build: shrink the release binary with fat LTO and one codegen unit

.github/workflows/release.yml

@@ -131,14 +131,17 @@ jobs:
 
       - name: Build release binary
         env:
-          # Smaller, more portable binaries. LTO is on by default in
-          # the `release` profile; this just makes binaries strip-friendly.
+          # LTO and codegen units come from `[profile.release]` in
+          # Cargo.toml. Symbols are left in here and removed by the
+          # target's own `strip` in the next step.
           CARGO_PROFILE_RELEASE_STRIP: none
+        # `--bin hoverfly`: sigcheck isn't shipped, so don't pay for a
+        # second fat-LTO link.
         run: |
           if [ "${{ matrix.use-cross }}" = "true" ]; then
-            cross build --release --locked --target ${{ matrix.target }}
+            cross build --release --locked --target ${{ matrix.target }} --bin hoverfly
           else
-            cargo build --release --locked --target ${{ matrix.target }}
+            cargo build --release --locked --target ${{ matrix.target }} --bin hoverfly
           fi
 
       - name: Strip binary

AGENTS.md

@@ -70,6 +70,11 @@ are in flight at once. Public API (`Behaviour`, `Control`, `IncomingStreams`,
 
 - Native: `cargo build` / `cargo build --release`. Release is ~2-5× faster on
   crypto paths but only ~10-15% end-to-end (network dominates).
+- `[profile.release]` is tuned for binary size: fat LTO, `codegen-units = 1`.
+  A release build is a full slow rebuild; iterate on `cargo build`. Don't
+  switch it to `panic = "abort"` for the extra ~2 MiB: unwinding is what
+  confines a panic to the one tokio task (peer session, daemon request,
+  relay connection) it happened in.
 - Edition **2024**; crate version tracked in `Cargo.toml` (`0.x`, currently
   `0.3.0`).
 - Default features are `cli`, `bridge`, `pusher`. `--no-default-features

Cargo.toml

@@ -189,3 +189,14 @@ tokio-test = "0.4"
 # from `proto/*.proto`. Not pulled into normal `cargo build` — invoke
 # explicitly via `scripts/regen-protos.sh`.
 prost-build = "0.13"
+
+# Binary size over build time. Without LTO each of the default 16 codegen
+# units carries its own copy of every async body and closure it touches;
+# fat LTO with a single unit merges them. cargo can't LTO the cdylib+rlib lib
+# target, so the wasm build only gets the single unit. Panics stay `unwind`
+# despite the unwind tables: each peer session drives its swarm in its own
+# task, and a panic there (say, a dependency choking on a peer's bytes) must
+# drop that session, not the whole upload, daemon or relay.
+[profile.release]
+lto = "fat"
+codegen-units = 1