+            s3_addressing_style: S3AddressingStyle::Path,
+            max_image_bytes: 50 * 1024 * 1024,
+            max_gif_bytes: 10 * 1024 * 1024,
+            max_video_bytes: 524_288_000,
+            max_file_bytes: 104_857_600,
+            public_base_url: "http://localhost:3000/media".to_string(),
+            upload_records_enabled: false,
+            upload_ip_header: None,
+            upload_port_header: None,
+        }
+    }
+
+    /// Static keys present: builds a client without touching the AWS
+    /// credential chain (no env/metadata access), and the signing region
+    /// comes from config rather than a hardcoded "us-east-1".
+    #[test]
+    fn static_keys_build_client_with_configured_region() {
+        let storage = MediaStorage::new(&storage_config("buzz_dev", "buzz_dev_secret"))
+            .expect("static creds should build a client");
+        match storage.bucket.region {
+            Region::Custom { ref region, .. } => assert_eq!(region, "us-west-2"),
+            other => panic!("expected Custom region, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn client_constructor_applies_both_addressing_styles() {
+        let path = MediaStorage::new(&storage_config("buzz_dev", "buzz_dev_secret"))
+            .expect("path-style client");
+        assert!(path.bucket.is_path_style());
+        assert_eq!(path.bucket.url(), "http://localhost:9000/buzz-media");
+
+        let mut virtual_config = storage_config("buzz_dev", "buzz_dev_secret");
+        virtual_config.s3_addressing_style = S3AddressingStyle::Virtual;
+        let virtual_hosted = MediaStorage::new(&virtual_config).expect("virtual-hosted client");
+        assert!(virtual_hosted.bucket.is_subdomain_style());
+        assert_eq!(
+            virtual_hosted.bucket.url(),
+            "http://buzz-media.localhost:9000"
+        );
+    }
+
+    #[test]
+    fn partial_static_keys_are_rejected() {
+        let err = match MediaStorage::new(&storage_config("buzz_dev", "")) {
+            Ok(_) => panic!("partial static creds must not silently use credential chain"),
+            Err(err) => err,
+        };
+        assert!(
+            err.to_string().contains("must be configured together"),
+            "unexpected error: {err}"
+        );
+
+        let err = match MediaStorage::new(&storage_config("", "buzz_dev_secret")) {
+            Ok(_) => panic!("partial static creds must not silently use credential chain"),
+            Err(err) => err,
+        };
+        assert!(
+            err.to_string().contains("must be configured together"),
+            "unexpected error: {err}"
+        );
+    }
+
+    #[test]
+    fn sidecar_keys_are_community_scoped() {
+        let a = tenant(1);
+        let b = tenant(2);
+        let sha = "f".repeat(64);
+
+        assert_eq!(
+            MediaStorage::ctx_sidecar_key(&a, &sha),
+            format!("_meta/{}/{sha}.json", a.community())
+        );
+        assert_ne!(
+            MediaStorage::ctx_sidecar_key(&a, &sha),
+            MediaStorage::ctx_sidecar_key(&b, &sha)
+        );
+        assert_ne!(
+            MediaStorage::ctx_sidecar_key(&a, &sha),
+            format!("_meta/{sha}.json")
+        );
+    }
+
+    /// Mutate-bite shape for the media substrate: same CAS bytes/hash can be
+    /// known in A and B, but the sidecar is the read/existence gate. If the
+    /// community segment is dropped from `sidecar_key`, B's metadata overwrites
+    /// A's in this map and A observes B's MIME (wrong answer, not absence).
+    #[test]
+    fn same_sha_sidecars_do_not_bleed_between_communities() {
+        let a = tenant(1);
+        let b = tenant(2);
+        let sha = "a".repeat(64);
+        let mut sidecars = HashMap::new();
+
+        sidecars.insert(MediaStorage::ctx_sidecar_key(&a, &sha), "image/png");
+        sidecars.insert(MediaStorage::ctx_sidecar_key(&b, &sha), "video/mp4");
+
+        assert_eq!(
+            sidecars[&MediaStorage::ctx_sidecar_key(&a, &sha)],
+            "image/png"
+        );
+        assert_eq!(
+            sidecars[&MediaStorage::ctx_sidecar_key(&b, &sha)],
+            "video/mp4"
+        );
+    }
+}
+
+/// Metadata returned by HEAD — just enough for BUD-01 response headers.
+pub struct BlobHeadMeta {
+    pub size: u64,
+}
+
+/// Full blob metadata — stored as sidecar JSON in `_meta/{community}/{sha256}.json`.
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+pub struct BlobMeta {
+    /// Pixel dimensions ("WxH").
+    pub dim: String,
+    /// Blurhash string.
+    pub blurhash: String,
+    /// Full URL to thumbnail.
+    pub thumb_url: String,
+    /// File extension (e.g. "jpg").
+    pub ext: String,
+    /// MIME type (e.g. "image/jpeg").
+    pub mime_type: String,
+    /// File size in bytes.
+    pub size: u64,
+    /// Unix timestamp when the blob was first uploaded.
+    #[serde(default)]
+    pub uploaded_at: i64,
+    /// Video duration in seconds. `None` for non-video blobs.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub duration_secs: Option<f64>,
+}

crates/buzz-media/src/thumbnail.rs

@@ -0,0 +1,51 @@
+//! Synchronous thumbnail generation and blurhash encoding.
+
+use std::io::Cursor;
+
+use image::ImageFormat;
+
+use crate::config::MediaConfig;
+use crate::error::MediaError;
+use crate::storage::BlobMeta;
+
+/// Generate thumbnail and blurhash from image bytes (CPU-bound, sync).
+///
+/// Returns `(metadata, optional thumbnail JPEG bytes)`.
+/// Caller handles S3 writes after `spawn_blocking` returns.
+pub fn generate_image_metadata_sync(
+    config: &MediaConfig,
+    sha256: &str,
+    bytes: &[u8],
+    mime: &str,
+    ext: &str,
+) -> Result<(BlobMeta, Option<Vec<u8>>), MediaError> {
+    if !mime.starts_with("image/") {
+        return Ok((BlobMeta::default(), None));
+    }
+
+    let img = image::load_from_memory(bytes)?;
+    let (w, h) = (img.width(), img.height());
+
+    // Thumbnail: 320px max dimension, preserve aspect ratio
+    let thumb = img.thumbnail(320, 320);
+    let mut thumb_bytes = Vec::new();
+    thumb.write_to(&mut Cursor::new(&mut thumb_bytes), ImageFormat::Jpeg)?;
+
+    // Blurhash from thumbnail (faster than full image)
+    let rgba = thumb.to_rgba8();
+    let bh =
+        blurhash::encode(4, 3, thumb.width(), thumb.height(), rgba.as_raw()).unwrap_or_default();
+
+    Ok((
+        BlobMeta {
+            dim: format!("{w}x{h}"),
+            blurhash: bh,
+            thumb_url: format!("{}/{sha256}.thumb.jpg", config.public_base_url),
+            ext: ext.to_string(),
+            mime_type: mime.to_string(),
+            size: bytes.len() as u64,
+            ..BlobMeta::default() // uploaded_at set by caller (process_upload)
+        },
+        Some(thumb_bytes),
+    ))
+}

crates/buzz-media/src/types.rs

@@ -0,0 +1,31 @@
+//! Blossom BUD-02 response types.
+
+use serde::{Deserialize, Serialize};
+
+/// Blossom BlobDescriptor — returned by `PUT /upload` and the legacy media alias.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct BlobDescriptor {
+    /// Full URL to the blob.
+    pub url: String,
+    /// SHA-256 hex hash (64 chars).
+    pub sha256: String,
+    /// File size in bytes.
+    pub size: u64,
+    /// MIME type.
+    #[serde(rename = "type")]
+    pub mime_type: String,
+    /// Unix timestamp of upload.
+    pub uploaded: i64,
+    /// Pixel dimensions ("WxH").
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub dim: Option<String>,
+    /// Blurhash string.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub blurhash: Option<String>,
+    /// Thumbnail URL.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub thumb: Option<String>,
+    /// Video duration in seconds. `None` for non-video blobs.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub duration: Option<f64>,
+}

crates/buzz-media/src/upload.rs

@@ -0,0 +1,733 @@
+//! Upload pipeline — validate, store, thumbnail, sidecar.
+
+use buzz_core::tenant::TenantContext;
+use bytes::Bytes;
+use sha2::{Digest, Sha256};
+use tokio::io::AsyncWriteExt;
+
+use crate::auth::verify_blossom_upload_auth;
+use crate::config::MediaConfig;
+use crate::error::MediaError;
+use crate::storage::{BlobMeta, MediaStorage};
+use crate::thumbnail::generate_image_metadata_sync;
+use crate::types::BlobDescriptor;
+use crate::upload_record::{record_upload_event, UploadAttribution, UploadEventFacts};
+use crate::validation::{
+    looks_like_mp4_iso_bmff, mime_to_ext, validate_content, validate_file_content,
+    validate_video_file,
+};
+
+/// Shared buffered-upload pipeline for the image and generic-file paths.
+///
+/// Both paths are identical except for two steps, which are injected:
+/// - `validate`: a CPU-bound check (run inside `spawn_blocking`) that returns
+///   the `(mime, ext)` pair for the body. Images derive `ext` from the MIME;
+///   generic files get both from the deny-list validator.
+/// - `prepare_metadata`: builds metadata and stores any derived artifacts such
+///   as a thumbnail, but deliberately does not write the sidecar. The sidecar
+///   is the media serve gate and is published only after the moderation record
+///   succeeds. It receives the already-computed
+///   `(sha256, ext, mime, uploaded_at)` so no work is repeated.
+///
+/// Everything else — hash, Blossom auth (10-minute window), content-addressed
+/// key, the both-exist idempotency short-circuit, blob store, orphan-blob
+/// handling, and descriptor build — is common. The streaming video path stays
+/// separate (see [`process_video_upload`]) because it never buffers in RAM.
+///
+/// `attribution` is `Some` when per-event upload records are enabled
+/// (`BUZZ_MEDIA_UPLOAD_RECORDS`): a record is then written for **every**
+/// accepted upload — including the idempotent short-circuit, which does no
+/// blob PUT and would otherwise be invisible to the moderation pipeline.
+/// For fresh uploads, the record is written after the blob and derived
+/// artifacts but before the sidecar. This preserves both contracts: record
+/// existence implies referenced objects are readable, while a record failure
+/// cannot publish media without triggering moderation.
+struct BufferedUploadInput<'a> {
+    storage: &'a MediaStorage,
+    config: &'a MediaConfig,
+    ctx: &'a TenantContext,
+    auth_event: &'a nostr::Event,
+    body: Bytes,
+    attribution: Option<UploadAttribution>,
+}
+
+async fn process_buffered_upload<V, M, Fut>(
+    input: BufferedUploadInput<'_>,
+    validate: V,
+    prepare_metadata: M,
+) -> Result<BlobDescriptor, MediaError>
+where
+    V: FnOnce(&Bytes, &MediaConfig) -> Result<(String, String), MediaError> + Send + 'static,
+    M: FnOnce(MetadataInput) -> Fut,
+    Fut: std::future::Future<Output = Result<BlobMeta, MediaError>>,
+{
+    let BufferedUploadInput {
+        storage,
+        config,
+        ctx,
+        auth_event,
+        body,
+        attribution,
+    } = input;
+
+    // CPU-bound: validate content, compute hash, verify auth.
+    let auth = auth_event.clone();
+    let bytes = body.clone();
+    let cfg = config.clone();
+    // Validate the Blossom `server` tag against the host this request was bound
+    // to (the per-request tenant), not a process-global domain — a relay serves
+    // many tenant hosts.
+    let bound_host = ctx.host().to_string();
+    let (mime, sha256, ext) = tokio::task::spawn_blocking(move || -> Result<_, MediaError> {
+        let (mime, ext) = validate(&bytes, &cfg)?;
+        let sha256 = hex::encode(Sha256::digest(&bytes));
+        // Buffered uploads (image + file): 10-minute auth window is plenty.
+        verify_blossom_upload_auth(&auth, &sha256, Some(bound_host.as_str()), 600)?;
+        Ok((mime, sha256, ext))
+    })
+    .await
+    .map_err(|_| MediaError::Internal)??;
+
+    let key = format!("{sha256}.{ext}");
+    let meta_key = MediaStorage::ctx_sidecar_key(ctx, &sha256);
+
+    // Idempotent: short-circuit only if BOTH sidecar and blob exist. If the
+    // sidecar exists but the blob is missing, fall through to re-upload.
+    let sidecar_exists = storage.head(&meta_key).await?;
+    let blob_exists = storage.head(&key).await?;
+    if sidecar_exists && blob_exists {
+        let meta = storage.get_sidecar(ctx, &sha256).await?;
+        // A re-upload of known bytes is still a distinct upload *event*: no
+        // blob PUT happens, so without this record the uploader would be
+        // invisible to the moderation pipeline (and takedown re-uploads
+        // would go unscanned).
+        if let Some(attribution) = &attribution {
+            record_upload_event(
+                storage,
+                ctx,
+                &auth_event.pubkey,
+                attribution,
+                UploadEventFacts {
+                    sha256: &sha256,
+                    ext: &ext,
+                    mime: &mime,
+                    size: body.len() as u64,
+                    uploaded_at: chrono::Utc::now().timestamp(),
+                },
+            )
+            .await?;
+        }
+        return Ok(build_descriptor(
+            config,
+            &sha256,
+            &ext,
+            &mime,
+            body.len() as u64,
+            Some(&meta),
+            meta.uploaded_at,
+        ));
+    }
+
+    // Compute uploaded_at once — single source of truth for sidecar and response.
+    let uploaded_at = chrono::Utc::now().timestamp();
+
+    // Store blob first, then metadata.
+    // On failure we intentionally do NOT delete the orphan blob — concurrent
+    // uploads of the same hash could race and delete a blob that another
+    // request is about to reference via its sidecar. Orphan blobs are
+    // content-addressed and bounded by the upload size limit, so the storage
+    // cost is negligible. A V2 background GC job can sweep blobs with no
+    // matching sidecar after a grace period.
+    storage.put(&key, &body, &mime).await?;
+
+    let meta = match prepare_metadata(MetadataInput {
+        sha256: sha256.clone(),
+        ext: ext.clone(),
+        mime: mime.clone(),
+        body: body.clone(),
+        uploaded_at,
+    })
+    .await
+    {
+        Ok(meta) => meta,
+        Err(e) => {
+            tracing::warn!(sha256 = %sha256, error = %e, "metadata generation failed; orphan blob left for GC");
+            return Err(e);
+        }
+    };
+
+    // The moderation record precedes the sidecar publish gate. If this write
+    // fails, the blob and any thumbnail remain orphaned but the media cannot be
+    // served. Conversely, record existence still implies those objects exist.
+    if let Some(attribution) = &attribution {
+        record_upload_event(
+            storage,
+            ctx,
+            &auth_event.pubkey,
+            attribution,
+            UploadEventFacts {
+                sha256: &sha256,
+                ext: &ext,
+                mime: &mime,
+                size: body.len() as u64,
+                uploaded_at,
+            },
+        )
+        .await?;
+    }
+    storage.put_sidecar(ctx, &sha256, &meta).await?;
+
+    Ok(build_descriptor(
+        config,
+        &sha256,
+        &ext,
+        &mime,
+        body.len() as u64,
+        Some(&meta),
+        uploaded_at,
+    ))
+}
+
+/// Inputs handed to a buffered-upload metadata builder, after the shared
+/// pipeline has already validated, hashed, and stored the blob. Owned so the
+/// builder's future doesn't borrow the pipeline's locals; `body` is a `Bytes`
+/// handle, so cloning it is a refcount bump, not a copy.
+struct MetadataInput {
+    sha256: String,
+    ext: String,
+    mime: String,
+    body: Bytes,
+    uploaded_at: i64,
+}
+
+/// Process an upload end-to-end: validate, store, thumbnail, return descriptor.
+///
+/// This is the image path — body is already fully buffered in RAM. Do NOT use
+/// this for video uploads; use [`process_video_upload`] instead.
+pub async fn process_upload(
+    storage: &MediaStorage,
+    config: &MediaConfig,
+    ctx: &TenantContext,
+    auth_event: &nostr::Event,
+    body: Bytes,
+    attribution: Option<UploadAttribution>,
+) -> Result<BlobDescriptor, MediaError> {
+    process_buffered_upload(
+        BufferedUploadInput {
+            storage,
+            config,
+            ctx,
+            auth_event,
+            body,
+            attribution,
+        },
+        |bytes, cfg| {
+            let mime = validate_content(bytes, cfg)?;
+            let ext = mime_to_ext(&mime).to_string();
+            Ok((mime, ext))
+        },
+        |input| async move { prepare_image_metadata(storage, config, input).await },
+    )
+    .await
+}
+
+/// Process a generic non-media file upload end-to-end.
+///
+/// This is the catch-all attachment path for documents, archives, text, and
+/// data. Recognized image, video, and audio formats fail closed instead of
+/// entering exact-byte storage without their format-specific location policy.
+/// The body is fully buffered in RAM (bounded by `config.max_file_bytes` at the
+/// transport layer), validated against the deny-list + size cap, stored, and
+/// recorded in a minimal sidecar. No thumbnail, dimensions, or duration.
+///
+/// The resulting blob is served with `Content-Disposition: attachment`, so the
+/// client always downloads it rather than rendering it inline.
+pub async fn process_file_upload(
+    storage: &MediaStorage,
+    config: &MediaConfig,
+    ctx: &TenantContext,
+    auth_event: &nostr::Event,
+    body: Bytes,
+    attribution: Option<UploadAttribution>,
+) -> Result<BlobDescriptor, MediaError> {
+    process_buffered_upload(
+        BufferedUploadInput {
+            storage,
+            config,
+            ctx,
+            auth_event,
+            body,
+            attribution,
+        },
+        |bytes, cfg| validate_file_content(bytes, cfg),
+        |input| async move {
+            // Minimal sidecar — no thumbnail/dim/blurhash/duration for generic files.
+            let meta = BlobMeta {
+                dim: String::new(),
+                blurhash: String::new(),
+                thumb_url: String::new(),
+                size: input.body.len() as u64,
+                ext: input.ext,
+                mime_type: input.mime,
+                uploaded_at: input.uploaded_at,
+                duration_secs: None,
+            };
+            Ok(meta)
+        },
+    )
+    .await
+}
+
+/// Process a video upload end-to-end using a streaming pipeline.
+///
+/// Unlike [`process_upload`], this function:
+/// 1. Streams the request body to a [`tempfile::NamedTempFile`] while computing
+///    SHA-256 incrementally — the full body is never in RAM simultaneously.
+/// 2. Verifies the Blossom auth event `x` tag against the computed hash.
+/// 3. Runs full MP4 validation (codec, duration, resolution, moov placement).
+/// 4. Stores the blob via [`MediaStorage::put_file`] (streaming read from disk).
+/// 5. Writes a sidecar with `duration_secs` (no thumbnail — desktop handles that).
+///
+/// Returns a [`BlobDescriptor`] with the `duration` field populated.
+pub async fn process_video_upload(
+    storage: &MediaStorage,
+    config: &MediaConfig,
+    ctx: &TenantContext,
+    auth_event: &nostr::Event,
+    body_stream: impl futures_core::Stream<Item = Result<Bytes, axum::Error>> + Send + 'static,
+    content_length: Option<u64>,
+    attribution: Option<UploadAttribution>,
+) -> Result<BlobDescriptor, MediaError> {
+    // --- 1. Stream body to temp file, compute SHA-256 incrementally ---
+    let tmp = tempfile::NamedTempFile::new().map_err(|e| MediaError::Io(e.to_string()))?;
+    let tmp_path = tmp.path().to_path_buf();
+
+    let max_bytes = config.max_video_bytes;
+
+    // Fast-fail: reject oversized uploads before streaming starts.
+    if let Some(cl) = content_length {
+        if cl > max_bytes {
+            return Err(MediaError::FileTooLarge {
+                size: cl,
+                max: max_bytes,
+            });
+        }
+    }
+
+    let (sha256_hex, file_size, first_bytes) = {
+        use tokio_util::io::StreamReader;
+
+        // Convert axum::Error stream to std::io::Error stream for StreamReader.
+        // Box::pin is required because StreamReader needs a pinned stream.
+        // Belt-and-suspenders body-limit detection: axum wraps LengthLimitError
+        // in its error chain but doesn't expose the inner type for downcasting.
+        // We check multiple Display strings so that if axum changes the wording,
+        // at least one pattern still matches. test_body_limit_error_detection
+        // will catch a regression if ALL patterns break.
+        let mapped = futures_util::StreamExt::map(body_stream, |r| {
+            r.map_err(|e| {
+                let msg = e.to_string();
+                if msg.contains("length limit")
+                    || msg.contains("body limit")
+                    || msg.contains("LengthLimitError")
+                {
+                    std::io::Error::new(std::io::ErrorKind::WriteZero, msg)
+                } else {
+                    std::io::Error::other(e)
+                }
+            })
+        });
+        let mut reader = StreamReader::new(Box::pin(mapped));
+
+        let mut file = tokio::fs::File::create(&tmp_path)
+            .await
+            .map_err(|e| MediaError::Io(e.to_string()))?;
+        let mut hasher = Sha256::new();
+        let mut total: u64 = 0;
+        // Accumulate enough leading bytes for magic-byte detection.
+        // 4 KiB is the standard sniff buffer — infer checks signatures at
+        // various offsets, and some formats need more than just the first few
+        // bytes. This is tiny relative to any real upload.
+        const MIN_SNIFF_BYTES: usize = 4096;
+        let mut sniff_buf: Vec<u8> = Vec::with_capacity(MIN_SNIFF_BYTES);
+        let mut buf = vec![0u8; 64 * 1024]; // 64 KiB read buffer
+
+        loop {
+            use tokio::io::AsyncReadExt;
+            let n = match reader.read(&mut buf).await {
+                Ok(n) => n,
+                Err(e) if e.kind() == std::io::ErrorKind::WriteZero => {
+                    // Body limit exceeded — return 413 instead of 500.
+                    // `total` is bytes received before the cutoff — honest, not exact.
+                    return Err(MediaError::FileTooLarge {
+                        size: total,
+                        max: max_bytes,
+                    });
+                }
+                Err(e) => return Err(MediaError::Io(e.to_string())),
+            };
+            if n == 0 {
+                break;
+            }
+            total += n as u64;
+            if total > max_bytes {
+                return Err(MediaError::FileTooLarge {
+                    size: total,
+                    max: max_bytes,
+                });
+            }
+            hasher.update(&buf[..n]);
+            file.write_all(&buf[..n])
+                .await
+                .map_err(|e| MediaError::Io(e.to_string()))?;
+            if sniff_buf.len() < MIN_SNIFF_BYTES {
+                let need = MIN_SNIFF_BYTES - sniff_buf.len();
+                sniff_buf.extend_from_slice(&buf[..n.min(need)]);
+            }
+        }
+        file.flush()
+            .await
+            .map_err(|e| MediaError::Io(e.to_string()))?;
+
+        let sha256_hex = hex::encode(hasher.finalize());
+        (sha256_hex, total, sniff_buf)
+    };
+
+    // --- 2. ISO-BMFF/MP4 structural check ---
+    // Do not depend on `infer`'s finite major-brand list: valid MP4 producers
+    // may use a proprietary major brand while declaring `isom` compatibility.
+    if !looks_like_mp4_iso_bmff(&first_bytes) {
+        return Err(MediaError::UnsupportedContainer);
+    }
+    let mime = "video/mp4".to_string();
+
+    // --- 3. Verify Blossom auth: x tag must match computed SHA-256 ---
+    let auth = auth_event.clone();
+    let sha256_for_auth = sha256_hex.clone();
+    // Validate the Blossom `server` tag against the bound tenant host (not a
+    // process-global domain) — a relay serves many tenant hosts.
+    let bound_host = ctx.host().to_string();
+    tokio::task::spawn_blocking(move || {
+        // Videos: 1-hour window — large uploads on slow connections need headroom.
+        verify_blossom_upload_auth(&auth, &sha256_for_auth, Some(bound_host.as_str()), 3600)
+    })
+    .await
+    .map_err(|_| MediaError::Internal)??;
+
+    // --- 4. Full MP4 validation on the temp file ---
+    let tmp_path_clone = tmp_path.clone();
+    let cfg = config.clone();
+    let video_meta =
+        tokio::task::spawn_blocking(move || validate_video_file(&tmp_path_clone, &cfg))
+            .await
+            .map_err(|_| MediaError::Internal)??;
+
+    let ext = "mp4";
+    let key = format!("{sha256_hex}.{ext}");
+    let meta_key = MediaStorage::ctx_sidecar_key(ctx, &sha256_hex);
+
+    // --- 5. Idempotency check ---
+    let sidecar_exists = storage.head(&meta_key).await?;
+    let blob_exists = storage.head(&key).await?;
+    if sidecar_exists && blob_exists {
+        let meta = storage.get_sidecar(ctx, &sha256_hex).await?;
+        // Re-upload of known bytes: still a distinct upload event — see the
+        // buffered path's short-circuit for the rationale.
+        if let Some(attribution) = &attribution {
+            record_upload_event(
+                storage,
+                ctx,
+                &auth_event.pubkey,
+                attribution,
+                UploadEventFacts {
+                    sha256: &sha256_hex,
+                    ext,
+                    mime: &mime,
+                    size: file_size,
+                    uploaded_at: chrono::Utc::now().timestamp(),
+                },
+            )
+            .await?;
+        }
+        return Ok(build_descriptor(
+            config,
+            &sha256_hex,
+            ext,
+            &mime,
+            file_size,
+            Some(&meta),
+            meta.uploaded_at,
+        ));
+    }
+
+    let uploaded_at = chrono::Utc::now().timestamp();
+
+    // --- 6. Stream blob from temp file to S3 ---
+    storage.put_file(&key, &tmp_path, &mime).await?;
+    drop(tmp); // Free temp file disk space immediately after S3 upload.
+
+    // --- 7. Build metadata (no thumbnail for video — desktop handles that) ---
+    let meta = BlobMeta {
+        dim: format!("{}x{}", video_meta.width, video_meta.height),
+        blurhash: String::new(),
+        thumb_url: String::new(),
+        ext: ext.to_string(),
+        mime_type: mime.clone(),
+        size: file_size,
+        uploaded_at,
+        duration_secs: Some(video_meta.duration_secs),
+    };
+
+    // Record before publishing the sidecar serve gate. See the buffered path.
+    if let Some(attribution) = &attribution {
+        record_upload_event(
+            storage,
+            ctx,
+            &auth_event.pubkey,
+            attribution,
+            UploadEventFacts {
+                sha256: &sha256_hex,
+                ext,
+                mime: &mime,
+                size: file_size,
+                uploaded_at,
+            },
+        )
+        .await?;
+    }
+    storage.put_sidecar(ctx, &sha256_hex, &meta).await?;
+
+    Ok(build_descriptor(
+        config,
+        &sha256_hex,
+        ext,
+        &mime,
+        file_size,
+        Some(&meta),
+        uploaded_at,
+    ))
+}
+
+/// Generate thumbnail and metadata without publishing the sidecar serve gate.
+/// Returns the completed [`BlobMeta`] on success.
+async fn prepare_image_metadata(
+    storage: &MediaStorage,
+    config: &MediaConfig,
+    input: MetadataInput,
+) -> Result<BlobMeta, MediaError> {
+    let body_ref = input.body.clone();
+    let mime_ref = input.mime.clone();
+    let ext_ref = input.ext.clone();
+    let sha256_ref = input.sha256.clone();
+    let cfg_ref = config.clone();
+    let (mut meta, thumb_bytes) = tokio::task::spawn_blocking(move || {
+        generate_image_metadata_sync(&cfg_ref, &sha256_ref, &body_ref, &mime_ref, &ext_ref)
+    })
+    .await
+    .map_err(|_| MediaError::Internal)??;
+
+    meta.uploaded_at = input.uploaded_at;
+
+    if let Some(ref tb) = thumb_bytes {
+        let thumb_key = format!("{}.thumb.jpg", input.sha256);
+        storage.put(&thumb_key, tb, "image/jpeg").await?;
+    }
+
+    Ok(meta)
+}
+
+fn build_descriptor(
+    config: &MediaConfig,
+    sha256: &str,
+    ext: &str,
+    mime: &str,
+    size: u64,
+    meta: Option<&BlobMeta>,
+    uploaded_at: i64,
+) -> BlobDescriptor {
+    let duration = meta.and_then(|m| m.duration_secs);
+    BlobDescriptor {
+        url: format!("{}/{sha256}.{ext}", config.public_base_url),
+        sha256: sha256.to_string(),
+        size,
+        mime_type: mime.to_string(),
+        uploaded: uploaded_at,
+        dim: meta.and_then(|m| (!m.dim.is_empty()).then(|| m.dim.clone())),
+        blurhash: meta.and_then(|m| (!m.blurhash.is_empty()).then(|| m.blurhash.clone())),
+        thumb: meta.and_then(|m| (!m.thumb_url.is_empty()).then(|| m.thumb_url.clone())),
+        duration,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn test_config() -> MediaConfig {
+        MediaConfig {
+            s3_endpoint: String::new(),
+            s3_access_key: String::new(),
+            s3_secret_key: String::new(),
+            s3_bucket: String::new(),
+            s3_region: "us-east-1".to_string(),
+            s3_addressing_style: crate::config::S3AddressingStyle::Path,
+            max_image_bytes: 50 * 1024 * 1024,
+            max_gif_bytes: 10 * 1024 * 1024,
+            max_video_bytes: 524_288_000,
+            max_file_bytes: 104_857_600,
+            public_base_url: "https://media.example.com".to_string(),
+            upload_records_enabled: false,
+            upload_ip_header: None,
+            upload_port_header: None,
+        }
+    }
+
+    #[test]
+    fn test_build_descriptor_video_omits_empty_thumb_and_blurhash() {
+        // Video uploads produce a BlobMeta with empty thumb_url and blurhash.
+        // build_descriptor must convert these to None so they're omitted from JSON.
+        let config = test_config();
+        let meta = BlobMeta {
+            dim: "320x240".to_string(),
+            blurhash: String::new(),  // empty — video has no blurhash
+            thumb_url: String::new(), // empty — video has no thumbnail
+            ext: "mp4".to_string(),
+            mime_type: "video/mp4".to_string(),
+            size: 5_000_000,
+            uploaded_at: 1700000000,
+            duration_secs: Some(29.5),
+        };
+
+        let desc = build_descriptor(
+            &config,
+            "abc123",
+            "mp4",
+            "video/mp4",
+            5_000_000,
+            Some(&meta),
+            1700000000,
+        );
+
+        // Empty strings must become None, not Some("")
+        assert!(
+            desc.blurhash.is_none(),
+            "blurhash should be None for video, got {:?}",
+            desc.blurhash
+        );
+        assert!(
+            desc.thumb.is_none(),
+            "thumb should be None for video, got {:?}",
+            desc.thumb
+        );
+        // Non-empty fields should be present
+        assert_eq!(desc.dim, Some("320x240".to_string()));
+        assert_eq!(desc.duration, Some(29.5));
+
+        // Verify JSON serialization omits the empty fields entirely
+        let json = serde_json::to_value(&desc).unwrap();
+        assert!(
+            json.get("blurhash").is_none(),
+            "blurhash should be absent from JSON"
+        );
+        assert!(
+            json.get("thumb").is_none(),
+            "thumb should be absent from JSON"
+        );
+        assert!(json.get("dim").is_some(), "dim should be present in JSON");
+        assert!(
+            json.get("duration").is_some(),
+            "duration should be present in JSON"
+        );
+    }
+
+    #[test]
+    fn test_build_descriptor_image_includes_thumb_and_blurhash() {
+        // Image uploads produce a BlobMeta with populated thumb_url and blurhash.
+        let config = test_config();
+        let hash = "a".repeat(64);
+        let meta = BlobMeta {
+            dim: "800x600".to_string(),
+            blurhash: "LEHV6nWB2yk8pyo0adR*.7kCMdnj".to_string(),
+            thumb_url: format!("https://media.example.com/{hash}.thumb.jpg"),
+            ext: "jpg".to_string(),
+            mime_type: "image/jpeg".to_string(),
+            size: 100_000,
+            uploaded_at: 1700000000,
+            duration_secs: None,
+        };
+
+        let desc = build_descriptor(
+            &config,
+            &hash,
+            "jpg",
+            "image/jpeg",
+            100_000,
+            Some(&meta),
+            1700000000,
+        );
+
+        assert_eq!(
+            desc.blurhash,
+            Some("LEHV6nWB2yk8pyo0adR*.7kCMdnj".to_string())
+        );
+        assert!(desc.thumb.is_some());
+        assert!(desc.duration.is_none());
+
+        // Verify JSON: duration should be absent, blurhash and thumb present
+        let json = serde_json::to_value(&desc).unwrap();
+        assert!(json.get("blurhash").is_some());
+        assert!(json.get("thumb").is_some());
+        assert!(
+            json.get("duration").is_none(),
+            "duration should be absent for images"
+        );
+    }
+
+    #[test]
+    fn test_body_limit_error_detection() {
+        // Verify that body-limit errors are mapped to WriteZero (which
+        // process_video_upload converts to FileTooLarge / 413).
+        // Must match the detection logic in process_video_upload exactly.
+        let detect = |msg: &str| -> std::io::ErrorKind {
+            if msg.contains("length limit")
+                || msg.contains("body limit")
+                || msg.contains("LengthLimitError")
+            {
+                std::io::ErrorKind::WriteZero
+            } else {
+                std::io::ErrorKind::Other
+            }
+        };
+
+        // All known patterns should trigger WriteZero.
+        assert_eq!(
+            detect("length limit exceeded"),
+            std::io::ErrorKind::WriteZero
+        );
+        assert_eq!(detect("body limit exceeded"), std::io::ErrorKind::WriteZero);
+        assert_eq!(detect("LengthLimitError"), std::io::ErrorKind::WriteZero);
+
+        // Non-limit errors should remain as Other.
+        assert_eq!(detect("connection reset"), std::io::ErrorKind::Other);
+    }
+
+    #[test]
+    fn test_build_descriptor_no_meta() {
+        // When meta is None, all optional fields should be None.
+        let config = test_config();
+        let desc = build_descriptor(
+            &config,
+            "abc123",
+            "jpg",
+            "image/jpeg",
+            100,
+            None,
+            1700000000,
+        );
+
+        assert!(desc.dim.is_none());
+        assert!(desc.blurhash.is_none());
+        assert!(desc.thumb.is_none());
+        assert!(desc.duration.is_none());
+    }
+}

crates/buzz-media/src/upload_record.rs

@@ -0,0 +1,419 @@
+//! Per-upload-event records — the moderation side channel.
+//!
+//! Content-addressed storage keys facts about *bytes*; moderation and legal
+//! reporting (e.g. NCMEC CyberTipline) need facts about *upload events*: who
+//! uploaded, when, from which network address. This module writes one small
+//! append-only JSON record per **accepted** upload — including the idempotent
+//! re-upload short-circuit, which does no blob PUT and is therefore invisible
+//! to any blob-creation-driven pipeline.
+//!
+//! Key layout (alongside the existing `_meta/` sidecar convention):
+//!
+//! ```text
+//! _uploads/{community}/{sha256}/{event_id}.json
+//! ```
+//!
+//! `event_id` is a ULID — unique and time-sortable, one record per accepted
+//! upload event. Records are unreachable through the media serve path by
+//! construction (`validate_media_path` requires a bare 64-hex first segment),
+//! and the bucket is only accessible via the relay's IAM role.
+//!
+//! The whole feature is **off by default** and gated behind
+//! `BUZZ_MEDIA_UPLOAD_RECORDS`. IP collection is a second, independent opt-in
+//! (`BUZZ_MEDIA_UPLOAD_IP_HEADER`) and is *fail-empty*: a missing, malformed,
+//! or non-public address records nothing — a wrong IP is worse than no IP,
+//! so absent is always preferable. The IP goes only into this
+//! record — never blob metadata, never the upload response, never the
+//! hash-chained audit log.
+//!
+//! ## Consumer contract (buzz-moderation)
+//!
+//! The moderation pipeline triggers on `ObjectCreated` events under the
+//! `_uploads/` prefix and parses this record instead of HEADing blobs:
+//!
+//! - For fresh uploads, the record is written after the blob and derived
+//!   artifacts but before the sidecar serve gate. Record existence therefore
+//!   implies the scan inputs are readable, while record failure cannot leave
+//!   unscanned media publicly servable.
+//! - `ext`, `mime_type`, and `size` are always present so the consumer can
+//!   derive the blob key (`{sha256}.{ext}`) and scan eligibility without
+//!   extra round-trips.
+//! - `uploader_name`, `ip`, and `port` are omitted (never `null`) when
+//!   unknown or when collection is disabled.
+//! - Consumers must tolerate unknown fields; `version` bumps only on
+//!   breaking changes to existing fields.
+
+use std::net::IpAddr;
+
+use buzz_core::tenant::TenantContext;
+use serde::{Deserialize, Serialize};
+
+/// Current record schema version. Additive fields do not bump this.
+pub const UPLOAD_RECORD_VERSION: u32 = 1;
+
+/// One accepted upload event. See module docs for the consumer contract.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct UploadRecord {
+    /// Schema version ([`UPLOAD_RECORD_VERSION`]).
+    pub version: u32,
+    /// ULID — unique per accepted upload, time-sortable. Also the key suffix.
+    pub event_id: String,
+    /// Content hash of the uploaded bytes (64 lowercase hex chars).
+    pub sha256: String,
+    /// Canonical extension — consumers derive the blob key `{sha256}.{ext}`.
+    pub ext: String,
+    /// Sniffed MIME type of the uploaded bytes.
+    pub mime_type: String,
+    /// Size of the uploaded bytes.
+    pub size: u64,
+    /// Unix seconds when the relay accepted *this* upload event. On an
+    /// idempotent re-upload this is the re-upload time, not the original
+    /// blob's `uploaded_at`.
+    pub uploaded_at: i64,
+    /// Server-resolved community id (UUID). Never client-supplied.
+    pub community_id: String,
+    /// Server-resolved tenant host the upload was bound to.
+    pub community_host: String,
+    /// Authenticated uploader pubkey (64 lowercase hex chars).
+    pub uploader_id: String,
+    /// Same pubkey, bech32 `npub` encoding.
+    pub uploader_npub: String,
+    /// Uploader's configured display name at upload time. Best-effort label;
+    /// `uploader_id` is authoritative. Omitted when unset.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub uploader_name: Option<String>,
+    /// Uploader's public IP as reported by the configured edge header.
+    /// Present only when `BUZZ_MEDIA_UPLOAD_IP_HEADER` is set AND the header
+    /// held a valid public address (fail-empty). Omitted, never `null`.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub ip: Option<String>,
+    /// Uploader's source port as reported by the configured edge header.
+    /// Standard edge headers don't carry the client port, so this is
+    /// best-effort and usually absent. Only recorded alongside `ip`.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub port: Option<u16>,
+}
+
+/// Network address facts extracted from trusted edge headers by the HTTP
+/// handler, already validated (fail-empty). `Default` is "nothing collected".
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct UploadNetworkInfo {
+    /// Validated public IP of the uploader, or `None`.
+    pub ip: Option<IpAddr>,
+    /// Source port of the uploader, or `None`. Ignored when `ip` is `None`.
+    pub port: Option<u16>,
+}
+
+/// Per-event upload attribution passed into the upload pipeline by the
+/// handler. Only consulted when upload records are enabled.
+#[derive(Debug, Clone, Default)]
+pub struct UploadAttribution {
+    /// Uploader's configured display name, if known (sanitized, bounded).
+    pub uploader_name: Option<String>,
+    /// Validated network facts from trusted edge headers.
+    pub net: UploadNetworkInfo,
+}
+
+/// Facts about the accepted upload, computed by the upload pipeline.
+#[derive(Debug, Clone, Copy)]
+pub struct UploadEventFacts<'a> {
+    /// Content hash (64 lowercase hex chars).
+    pub sha256: &'a str,
+    /// Canonical extension.
+    pub ext: &'a str,
+    /// Sniffed MIME type.
+    pub mime: &'a str,
+    /// Uploaded byte size.
+    pub size: u64,
+    /// Unix seconds this upload event was accepted.
+    pub uploaded_at: i64,
+}
+
+/// Build and store the per-event record for one accepted upload.
+///
+/// Called after blob and derived-artifact durability but before the sidecar
+/// publish gate on fresh uploads; called on the existing published state for
+/// idempotent re-uploads. A write failure propagates and fails the upload. The
+/// record's `ObjectCreated` event is the moderation pipeline's only scan
+/// trigger, so no newly published media may exist without a record.
+pub async fn record_upload_event(
+    storage: &crate::storage::MediaStorage,
+    ctx: &TenantContext,
+    uploader: &nostr::PublicKey,
+    attribution: &UploadAttribution,
+    facts: UploadEventFacts<'_>,
+) -> Result<(), crate::error::MediaError> {
+    use nostr::ToBech32;
+
+    let event_id = ulid::Ulid::new().to_string();
+    // Ports are only meaningful next to the address they were observed with.
+    let ip = attribution.net.ip;
+    let port = ip.and(attribution.net.port);
+    let record = UploadRecord {
+        version: UPLOAD_RECORD_VERSION,
+        event_id: event_id.clone(),
+        sha256: facts.sha256.to_string(),
+        ext: facts.ext.to_string(),
+        mime_type: facts.mime.to_string(),
+        size: facts.size,
+        uploaded_at: facts.uploaded_at,
+        community_id: ctx.community().to_string(),
+        community_host: ctx.host().to_string(),
+        uploader_id: uploader.to_hex(),
+        uploader_npub: uploader.to_bech32().map_err(|e| {
+            // Unreachable for a valid pubkey; surfaced rather than unwrapped.
+            crate::error::MediaError::StorageError(format!("npub encoding failed: {e}"))
+        })?,
+        uploader_name: attribution.uploader_name.clone(),
+        ip: ip.map(|addr| addr.to_string()),
+        port,
+    };
+    let key = upload_record_key(ctx, facts.sha256, &event_id);
+    let json = serde_json::to_vec(&record)?;
+    storage.put(&key, &json, "application/json").await
+}
+
+/// Build the per-event record key:
+/// `_uploads/{community}/{sha256}/{event_id}.json`.
+///
+/// `ctx` must be the server-resolved request tenant — same fence as the
+/// `_meta/` sidecar key (see [`crate::storage::MediaStorage::sidecar_key`]).
+pub fn upload_record_key(ctx: &TenantContext, sha256: &str, event_id: &str) -> String {
+    format!("_uploads/{}/{sha256}/{event_id}.json", ctx.community())
+}
+
+/// Parse an IP header value, accepting only public addresses (fail-empty).
+///
+/// Returns `None` — record nothing — for anything that is not a single,
+/// syntactically valid, public IP: garbage, comma lists, private ranges,
+/// loopback, link-local, CGNAT, multicast, documentation, ULA, etc. Never
+/// guesses and never falls back to the socket address.
+pub fn parse_public_ip(raw: &str) -> Option<IpAddr> {
+    let ip: IpAddr = raw.trim().parse().ok()?;
+    is_public_ip(&ip).then_some(ip)
+}
+
+/// Parse a port header value: a single decimal u16, non-zero.
+pub fn parse_port(raw: &str) -> Option<u16> {
+    raw.trim().parse::<u16>().ok().filter(|&p| p != 0)
+}
+
+/// Conservative "is this a public, routable address" check.
+///
+/// `IpAddr::is_global` is unstable, so this enumerates the reserved ranges
+/// explicitly. Anything not recognizably public is rejected — the cost of a
+/// false negative is an absent field; the cost of a false positive is a wrong
+/// address in a federal report.
+fn is_public_ip(ip: &IpAddr) -> bool {
+    match ip {
+        IpAddr::V4(v4) => {
+            let octets = v4.octets();
+            !(v4.is_private()
+                || v4.is_loopback()
+                || v4.is_link_local()
+                || v4.is_broadcast()
+                || v4.is_documentation()
+                || v4.is_multicast()
+                || v4.is_unspecified()
+                // This network 0.0.0.0/8 (RFC 1122)
+                || octets[0] == 0
+                // CGNAT 100.64.0.0/10 (RFC 6598)
+                || (octets[0] == 100 && (octets[1] & 0b1100_0000) == 64)
+                // Reserved 240.0.0.0/4 (RFC 1112) — is_broadcast covers .255 only
+                || octets[0] >= 240
+                // Benchmarking 198.18.0.0/15 (RFC 2544)
+                || (octets[0] == 198 && (octets[1] & 0xFE) == 18)
+                // IETF protocol assignments 192.0.0.0/24 (RFC 6890)
+                || (octets[0] == 192 && octets[1] == 0 && octets[2] == 0))
+        }
+        IpAddr::V6(v6) => {
+            let seg = v6.segments();
+            !(v6.is_loopback()
+                || v6.is_multicast()
+                || v6.is_unspecified()
+                // Unique local fc00::/7 (RFC 4193)
+                || (seg[0] & 0xFE00) == 0xFC00
+                // Link-local fe80::/10 (RFC 4291)
+                || (seg[0] & 0xFFC0) == 0xFE80
+                // Discard-only 100::/64 (RFC 6666)
+                || (seg[0] == 0x0100 && seg[1..4] == [0, 0, 0])
+                // Teredo 2001::/32 (RFC 4380)
+                || (seg[0] == 0x2001 && seg[1] == 0)
+                // Benchmarking 2001:2::/48 (RFC 5180)
+                || (seg[0] == 0x2001 && seg[1] == 2 && seg[2] == 0)
+                // Documentation 2001:db8::/32 (RFC 3849)
+                || (seg[0] == 0x2001 && seg[1] == 0x0DB8)
+                // 6to4 2002::/16 (RFC 3056)
+                || seg[0] == 0x2002
+                // Documentation 3fff::/20 (RFC 9637)
+                || (seg[0] & 0xFFF0) == 0x3FF0
+                // IPv4-mapped ::ffff:0:0/96 — the embedded v4 was already
+                // rejected above if it arrived as dotted quad; reject the
+                // mapped form outright rather than re-deriving it.
+                || v6.to_ipv4_mapped().is_some())
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use buzz_core::tenant::CommunityId;
+
+    fn tenant() -> TenantContext {
+        TenantContext::resolved(
+            CommunityId::from_uuid(uuid::Uuid::from_u128(7)),
+            "chat.example.com",
+        )
+    }
+
+    #[test]
+    fn record_key_matches_spec_layout() {
+        let ctx = tenant();
+        let sha = "a".repeat(64);
+        let key = upload_record_key(&ctx, &sha, "01J9W3ULIDULIDULIDULIDULID");
+        assert_eq!(
+            key,
+            format!(
+                "_uploads/{}/{sha}/01J9W3ULIDULIDULIDULIDULID.json",
+                ctx.community()
+            )
+        );
+    }
+
+    #[test]
+    fn record_serializes_full_shape() {
+        let record = UploadRecord {
+            version: UPLOAD_RECORD_VERSION,
+            event_id: "01J9W3TEST".into(),
+            sha256: "b".repeat(64),
+            ext: "png".into(),
+            mime_type: "image/png".into(),
+            size: 12345,
+            uploaded_at: 1_783_358_352,
+            community_id: uuid::Uuid::from_u128(7).to_string(),
+            community_host: "chat.example.com".into(),
+            uploader_id: "c".repeat(64),
+            uploader_npub: "npub1example".into(),
+            uploader_name: Some("alice".into()),
+            ip: Some("203.0.113.7".into()),
+            port: Some(51234),
+        };
+        let json = serde_json::to_value(&record).unwrap();
+        assert_eq!(json["version"], 1);
+        assert_eq!(json["ext"], "png");
+        assert_eq!(json["mime_type"], "image/png");
+        assert_eq!(json["size"], 12345);
+        assert_eq!(json["ip"], "203.0.113.7");
+        assert_eq!(json["port"], 51234);
+        assert_eq!(json["uploader_name"], "alice");
+    }
+
+    #[test]
+    fn record_omits_absent_optionals_entirely() {
+        let record = UploadRecord {
+            version: UPLOAD_RECORD_VERSION,
+            event_id: "01J9W3TEST".into(),
+            sha256: "b".repeat(64),
+            ext: "mp4".into(),
+            mime_type: "video/mp4".into(),
+            size: 1,
+            uploaded_at: 0,
+            community_id: "cid".into(),
+            community_host: "h".into(),
+            uploader_id: "c".repeat(64),
+            uploader_npub: "npub1example".into(),
+            uploader_name: None,
+            ip: None,
+            port: None,
+        };
+        let json = serde_json::to_value(&record).unwrap();
+        // Omitted, not null — the consumer contract.
+        assert!(json.get("uploader_name").is_none());
+        assert!(json.get("ip").is_none());
+        assert!(json.get("port").is_none());
+    }
+
+    #[test]
+    fn record_deserialization_tolerates_unknown_fields() {
+        // Forward-compat: additive relay fields must not break older parsers
+        // of the same version (mirrors the consumer's requirement).
+        let json = r#"{
+            "version": 1, "event_id": "01J", "sha256": "ab", "ext": "png",
+            "mime_type": "image/png", "size": 1, "uploaded_at": 2,
+            "community_id": "c", "community_host": "h",
+            "uploader_id": "u", "uploader_npub": "n",
+            "some_future_field": {"nested": true}
+        }"#;
+        let record: UploadRecord = serde_json::from_str(json).unwrap();
+        assert_eq!(record.version, 1);
+        assert_eq!(record.ip, None);
+    }
+
+    #[test]
+    fn public_ips_accepted() {
+        for raw in [
+            "8.8.8.8",
+            "1.1.1.1",
+            "  93.184.216.34  ", // trims whitespace
+            "2600:1f18::1",
+            "2a00:1450:4009:81f::200e",
+        ] {
+            assert!(parse_public_ip(raw).is_some(), "should accept {raw}");
+        }
+    }
+
+    #[test]
+    fn non_public_ips_fail_empty() {
+        for raw in [
+            "",
+            "not-an-ip",
+            "10.0.0.1",         // private
+            "172.16.0.1",       // private
+            "192.168.1.1",      // private
+            "127.0.0.1",        // loopback
+            "169.254.1.1",      // link-local
+            "100.64.0.1",       // CGNAT
+            "100.127.255.255",  // CGNAT upper edge
+            "0.0.0.0",          // unspecified
+            "0.1.2.3",          // this network 0.0.0.0/8
+            "255.255.255.255",  // broadcast
+            "224.0.0.1",        // multicast
+            "240.0.0.1",        // reserved
+            "198.18.0.1",       // benchmarking
+            "192.0.0.1",        // IETF assignments
+            "203.0.113.7",      // documentation (TEST-NET-3)
+            "198.51.100.1",     // documentation (TEST-NET-2)
+            "192.0.2.1",        // documentation (TEST-NET-1)
+            "::1",              // v6 loopback
+            "::",               // v6 unspecified
+            "fe80::1",          // v6 link-local
+            "fc00::1",          // v6 ULA
+            "fd12:3456::1",     // v6 ULA
+            "ff02::1",          // v6 multicast
+            "100::1",           // v6 discard-only
+            "2001::1",          // Teredo
+            "2001:2::1",        // v6 benchmarking
+            "2001:db8::1",      // v6 documentation
+            "2002::1",          // 6to4
+            "3fff::1",          // v6 documentation
+            "::ffff:8.8.8.8",   // v4-mapped — reject the mapped form
+            "8.8.8.8, 1.1.1.1", // comma list — not a single IP
+            "8.8.8.8:443",      // ip:port — not a bare IP
+        ] {
+            assert!(parse_public_ip(raw).is_none(), "should reject {raw:?}");
+        }
+    }
+
+    #[test]
+    fn port_parses_single_nonzero_u16() {
+        assert_eq!(parse_port("51234"), Some(51234));
+        assert_eq!(parse_port(" 443 "), Some(443));
+        assert_eq!(parse_port("0"), None);
+        assert_eq!(parse_port("65536"), None);
+        assert_eq!(parse_port("-1"), None);
+        assert_eq!(parse_port("443, 444"), None);
+        assert_eq!(parse_port("abc"), None);
+        assert_eq!(parse_port(""), None);
+    }
+}

crates/buzz-media/tests/fixtures/android/README.md

@@ -0,0 +1,55 @@
+# Android Bitmap media fixtures
+
+These 3 x 2 fixtures were produced by Android 16 (API 36) `Bitmap.compress`, not by a generic image encoder.
+
+## Regeneration
+
+1. Compile and run the following program on an API 36 emulator:
+
+   ```java
+   import android.graphics.Bitmap;
+   import android.graphics.Color;
+   import android.graphics.ColorSpace;
+   import java.io.FileOutputStream;
+
+   public final class Main {
+     private static void write(Bitmap bitmap, Bitmap.CompressFormat format, String stem)
+         throws Exception {
+       String extension = format == Bitmap.CompressFormat.PNG ? "png" : "jpg";
+       try (FileOutputStream output =
+           new FileOutputStream("/data/local/tmp/" + stem + "." + extension)) {
+         if (!bitmap.compress(format, 100, output)) {
+           throw new IllegalStateException("Bitmap.compress failed for " + stem);
+         }
+       }
+     }
+
+     public static void main(String[] args) throws Exception {
+       Bitmap srgb = Bitmap.createBitmap(3, 2, Bitmap.Config.ARGB_8888);
+       srgb.setPixels(new int[] {
+         Color.argb(255, 255, 0, 0), Color.argb(255, 0, 255, 0),
+         Color.argb(255, 0, 0, 255), Color.argb(128, 255, 255, 0),
+         Color.argb(64, 0, 255, 255), Color.argb(0, 255, 0, 255),
+       }, 0, 3, 0, 0, 3, 2);
+       write(srgb, Bitmap.CompressFormat.PNG, "bitmap-srgb");
+       write(srgb, Bitmap.CompressFormat.JPEG, "bitmap-srgb");
+
+       Bitmap displayP3 = Bitmap.createBitmap(
+           3, 2, Bitmap.Config.RGBA_F16, true,
+           ColorSpace.get(ColorSpace.Named.DISPLAY_P3));
+       displayP3.eraseColor(Color.pack(
+           1.0f, 0.0f, 0.0f, 1.0f,
+           ColorSpace.get(ColorSpace.Named.DISPLAY_P3)));
+       write(displayP3, Bitmap.CompressFormat.PNG, "bitmap-display-p3");
+       write(displayP3, Bitmap.CompressFormat.JPEG, "bitmap-display-p3");
+     }
+   }
+   ```
+
+2. Pull the four files from `/data/local/tmp/` into this directory. Copy all four into `mobile/android/app/src/test/resources/fixtures/android/`, and copy `bitmap-display-p3.png` and `bitmap-srgb.png` into `mobile/android/app/src/androidTest/resources/fixtures/android/`.
+3. Run `cmp` on every copied fixture to confirm it is byte-identical.
+4. Run the app's `AndroidImageProcessor.decodeSrgbBitmap` and `encodeAndScrub` path for both source color spaces and both formats on the API 36 emulator. Save the four outputs under `sanitized/` with the `-sanitized` suffix.
+5. Run `cargo test -p buzz-media android_` to verify the relay accepts every sanitized fixture while rejecting the three unsanitized fixtures that contain forbidden metadata.
+6. Run `cd mobile/android && ./gradlew app:testDebugUnitTest app:connectedDebugAndroidTest` with an API 36 emulator to verify structural scrubbing and the Display-P3 to sRGB conversion for sanitized PNG and JPEG output.
+
+Regenerate both the encoded inputs and sanitized outputs whenever Android `Bitmap.compress` behavior or `AndroidMediaSanitizer` changes. Do not update only the sanitized files, because the tests cover the exact encoder-to-sanitizer boundary.

crates/buzz-media/tests/fixtures/ios/README.md

@@ -0,0 +1,51 @@
+# UIKit media fixtures
+
+These 2 x 2 fixtures were produced on an iOS simulator with UIKit, not by a generic image encoder.
+
+## Regeneration
+
+1. Create a small source image and run this program against the simulator SDK:
+
+   ```swift
+   import Foundation
+   import UIKit
+
+   let arguments = CommandLine.arguments
+   let source = try Data(contentsOf: URL(fileURLWithPath: arguments[1]))
+   guard
+     let image = UIImage(data: source),
+     let png = image.pngData(),
+     let jpeg = image.jpegData(compressionQuality: 1.0)
+   else {
+     fatalError("UIKit could not encode the source image")
+   }
+   try png.write(to: URL(fileURLWithPath: arguments[2]))
+   try jpeg.write(to: URL(fileURLWithPath: arguments[3]))
+   ```
+
+   Compile and run it with the active Xcode toolchain:
+
+   ```sh
+   SDK_PATH="$(xcrun --sdk iphonesimulator --show-sdk-path)"
+   xcrun --sdk iphonesimulator swiftc \
+     -sdk "$SDK_PATH" \
+     -target arm64-apple-ios16.0-simulator \
+     reencode.swift -o reencode
+   xcrun simctl spawn booted ./reencode \
+     source.png uikit-encoded.png uikit-encoded.jpg
+   ```
+
+2. Copy the encoded files into both fixture directories:
+
+   ```sh
+   cp uikit-encoded.png mobile/ios/RunnerTests/Fixtures/UIKitEncoded.png
+   cp uikit-encoded.jpg mobile/ios/RunnerTests/Fixtures/UIKitEncoded.jpg
+   cp uikit-encoded.png crates/buzz-media/tests/fixtures/ios/
+   cp uikit-encoded.jpg crates/buzz-media/tests/fixtures/ios/
+   ```
+
+3. Add a temporary Runner test that loads `UIKitEncoded.png` and `UIKitEncoded.jpg`, calls `MediaSanitizer.scrubPng` and `MediaSanitizer.scrubJpeg`, and writes those outputs to `uikit-sanitized.png` and `uikit-sanitized.jpg`. Run it once, copy the files here, then remove the temporary test.
+4. Run `cmp` on each encoded copy to confirm that the Runner and Rust fixtures are byte-identical.
+5. Run `cargo test -p buzz-media test_ios_uikit` to verify that UIKit's encoded output is rejected and the matching sanitizer output is accepted by the relay contract.
+
+Regenerate both encoded and sanitized pairs whenever UIKit encoding or `MediaSanitizer` changes. Do not update only the sanitized files, because the test is intended to cover the exact encoder-to-sanitizer boundary.

crates/buzz-media/tests/static_creds_minio.rs

@@ -0,0 +1,80 @@
+//! Live round-trip test for the **static-credentials** S3 path against an
+//! S3-compatible service. It is guarded by `#[ignore]`.
+//!
+//! This is the path local/dev and any static-key deployment uses
+//! (`s3_access_key`/`s3_secret_key` both non-empty -> `Credentials::new`). It
+//! exists to prove that adding the IRSA/credential-chain fallback did **not**
+//! regress hardcoded credentials.
+//!
+//! Run it against the docker-compose MinIO (creds `buzz_dev`/`buzz_dev_secret`,
+//! bucket `buzz-media`, endpoint `http://localhost:9000`):
+//!
+//! ```bash
+//! docker compose up -d minio minio-init
+//! cargo test -p buzz-media --test static_creds_minio -- --ignored
+//! ```
+//!
+//! Overridable via `BUZZ_S3_ENDPOINT` / `BUZZ_S3_ACCESS_KEY` /
+//! `BUZZ_S3_SECRET_KEY` / `BUZZ_S3_BUCKET` / `BUZZ_S3_REGION` /
+//! `BUZZ_S3_ADDRESSING_STYLE`. The default remains `path` for MinIO.
+
+use buzz_media::config::MediaConfig;
+use buzz_media::storage::MediaStorage;
+
+fn minio_config() -> MediaConfig {
+    MediaConfig {
+        s3_endpoint: std::env::var("BUZZ_S3_ENDPOINT")
+            .unwrap_or_else(|_| "http://localhost:9000".to_string()),
+        s3_access_key: std::env::var("BUZZ_S3_ACCESS_KEY")
+            .unwrap_or_else(|_| "buzz_dev".to_string()),
+        s3_secret_key: std::env::var("BUZZ_S3_SECRET_KEY")
+            .unwrap_or_else(|_| "buzz_dev_secret".to_string()),
+        s3_bucket: std::env::var("BUZZ_S3_BUCKET").unwrap_or_else(|_| "buzz-media".to_string()),
+        s3_region: std::env::var("BUZZ_S3_REGION").unwrap_or_else(|_| "us-east-1".to_string()),
+        s3_addressing_style: std::env::var("BUZZ_S3_ADDRESSING_STYLE")
+            .unwrap_or_else(|_| "path".to_string())
+            .parse()
+            .expect("BUZZ_S3_ADDRESSING_STYLE must be path or virtual"),
+        max_image_bytes: 50 * 1024 * 1024,
+        max_gif_bytes: 10 * 1024 * 1024,
+        max_video_bytes: 524_288_000,
+        max_file_bytes: 104_857_600,
+        public_base_url: "http://localhost:3000/media".to_string(),
+        upload_records_enabled: false,
+        upload_ip_header: None,
+        upload_port_header: None,
+    }
+}
+
+#[tokio::test]
+#[ignore = "requires a live MinIO (docker compose up -d minio minio-init)"]
+async fn static_creds_round_trip_against_minio() {
+    let storage =
+        MediaStorage::new(&minio_config()).expect("static creds should build a storage client");
+
+    let key = format!("_test/static-creds-{}.bin", std::process::id());
+    let body = b"hardcoded-creds-still-work";
+
+    // PUT
+    storage
+        .put(&key, body, "application/octet-stream")
+        .await
+        .expect("put with static creds should succeed");
+
+    // HEAD -> exists with correct size
+    assert!(storage.head(&key).await.expect("head should succeed"));
+    let meta = storage
+        .head_with_metadata(&key)
+        .await
+        .expect("head_with_metadata should succeed")
+        .expect("object should exist");
+    assert_eq!(meta.size, body.len() as u64);
+
+    // GET round-trips the bytes
+    let got = storage.get(&key).await.expect("get should succeed");
+    assert_eq!(got, body);
+
+    // DELETE, then HEAD reports absence
+    storage.delete(&key).await.expect("delete should succeed");
+    assert!(!storage.head(&key).await.expect("head after delete"));
+}