diff --git a/core/dr-decode/src/decoder.rs b/core/dr-decode/src/decoder.rs
new file mode 100644
index 0000000..ca46877
--- /dev/null
+++ b/core/dr-decode/src/decoder.rs
@@ -0,0 +1,120 @@
+//! TRACES: FR-RAW-2
+//! The seam a second decoder plugs into.
+//!
+//! D2 keeps LibRaw as the fallback for bodies rawler does not cover. Adding
+//! it later should be a new `impl Decoder`, not an edit to every caller that
+//! reads a header, cuts a thumbnail or opens a photograph for export — which
+//! is what the free functions alone would have made it. So the callers take a
+//! `&dyn Decoder`, and only the places that start a job name [`default`].
+//!
+//! Bytes in, always. Nothing here takes a path or a `SourceRef`: resolving a
+//! file to bytes is `Storage`'s job at the caller, so the same decoder serves a
+//! local file, an Android document and a range fetched from Nextcloud. The
+//! decoder's part in that is to say how much of a file it needs
+//! ([`Decoder::header_bytes`]) and where its preview sits
+//! ([`Decoder::locate_preview`]); the storage layer fetches exactly that.
+//!
+//! What stays a free function is what is not a decoder's to vary: recognising
+//! a JPEG ([`crate::probe`]), decoding one ([`crate::decode_jpeg`]) and
+//! checking one is whole ([`crate::is_complete_jpeg`]). A second RAW decoder
+//! would not read a JPEG differently.
+
+use dr_types::Orientation;
+
+use crate::{DecodeError, Metadata, Preview, PreviewLocation, PreviewSize, RawImage};
+
+/// TRACES: FR-RAW-2
+/// A RAW decoder, over bytes.
+///
+/// Object-safe so a caller can hold `&dyn Decoder` without becoming generic,
+/// and `Send + Sync` because the callers that need one most — the thumbnail
+/// lanes, the export worker — run off the UI thread.
+pub trait Decoder: Send + Sync {
+    /// How much of the start of a file [`Self::metadata`] and
+    /// [`Self::locate_preview`] need. A caller reading over a network fetches
+    /// this range and no more.
+    fn header_bytes(&self) -> u64;
+
+    /// Capture metadata, from a header or a whole file, without touching
+    /// sensor data.
+    fn metadata(&self, bytes: &[u8]) -> Result<Metadata, DecodeError>;
+
+    /// How the stored pixels are turned, from a header. `None` where the file
+    /// does not say, which callers take as upright.
+    fn orientation(&self, header: &[u8]) -> Option<Orientation>;
+
+    /// Where the embedded preview best suited to a thumbnail sits in the file,
+    /// from its header, so a remote caller can fetch that range alone.
+    fn locate_preview(&self, header: &[u8], file_len: u64) -> Option<PreviewLocation>;
+
+    /// The embedded preview at the size asked for, falling through the ladder
+    /// to the next size where the file lacks it.
+    fn preview(&self, bytes: &[u8], size: PreviewSize) -> Result<Preview, DecodeError>;
+
+    /// Sensor data, for develop and export. The expensive path.
+    fn decode(&self, bytes: &[u8]) -> Result<RawImage, DecodeError>;
+}
+
+/// TRACES: FR-RAW-2
+/// The decoder the application ships: rawler for sensor data and the
+/// previews it knows, DarkRoom's own container walk for headers and ranges.
+///
+/// Its methods are the crate's free functions, unchanged. They stay public
+/// for the tools and examples that read one file and have no caller to keep
+/// decoder-agnostic.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Rawler;
+
+impl Decoder for Rawler {
+    fn header_bytes(&self) -> u64 {
+        crate::HEADER_BYTES
+    }
+
+    fn metadata(&self, bytes: &[u8]) -> Result<Metadata, DecodeError> {
+        crate::metadata(bytes)
+    }
+
+    fn orientation(&self, header: &[u8]) -> Option<Orientation> {
+        crate::orientation(header)
+    }
+
+    fn locate_preview(&self, header: &[u8], file_len: u64) -> Option<PreviewLocation> {
+        crate::locate_preview(header, file_len)
+    }
+
+    fn preview(&self, bytes: &[u8], size: PreviewSize) -> Result<Preview, DecodeError> {
+        crate::extract_preview(bytes, size)
+    }
+
+    fn decode(&self, bytes: &[u8]) -> Result<RawImage, DecodeError> {
+        crate::decode(bytes)
+    }
+}
+
+/// TRACES: FR-RAW-2
+/// The decoder a job uses unless it was handed another.
+///
+/// Named by the places that start work — a thread, a UI handler — and by
+/// nothing below them. Returning `&'static dyn Decoder` rather than `Rawler`
+/// is the point: a caller that only has this cannot reach past the trait.
+pub fn default() -> &'static dyn Decoder {
+    static RAWLER: Rawler = Rawler;
+    &RAWLER
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The default is the shipped decoder, reached through the trait: same
+    /// header budget, and the same answer to bytes neither can read.
+    #[test]
+    fn the_default_is_rawler_behind_the_trait() {
+        let d = default();
+        assert_eq!(d.header_bytes(), crate::HEADER_BYTES);
+        let junk = [0u8; 64];
+        assert_eq!(d.metadata(&junk).is_err(), crate::metadata(&junk).is_err());
+        assert!(d.decode(&junk).is_err());
+        assert_eq!(d.orientation(&junk), crate::orientation(&junk));
+    }
+}
diff --git a/core/dr-decode/src/lib.rs b/core/dr-decode/src/lib.rs
index ddb5888..2334f84 100644
--- a/core/dr-decode/src/lib.rs
+++ b/core/dr-decode/src/lib.rs
@@ -11,14 +11,20 @@
 //!
 //! Fusing them would force a full decode where a header read suffices, which
 //! is exactly why Lightroom stalls ~2 s per image during culling.
+//!
+//! Callers reach these through the [`Decoder`] trait rather than by name, so a
+//! second decoder can be put behind them without changing any of them
+//! (FR-RAW-2). [`Rawler`] is the one that ships; [`default`] hands it out.
 
 pub mod base_curve;
+mod decoder;
 mod error;
 mod locate;
 mod preview;
 pub mod profile;
 
 pub use base_curve::BaseCurve;
+pub use decoder::{default, Decoder, Rawler};
 pub use error::DecodeError;
 pub use locate::{
     defects, is_complete_jpeg, jpeg_metadata, locate_preview, tiff_metadata, BadLine, BadPixel,