diff --git a/config/derivations.json b/config/derivations.json
index 4d670f267..4fa5eadbb 100644
--- a/config/derivations.json
+++ b/config/derivations.json
@@ -113,7 +113,7 @@
       },
       "regenerate": "bash tools/photos/extract-photo-metadata.sh \"$SOOC_DIR\" && bun tools/photos/build-histogram-index.ts",
       "recorded": {
-        "inputs": "4f8e62b9dd95c62238aea1ee278bd4b3e7e8b1a926c9bbc295585ccf4470bb28",
+        "inputs": "025a02d09757cb99791760ce11cddf1fc5d636942d254124409ad3f12e9973ba",
         "count": 261
       }
     },
diff --git a/config/derivations.lock.json b/config/derivations.lock.json
index e25f14c92..10680b585 100644
--- a/config/derivations.lock.json
+++ b/config/derivations.lock.json
@@ -1551,8 +1551,8 @@
     "public/i/XT509996.47144221.avif": "471442213dffcaf2ad49d282c3a2c6058f555c6a57bdd1de0fa7fe3d2af84612",
     "public/i/XT509996.7c90b12d.jpg": "7c90b12dc70fc92e8f5e836e0b28248f7f07929b2df5607bf8a601b2a4b1b7c4",
     "tools/photos/shell-data.ts": "9f579625d015d884ac8cae51ee6075df9993b3b658f894cadbd168cde05b29e1",
-    "tools/photos/zenc/Cargo.lock": "f8461ff344258f61c9ca715e93c3f12a549be5cc0d7dd89364de27cb143e5897",
+    "tools/photos/zenc/Cargo.lock": "f7a8fc33f3c9da7a2f287dbf239758d567a9ab6dc87fd4a20870b77b7f01b207",
     "tools/photos/zenc/src/histogram.rs": "59e46d9f6e65905da2f6019714963cfc0d2d0cd40b518f5f57c5a14377b6612a",
-    "tools/photos/zenc/src/pixels.rs": "4ba5b98e36b3013260e2368f15a4a25adf7d3bce8dba4b326accab66a5dbd786"
+    "tools/photos/zenc/src/pixels.rs": "0185c8a11c7155d01993669a2e2dfb603e1ce72561d27f15ad8d32117b1ed2ab"
   }
 }
diff --git a/tools/photos/zenc/Cargo.lock b/tools/photos/zenc/Cargo.lock
index cd9e66d8a..ed39afaf8 100644
--- a/tools/photos/zenc/Cargo.lock
+++ b/tools/photos/zenc/Cargo.lock
@@ -194,7 +194,7 @@ dependencies = [
 [[package]]
 name = "halflight"
 version = "0.1.0"
-source = "git+https://github.com/oddharsh/halflight.git?rev=80e9f344f669f85cbf4e85d9337b11405691f7ba#80e9f344f669f85cbf4e85d9337b11405691f7ba"
+source = "git+https://github.com/oddharsh/halflight.git?rev=871487e15be16dae0da39ddd882d94d188a5dc5f#871487e15be16dae0da39ddd882d94d188a5dc5f"
 
 [[package]]
 name = "hashbrown"
diff --git a/tools/photos/zenc/Cargo.toml b/tools/photos/zenc/Cargo.toml
index 2c73ac9f5..7bee1c07e 100644
--- a/tools/photos/zenc/Cargo.toml
+++ b/tools/photos/zenc/Cargo.toml
@@ -56,7 +56,14 @@ zenjpeg = { version = "0.8.4", features = ["trellis"] }
 # transfer curves) and the histogram bake over all 165 stems came back
 # identical from the binary built either way, and the same A/B is the gate for
 # moving this rev.
-halflight = { git = "https://github.com/oddharsh/halflight.git", rev = "80e9f344f669f85cbf4e85d9337b11405691f7ba" }
+#
+# Moved 2026-09-22 to 871487e for resample_oriented, which square now calls so
+# no upright copy of a full frame is built. `bun run zenc:bench` is that A/B
+# now: all 24 square outputs and 258 histograms byte-identical, and the
+# histograms regenerated from this build matched the committed file. Pin the
+# rev to halflight's branch commit and merge that PR with a merge commit (not a
+# squash), as #1 was, so this sha stays reachable from halflight's main.
+halflight = { git = "https://github.com/oddharsh/halflight.git", rev = "871487e15be16dae0da39ddd882d94d188a5dc5f" }
 # Input decode only. `tiff` was added on 2026-08-25 for the Instagram export,
 # which needs a LOSSLESS decode of a HEIF source and cannot use JPEG for it.
 # Measured on a 7728x5152 HIF: `sips -s format png` takes 6222ms because it
diff --git a/tools/photos/zenc/README.md b/tools/photos/zenc/README.md
index 2fc89558e..77040abcf 100644
--- a/tools/photos/zenc/README.md
+++ b/tools/photos/zenc/README.md
@@ -95,9 +95,12 @@ whole rows for the flips and walks 64px tiles for the four transposes, which
 took the decode-to-resample floor from 731/774/859 ms to 572/577/601 ms at
 orientations 3/6/8. On the bench's corpus, rotated sources run 1.09-1.17x
 faster and use 5-9.5% less CPU, each at p = 0.0006, with every output
-byte-identical. What orientation still costs over upright is the copy into a
-fresh full-frame buffer. Removing it means reading through the orientation
-inside halflight's resample, which is a halflight API change and its own A/B.
+byte-identical. The second pass removed the copy itself: `square` now hands the
+STORED frame to halflight's `resample_oriented`, which reads it upright inside
+the resample and is bitwise equal to orienting first. On one 26 MP JPEG the
+floor reads 491 ms upright and 491/445/461 ms at orientations 3/6/8, so a
+rotated photo now costs nothing extra, and the axis-swapping ones are cheaper
+than upright. `resize` and `frame` still orient with the tiled copy.
 
 ## Remaining work
 
diff --git a/tools/photos/zenc/examples/heif-memory.rs b/tools/photos/zenc/examples/heif-memory.rs
index b74626445..7cac3cad0 100644
--- a/tools/photos/zenc/examples/heif-memory.rs
+++ b/tools/photos/zenc/examples/heif-memory.rs
@@ -72,7 +72,7 @@ fn run(args: &[String]) -> Result<(), String> {
             let root = Path::new(&args[2]);
             std::fs::create_dir_all(root).map_err(|e| e.to_string())?;
             for size in [600, 400, 200] {
-                let tier = square::tier(&src, size, halflight::Filter::Box);
+                let tier = square::tier(&src, pixels::Orientation::Upright, size, halflight::Filter::Box);
                 avif::write(&tier, &root.join(format!("{size}.avif")))?;
                 if size == 600 {
                     let jpg = jpeg::encode(&tier, 84, zenjpeg::encoder::ChromaSubsampling::Quarter)?;
diff --git a/tools/photos/zenc/examples/memory-encode.rs b/tools/photos/zenc/examples/memory-encode.rs
index 9c8dbe2c9..9d6d0c759 100644
--- a/tools/photos/zenc/examples/memory-encode.rs
+++ b/tools/photos/zenc/examples/memory-encode.rs
@@ -27,7 +27,7 @@ fn run(args: &[String]) -> Result<(), String> {
             let src = pixels::orient(pixels::load_linear(&args[1], pixels::TransferOption::Auto)?, orientation);
             std::fs::create_dir_all(root).map_err(|e| e.to_string())?;
             for size in [600, 400, 200] {
-                let tier = square::tier(&src, size, halflight::Filter::Box);
+                let tier = square::tier(&src, pixels::Orientation::Upright, size, halflight::Filter::Box);
                 avif::write(&tier, &root.join(format!("{size}.avif")))?;
                 if size == 600 {
                     let encoded = jpeg::encode(&tier, 84, ChromaSubsampling::Quarter)?;
diff --git a/tools/photos/zenc/src/pixels.rs b/tools/photos/zenc/src/pixels.rs
index b2f629547..078547507 100644
--- a/tools/photos/zenc/src/pixels.rs
+++ b/tools/photos/zenc/src/pixels.rs
@@ -71,6 +71,20 @@ pub enum Orientation {
     Rotate270,
 }
 
+impl Orientation {
+    /// Whether the upright frame's width is the stored frame's height.
+    pub fn swaps_axes(self) -> bool {
+        self as u8 >= 5
+    }
+
+    /// The same transform in halflight's terms. Both enums number the eight
+    /// cases as EXIF does and map a stored pixel identically, which zenc:bench
+    /// gates on real photos byte for byte.
+    fn halflight(self) -> halflight::Orientation {
+        halflight::Orientation::from_exif(self as u8).expect("Orientation admits exactly EXIF 1-8")
+    }
+}
+
 impl TryFrom<u8> for Orientation {
     type Error = &'static str;
     fn try_from(value: u8) -> Result<Self, Self::Error> {
@@ -310,11 +324,11 @@ pub fn orient(f: Frame, orientation: Orientation) -> Frame {
     //   orientation 6   774 -> 577 ms   (tiled transpose)
     //   orientation 8   859 -> 601 ms   (tiled transpose)
     //
-    // What is left over upright is the copy into a fresh full-frame buffer.
-    // Removing that means reading through the orientation inside halflight's
-    // resample, which is a halflight API change. Output is byte-identical,
-    // since this moves the same values to the same places, and zenc:bench gates
-    // that on real photos.
+    // The ingest no longer calls this: `square` reads the stored frame upright
+    // inside the resample (scale_oriented, halflight::resample_oriented), which
+    // skips the copy this still makes. `resize` and `frame` use it. Output is
+    // byte-identical, since this moves the same values to the same places, and
+    // zenc:bench gates that on real photos.
     //
     // The channel count is a const generic so each pixel copy compiles to a
     // fixed-size move.
@@ -389,6 +403,20 @@ pub fn scale(f: &Frame, dw: u32, dh: u32, filter: Filter) -> Frame {
     }
 }
 
+/// Resample a STORED frame to an exact UPRIGHT size, reading it through the
+/// orientation inside the resample rather than orienting a full-frame copy
+/// first. halflight holds this bitwise equal to `scale(&orient(f, o), ..)`.
+pub fn scale_oriented(f: &Frame, o: Orientation, dw: u32, dh: u32, filter: Filter) -> Frame {
+    let ch = if f.gray { 1 } else { 3 };
+    Frame {
+        w: dw,
+        h: dh,
+        gray: f.gray,
+        transfer: f.transfer,
+        data: halflight::resample_oriented(&f.data, f.w as usize, f.h as usize, ch, o.halflight(), dw as usize, dh as usize, filter),
+    }
+}
+
 /// Quantize once after resampling; the returned pixels can feed several encoders.
 pub fn encoded(f: &Frame) -> DynamicImage {
     let t = f.transfer;
diff --git a/tools/photos/zenc/src/square.rs b/tools/photos/zenc/src/square.rs
index 72190a878..82fddafb2 100644
--- a/tools/photos/zenc/src/square.rs
+++ b/tools/photos/zenc/src/square.rs
@@ -17,20 +17,23 @@
 // `--transfer` names the SOURCE's curve (srgb, or g22 for the Monochrom's
 // Gray Gamma 2.2) and is used for decode and encode both, so unaveraged values
 // pass through exactly.
-use crate::pixels::{crop, load_linear, orient, parse_transfer, encoded, scale, Frame, Orientation, TransferOption};
+use crate::pixels::{crop, load_linear, parse_transfer, encoded, scale_oriented, Frame, Orientation, TransferOption};
 use halflight::Filter;
 use std::path::Path;
 
-/// The production geometry, shared with the benchmark front end.
-pub fn tier(src: &Frame, size: u32, filter: Filter) -> image::DynamicImage {
-    let (w, h) = (src.w, src.h);
+/// The production geometry, shared with the benchmark front end. `src` is the
+/// STORED frame and `orientation` brings it upright inside the resample, so no
+/// upright copy of the full frame is ever built. Geometry is computed on the
+/// upright size, exactly as it was when this took an oriented frame.
+pub fn tier(src: &Frame, orientation: Orientation, size: u32, filter: Filter) -> image::DynamicImage {
+    let (w, h) = if orientation.swaps_axes() { (src.h, src.w) } else { (src.w, src.h) };
     let (nw, nh) = if w <= h {
         (size, (h as u64 * size as u64).div_ceil(w as u64) as u32)
     } else {
         ((w as u64 * size as u64).div_ceil(h as u64) as u32, size)
     };
     let (cx, cy) = ((nw.saturating_sub(size)) / 2, (nh.saturating_sub(size)) / 2);
-    let scaled = scale(src, nw, nh, filter);
+    let scaled = scale_oriented(src, orientation, nw, nh, filter);
     encoded(&crop(&scaled, cx, cy, size.min(nw), size.min(nh)))
 }
 
@@ -114,13 +117,14 @@ pub fn run(args: &[String]) -> i32 {
         Ok(f) => f,
         Err(e) => return err(&e),
     };
-    // Orient BEFORE the resample, so the crop math sees the frame the viewer
-    // will. For orientation 1 the decoded buffer moves without copying.
-    let src = orient(src, exif);
+    // The orientation is applied INSIDE each tier's resample (tier() and
+    // halflight::resample_oriented), so the crop math still sees the frame the
+    // viewer will and no upright copy of the full frame is built. That copy was
+    // a 311 MB f32 buffer for a 26 MP photo, rebuilt once per ingest.
     for output in outputs {
         // Short edge lands on `size`. sips reaches the same crop from the long
         // edge, which is one more piece of arithmetic to get wrong.
-        let pixels = tier(&src, output.size, filter);
+        let pixels = tier(&src, exif, output.size, filter);
         if let Some(path) = output.png {
             if let Err(e) = pixels.save(path) { return err(&format!("cannot write {path}: {e}")); }
         }