perf: switch release builds to Thin LTO

.github/copilot-instructions.md

@@ -220,15 +220,15 @@ The workspace ships with an aggressive release profile (`Cargo.toml`):
 
 ```toml
 [profile.release]
-lto = true            # Full link-time optimization
+lto = "thin"          # Thin link-time optimization
 codegen-units = 1     # Maximum optimization (slower compile)
 panic = "abort"       # No unwinding - smaller binary, but panics terminate immediately
 strip = true          # Strip debug symbols
 ```
 
 - **`panic = "abort"` means panics kill the process instantly** - reinforcing why `unwrap`/`expect` are banned in library code.
 - Always validate performance claims in `--release` mode. Debug builds are not representative.
-- Be aware that LTO + single codegen unit makes release builds slow. Use `cargo test` (debug) for iteration, `cargo build --release` for final validation.
+- Thin LTO with a single codegen unit keeps cross-crate optimization while reducing release build time compared with full LTO. Use `cargo test` (debug) for iteration, `cargo build --release` for final validation.
 
 ---
 

Cargo.toml

@@ -111,7 +111,7 @@ unnecessary_box_returns = "deny"
 unsafe_code = "deny"
 
 [profile.release]
-lto = true
+lto = "thin"
 codegen-units = 1
 panic = "abort"
 strip = true

xtask/src/main.rs

@@ -669,12 +669,12 @@ impl Step {
         run: || {
             // Use the dev profile (not the default bench profile) for the
             // criterion `--test` smoke run. The bench profile inherits the
-            // release profile's `lto = true, codegen-units = 1`, which spends
-            // ~40s compiling the full graph just to assert criterion's
+            // release profile's `lto = "thin", codegen-units = 1`, which still
+            // compiles the full optimized graph just to assert criterion's
             // `--test` entry point runs once. The dev profile reuses the
             // already-built unit-test artifacts and finishes in seconds. Real
             // benchmark measurements (via `cargo xtask bench`) continue to
-            // use the unchanged bench profile so their numbers stay accurate.
+            // use the release-derived bench profile so their numbers stay accurate.
             run_command_quiet(
                 "cargo",
                 &[